Model-based predictive performance evaluation of machine-learned classifiers

Abstract

A computing system can obtain a labeled dataset correlating a plurality of classification labels with a plurality of data items. The computing system can obtain, for each data item of the plurality of data items, one or more machine-learned confidence values indicative of a confidence that the data item belongs to one or more classes. The computing system can assign, based on the confidence values, each data item to a confidence bin. The computing system can determine, based on the confidence bins and classification labels, a count matrix indicative of a plurality of classification label counts, each associated respectively with a corresponding confidence bin and class. The computing system can determine, based at least in part on the count matrix, a probability matrix correlating each respective confidence bin of the plurality of confidence bins with one or more corresponding classification probabilities associated with the respective confidence bin.

Description

MODEL-BASED PREDICTIVE PERFORMANCE EVALUATION OF MACHINE- LEARNED CLASSIFIERSPRIORITY CLAIM

[0001] The present application is based on and claims priority to United States Provisional Application Number 63 / 789,679 having a filing date of April 16, 2025. The present application claims priority to and the benefit of each of such applications and incorporates all such applications herein by reference in their entirety.BACKGROUND

[0002] The present disclosure relates generally to machine learning processes and machine-learned devices and systems. A computer can receive input(s). The computer can execute instructions to process the input(s) to generate output(s) using a parameterized model. The computer can obtain feedback on its performance in generating the outputs with the model. The computer can generate feedback by evaluating its performance. The computer can receive feedback from an external source. The computer can update parameters of the model based on the feedback to improve its performance. In this manner, the computer can iteratively ‘'learn” to generate the desired outputs. The resulting model is often referred to as a machine-learned model.SUMMARY

[0003] Aspects and advantages of embodiments of the present disclosure will be set forth in part in the following description, or can be learned from the description, or can be learned through practice of the embodiments.

[0004] Example aspects of the present disclosure provide an example method. In some implementations, the example method can include obtaining, by a computing system comprising one or more computing devices, a dataset comprising a plurality of data entries. In the example method, each data entry of the plurality of data entries can correlate a first data item of the data entry to one or more ground truth classification labels associated with the first data item. The example method can include obtaining, by the computing system, a plurality of machine-learned confidence values comprising, for each data entry of the plurality of data entries, one or more respective machine-learned confidence values indicative of a confidence that the first data item of the data entry belongs to one or more respectiveclasses associated with the one or more ground truth classification labels. The example method can include assigning, by the computing system based at least in part on the plurality of machine-learned confidence values, each data entry of the plurality of data entries to a corresponding confidence bin of a plurality of confidence bins. The example method can include determining, by the computing system based at least in part on the plurality of confidence bins and the one or more ground truth classification labels of the plurality of data entries, a count matrix indicative of a plurality of ground truth classification label counts. In the example method, each count of the plurality of ground truth classification label counts can be associated respectively with a corresponding confidence bin of the plurality of confidence bins and a corresponding class of a plurality of possible classes. The example method can include determining, by the computing system based at least in part on the count matrix, a probability matrix correlating each respective confidence bin of the plurality of confidence bins with one or more corresponding classification probabilities associated with the respective confidence bin.

[0005] In the example method, determining the probability matrix can include determining one or more parameters of a parametrized probability matrix based at least in part on a relationship between the one or more parameters and a probability of randomly sampling the count matrix given the probability matrix.

[0006] In the example method, the probability matrix can be a matrix product of a fixed basis matrix and a matrix of coefficient values. In the example method, determining the probability matrix can include determining the matrix of coefficient values using fixed point iteration.

[0007] In the example method, the fixed point iteration can include initializing, by the computing system, an initial value for the matrix of coefficient values. In the example method, the fixed point iteration can include, for each of a plurality of iterations, determining, by the computing system, an updated value for the matrix of coefficient values. In the example method, the updated value can be determined based at least in part on a current value of the matrix of coefficient values.

[0008] In the example method, determining the probability matrix can include Poisson factorization.

[0009] In the example method, the count matrix can include a weighted sum of count values.

[0010] In the example method, the plurality of machine-learned confidence values can include values that were generated using a first machine-learned model. In the examplemethod, the weighted sum of count values can include one or more count values indicative of one or more classification labels generated by a second machine-learned model different from the first machine-learned model.

[0011] The example method can include determining, by the computing system based at least in part on the probability matrix, an estimated performance metric associated with one or more first machine-learned models. The example method can include selecting, by the computing system based at least in part on the estimated performance metric, an action. The example method can include causing, by the computing system, the action to be performed.

[0012] In the example method, the estimated performance metric can be associated with one or more outputs of the one or more first machine-learned models. In the example method, selecting the action can include selecting, based at least in part on a determination that the estimated performance metric satisfies a performance metric threshold, an action to output the one or more outputs.

[0013] In the example method, the estimated performance metric can be associated with one or more first outputs of the one or more first machine-learned models. In the example method, selecting the action can include selecting, based at least in part on a determination that the estimated performance metric does not satisfy a performance metric threshold, a corrective action.

[0014] In the example method, the corrective action can include generating one or more second outputs.

[0015] In the example method, generating the one or more second outputs can include generating the one or more second outputs using one or more second machine-learned models different from the one or more first machine-learned models.

[0016] In the example method, the action can be a machine-learned inference operation. In the example method, selecting the action can include selecting, based at least in part on the estimated performance metric, whether to use the one or more first machine-learned models to perform the machine-learned inference operation.

[0017] In the example method, the estimated performance metric can include an estimated likelihood that one or more outputs of the one or more first machine-learned models belongs to a set of one or more target classes.

[0018] In the example method, the one or more ground truth classification labels can include one or more of: one or more entailment classification labels; one or more relevance classification labels; one or more risk classification labels; one or more conciseness classification labels; one or more creativity classification labels; one or more genreclassification labels; and one or more classification labels indicative of a degree to which a first portion of the first data item complies with instructions contained in a second portion of the first data item.

[0019] The example method can include determining, by the computing system based at least in part on a number of data entries of the plurality of data entries, a number of bins of the plurality of confidence bins.

[0020] Example aspects of the present disclosure provide an example computing system that includes one or more processors and one or more example non -transitory computer-readable media storing instructions that are executable by one or more processors to cause a computing system to perform example operations. In some implementations, the example operations can include obtaining a dataset comprising a plurality of data entries. In the example operations, each data entry of the plurality of data entries can correlate a first data item of the data entry to one or more ground truth classification labels associated with the first data item. The example operations can include obtaining a plurality' of machine-learned confidence values comprising, for each data entry of the plurality of data entries, one or more respective machine-learned confidence values indicative of a confidence that the first data item of the data entry belongs to one or more respective classes associated with the one or more ground truth classification labels. The example operations can include assigning, based at least in part on the plurality of machine-learned confidence values, each data entry of the plurality of data entries to a corresponding confidence bin of a plurality of confidence bins. The example operations can include determining, based at least in part on the plurality of confidence bins and the one or more ground truth classification labels of the plurality of data entries, a count matrix indicative of a plurality of ground truth classification label counts. In the example operations, each count of the plurality of ground truth classification label counts can be associated respectively with a corresponding confidence bin of the plurality of confidence bins and a corresponding class of a plurality of possible classes. The example operations can include determining, based at least in part on the count matrix, a probability matrix correlating each respective confidence bin of the plurality of confidence bins with one or more corresponding classification probabilities associated with the respective confidence bin.

[0021] In the example computing system, determining the probability' matrix can include Poisson factorization.

[0022] In the example computing system, the example operations can include determining, based at least in part on the probability’ matrix, an estimated performance metricassociated with one or more first machine-learned models. The example operations can include selecting, based at least in part on the estimated performance metric, an action. The example operations can include causing the action to be performed.

[0023] Example aspects of the present disclosure provide one or more example non-transitory computer-readable media storing instructions that are executable by one or more processors to cause a computing system to perform example operations. In some implementations, the example operations can include obtaining a dataset comprising a plurality of data entries. In the example operations, each data entry of the plurality of data entries can correlate a first data item of the data entry to one or more ground truth classification labels associated with the first data item. The example operations can include obtaining a plurality of machine-learned confidence values comprising, for each data entry of the plurality of data entries, one or more respective machine-learned confidence values indicative of a confidence that the first data item of the data entry belongs to one or more respective classes associated with the one or more ground truth classification labels. The example operations can include assigning, based at least in part on the plurality of machine-learned confidence values, each data entry of the plurality of data entries to a corresponding confidence bin of a plurality of confidence bins. The example operations can include determining, based at least in part on the plurality7of confidence bins and the one or more ground truth classification labels of the plurality of data entries, a count matrix indicative of a plurality of ground truth classification label counts. In the example operations, each count of the plurality of ground truth classification label counts can be associated respectively with a corresponding confidence bin of the plurality of confidence bins and a corresponding class of a plurality of possible classes. The example operations can include determining, based at least in part on the count matrix, a probability matrix correlating each respective confidence bin of the plurality of confidence bins with one or more corresponding classification probabilities associated with the respective confidence bin.

[0024] Other example aspects of the present disclosure are directed to other systems, methods, apparatuses, tangible non-transitory computer-readable media, and devices for performing functions described herein. These and other features, aspects, and advantages of various implementations will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate implementations of the present disclosure and, together with the description, help explain the related principles.BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a block diagram of an example system for generating a probability¬ matrix according to example implementations of aspects of the present disclosure;

[0026] Figure 2 is a block diagram of an example system for generating an intensity matrix according to example implementations of aspects of the present disclosure;

[0027] Figure 3 is a block diagram of an example system for generating a weighted count matrix from partially labeled data according to example implementations of some aspects of the present disclosure;

[0028] Figure 4 is a block diagram of an example system for selecting an action based on an estimated performance metric according to example implementations of aspects of the present disclosure;

[0029] Figure 5 is a block diagram of an example system for modifying one or more machine-learned outputs based on a performance metric according to example implementations of aspects of the present disclosure;

[0030] Figure 6 is a block diagram of an example system for generating a plurality of probability matrices associated with a plurality of subsets of a dataset according to example implementations of aspects of the present disclosure;

[0031] Figure 7 is a flow chart diagram of an example method for generating a probability matrix according to example implementations of aspects of the present disclosure;

[0032] Figure 8 is a flow chart diagram illustrating an example method for training a machine-learned model according to example implementations of aspects of the present disclosure;

[0033] Figure 9 is a block diagram of an example processing flow for using machine-learned model(s) to process input(s) to generate output(s) according to example implementations of aspects of the present disclosure;

[0034] Figure 10 is a block diagram of an example sequence processing model according to example implementations of aspects of the present disclosure;

[0035] Figure 11 is a block diagram of an example technique for populating an example input sequence for processing by a sequence processing model according to example implementations of aspects of the present disclosure;

[0036] Figure 12 is a block diagram of an example model development platform according to example implementations of aspects of the present disclosure;

[0037] Figure 13 is a block diagram of an example training workflow for training a machine-learned model according to example implementations of aspects of the present disclosure;

[0038] Figure 14 is a block diagram of an inference system for operating one or more machine-learned model(s) to perform inference according to example implementations of aspects of the present disclosure;

[0039] Figure 15 is a block diagram of an example networked computing system according to example implementations of aspects of the present disclosure;

[0040] Figure 16 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure; and

[0041] Figure 17 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure.DETAILED DESCRIPTION

[0042] Generally, the present disclosure is directed to systems and methods for evaluating and using machine-learned classifiers, such as machine-learned “critic” classifiers configured to evaluate an output of another machine-learned model (e.g.. language model, etc.). More particularly, systems and methods according to some aspects of the present disclosure can use classification labels contained in labeled data (e.g., human-labeled data, machine-labeled, etc.) to evaluate the accuracy of a machine-learned classifier model, and can use the accuracy data in a variety of ways. For example, a computing system can obtain labeled data; obtain machine-learned classification confidence values associated with the labeled data; and generate, based on the labeled data and the confidence values, a probability matrix correlating a plurality of confidence value ranges with corresponding class probabilities associated with the confidence value ranges.

[0043] In some instances, generating a probability matrix (e.g., intensity matrix, etc.) can include sorting the labeled data into a plurality of confidence bins and generating the probability matrix based on the confidence bins. For example, a labeled data example can include a data item to be classified (e.g., image, video, audio, text data, data pair, etc.) and a corresponding classification label. Continuing the example, confidence value data can include one or more machine-learned classification confidence values for each data item generated. Continuing the example, the data items to be classified can be sorted into confidence bins based on the corresponding confidence values, and a probability matrix can be generated based on the confidence bins.

[0044] In some instances, generating a probability matrix based on confidence bins can include generating a count matrix, and generating a probability matrix based on the count matrix. A count matrix can be a matrix indicating, for each confidence bin, how many data items in each confidence bin have classification labels belonging to a particular class. As a non-limiting illustrative example, a machine-learned classifier can include an entailment classifier indicating whether a first statement (e.g., a statement included in a machine-learned summary, etc.) is entailed by one or more second statements (e.g., a related statement in a document being summarized, etc ). Continuing the example, a plurality of statement pairs can be provided to the entailment classifier, which can generate a plurality of confidence values indicating a confidence that a first input is entailed or not entailed by the second input.Continuing the example, the plurality of statement pairs can be sorted into confidence bins based on the confidence values. For each confidence bin, a computing system can count the number of statement pairs in the bin that have been labeled by one or more human experts as “entailed,” and the number of statement pairs in the same bin that have been labeled by the human expert(s) as “not entailed.”

[0045] Generating a probability matrix based on a count matrix can include, for example, parametrizing a probability matrix and determining one or more parameters of the parametrized probability matrix. In some instances, determining the parameter(s) can include determining a set of parameters based at least in part on a likelihood of randomly sampling the count matrix given the set of parameters. For example, in some instances, determining the parameter(s) can include determining a set of parameters that maximize a likelihood of randomly sampling the count matrix. Further details of some example methods (e.g., fixed-point iteration, Poisson factorization, etc.) for parametrizing a probability matrix and determining one or more parameters for the parametrized probability matrix are provided below with respect to the Figures.

[0046] In some instances, labeled data can include human-labeled data, machine-labeled data, or a combination of both human-labeled and machine-labeled data. For example, in some instances, machine-labeled data can include a plurality of classifier labels generated by a different machine-learned model from the machine-learned model being evaluated, such as a higher-powered model (e g., higher parameter count, etc ), a specialized model associated with a particular classification task, or another model. In some instances, labeled data from different sources (e.g., human-labeled and machine-labeled, machine-labeled from a first machine-learned model and second machine-learned model, etc.) can be assigned weights, and a weighted count matrix can be generated based on the assignedweights. As a non-limiting illustrative example, in some instances, human-labeled data can be assigned a higher weight than machine-labeled data. For example, in some example experiments according to some aspects of the present disclosure, good results were achieved weighting machine-generated labels with a weight of 0.0005 and human-generated labels with a weight of 1.

[0047] In some instances, a machine-learned classifier being evaluated can include a machine-learned ' critic ' configured to evaluate one or more aspects of a machine-learned output of another machine-learned model (e.g., generative language model, image generation model, video generation model, etc.). As a non-limiting illustrative example, an entailment critic can analyze machine-generated summary outputs to determine whether the summary is entailed by the content being summarized. Other example critics for evaluating a language model output can include one or more of a conciseness critic classifying whether an output (or output portion) is concise; a relevance critic classify ing whether an output is relevant to an input (e.g., user input); a risk critic indicating whether an output is associated with one or more risks (e.g., whether an output describes an action for which safety precautions should be suggested, etc.); a style, genre, or creativity critic (e.g., whether an image output corresponds to a “cartoon” image style, etc ); an instruction compliance critic classifying whether an output complies with one or more instructions of an input; a syntax critic evaluating whether an output complies with one or more rules (e.g., syntax rules, function or variable name validity rules, style rules, dependency rules, etc.) associated with a computer programming language; or other machine-learned critic.

[0048] In some instances, a dataset can be divided into subsets, and a plurality of subset-specific probability matrices can be generated. As a non-limiting illustrative example, an entailment dataset can be divided into a plurality of domain-specific subsets, such as subsets associated with one or more scientific domains (e.g., physics, chemistry, geology, quantum computing, etc.) or other domain types (e.g., literary or artistic domains, geographical domains, natural language domains such as English-language or Frenchlanguage, input data type domains such as video or text, etc.), and an entailment classifier can be evaluated for each of the subsets.

[0049] In some instances, a computing system can use probability matrices to perform various actions. For example, a computing system can obtain (e.g., receive, generate, etc.) a confidence value associated with a first action (e.g., machine-learned inference action, etc.); determine, based on the confidence value and the probability matrix, a performance metric (e.g., estimated probability that a machine-generated output belongs to a particular class,etc.); and select, based on the performance metric, an action to perform. For example, in some instances, a computing system can determine a performance metric associated with the machine-learned classifier that generated the confidence values, and can select whether or not to use the machine-learned classifier in one or more actions based on the performance metric. For example, in some instances, a plurality' of performance metrics can be generated for a plurality of machine-learned classifiers, and a machine-learned classifier (e.g., critic, autorater, entailment classifier, etc.) can be selected for use based on the plurality- of performance metrics.

[0050] As another example, in some instances, the computing system can compare a performance metric or confidence value to a threshold (e.g., entailment probability threshold, etc.), and select an action based on the comparison. As a non-limiting illustrative example, if a machine-learned inference output does not satisfy one or more qualify thresholds (e.g., entailment thresholds, relevance thresholds, style thresholds, risk thresholds, etc.), then a computing system can perform one or more corrective actions, such as editing or regenerating the output (e.g., using another machine-learned model such as a higher-powered but more computationally expensive model compared to a first model that performed the first action, etc.). As another example, in some instances, if a machine-learned output (e.g., computer code output generated by a machine-learned agent configured to use application programming interface (API) calls, etc.) satisfies one or more thresholds (e.g., syntax validity probability threshold, instruction compliance probability threshold, etc.), a computing system can cause an appropriate action to be performed (e.g., providing a machine-generated API call to an API; outputting another machine-learned inference output to a user; etc.).

[0051] Systems and methods according to some aspects of the present disclosure can provide a variety of technical effects and benefits, such as improvements to computing technology (e.g., machine learning technology ). For example, in some instances, systems and methods according to some aspects of the present disclosure can provide more accurate (e.g., lower root mean square error (RMSE), etc.) evaluations of machine-learned classifier performance metrics and classification outputs compared to some alternative implementations. As another example, in some instances, systems and methods according to some aspects of the present disclosure can provide performance estimations of a given accuracy (e.g., RMSE, etc.) at reduced cost (e.g., computational cost, data collection cost, etc.) compared to some alternative implementations. Additionally, in some instances, systems and methods according to some aspects of the present disclosure can reduce a cost ofperforming some other actions, or can improve a technical performance (e.g., accuracy, etc.) of performing such other actions.

[0052] In some instances, systems and methods according to some aspects of the present disclosure can provide more accurate (e.g., lower root mean square error (RMSE), etc.) evaluations of machine-learned classifier performance metrics or classification outputs compared to some alternative implementations. For example, in some example experiments according to aspects of the present disclosure, systems and methods according to some aspects of the present disclosure outperformed some alternative systems and methods when evaluating classifier performance using only a small number of ground truth labels. In the example experiments, a large labeled dataset with a large number of labelled data items was obtained. A computing system was given access to only a small subset of the labels of the labeled dataset. Based on the small subset of labels, machine-learned classifiers were evaluated by systems and methods according to some aspects of the present disclosure, and by some alternative methods to generate a plurality of evaluation metrics. The true performance of the machine-learned classifiers was then evaluated using all of the labels of the labeled dataset. In the example experiments, systems and methods according to aspects of the present disclosure evaluated the classifiers more accurately (e.g., according to an RSME metric, etc.) than alternative methods using the same number of ground truth labels (e.g., Monte Carlo methods, variance reduction methods based on control variates, etc.).Additionally, in an example analysis according to aspects of the present disclosure, some example systems and methods described herein had an improved asymptotic root mean square error compared to an alternative Monte Carlo estimation approach.

[0053] In some instances, systems and methods according to some aspects of the present disclosure can provide performance metrics of a given accuracy (e.g., RMSE, etc.) at reduced cost (e.g., computational cost, data collection cost, etc.) compared to some alternative implementations. For example, it will be understood that an accuracy of an estimated performance metric can be increased by increasing a number of datapoints (e.g., number of human-labelled data examples, etc.) used to generate the estimated performance metric. Therefore, systems and methods that can provide higher accuracy using a given amount of data can be adapted to provide a given amount of accuracy using a reduced amount of labeled data compared to some alternative disclosures. By reducing an amount of labeled data needed to generate performance metrics of a given accuracy, systems and methods according to some aspects of the present disclosure can reduce a cost (e.g., computational cost, labor cost, etc.) of generating such performance metrics. For example, in someinstances, collecting and processing additional data can incur various costs, such as computational costs (e.g., storage cost, memory cost, processor usage cost, etc.) and other costs (e.g., time cost or labor cost of obtaining human-generated labels, etc.).Advantageously, systems and methods according to aspects of the present disclosure can reduce such costs, thereby improving the functioning of the computing system and machine learning technology compared to some alternative implementations.

[0054] In some instances, systems and methods according to some aspects of the present disclosure can reduce a cost of performing some other actions, or can improve a technical performance (e.g., accuracy, etc.) of performing such other actions. For example, in some instances, performance metrics generated according to some aspects of the present disclosure can be used to select one or more actions, such as selecting which of a plurality of machine-learned models (e.g., critics, etc.) to use, selecting whether to edit or regenerate a machine-learned output (e.g., language model output, multimodal generative output, etc.) that does not satisfy one or more rules or goals (e.g., entailment, relevance, style, safety, etc.), selecting whether to execute one or more computer-readable instructions (e.g., API calls, etc.) associated with a machine-learned output, or the like. Continuing the example, in some instances, more accurately estimating a performance metric can reduce a computational cost of generating a satisfactory machine-learned inference output by reducing a rate of “false positive’' decisions to edit or regenerate a response that is already satisfactory. Continuing the example, in some instances, more accurately estimating a performance metric can increase an accuracy of a final output (e.g., output provided to a user, computer code output provided to an API, etc.) by reducing a rate of “false negative’’ decisions in which an unsatisfactory output is accepted as satisfactory. Additionally, in the case of machine-generated computer code (e.g., API calls, etc ), reducing a rate of “false negatives” can reduce a number of times that unsatisfactory computer code is executed by a computing system, thereby saving computational resources and reducing a computational cost of machine-learned agentic action. In this manner, for instance, systems and methods according to some aspects of the present disclosure can improve the functioning of the computing system itself, and can improve machine learning technology (e.g., machine-learned agent technology, language generation technology, image or video generation technology, etc.).

[0055] Further details of some example experiments according to aspects of the present disclosure, and some example technical effects and benefits provided by some systems and methods according to some aspects of the present disclosure, are provided inAppendix A, which is incorporated by reference herein in its entirety and forms a part of this disclosure.

[0056] Various example implementations are described herein with respect to the accompanying Figures.

[0057] Figure 1 is a block diagram of an example system for generating a probability matrix according to example implementations of aspects of the present disclosure. A computing system 102 can obtain labeled data 104 and one or more machine-learned confidence values 106. Based on the labeled data 104 and the one or more machine-learned confidence values 106, the computing system 102 can determine a probability matrix 108.

[0058] A computing system 102 can be or include one or more software, firmware, or hardware components configured to generate a probability matrix based on labelled data. In some instances, the computing system 102 can be, comprise, be comprised by, or share one or more properties with a computing device or system described below with respect to Figures 15-17 (e.g., computing device 50, third-party system 80, computing device 98, computing device 99, etc.).

[0059] Labelled data 104 can generally include or otherwise represent vanous types of data. Labelled data 104 can include one type or many different types of data. Example data types for labelled data 104 can include, for example, any data type described below with respect to Figures 9-10 and inputs 2, such as language data (e.g., natural language data such as text or speech data, programming language data, etc.), sequence data (e.g.. language sequence, time series, etc ), image data, audio data, video data, or another data type.

[0060] In some instances, labeled data can include data items paired with corresponding labels (e.g., ground truth labels assigned by human experts or other trusted sources, estimated labels provided by a machine-learned model, etc.). Labels can represent, for example, a correct or desired output (e.g., machine-learned classification output, etc.) or a correct or desired classification of a given data item. For example, in some instances, a label can include a ground truth label (e.g., human-generated ground truth label, etc.) indicative of a ground truth classification indicating a class to which the data item belongs. In some instances, a label can represent a correct or desired output (e.g., machine-learned classification output, etc.) for a machine-learned model presented with an input comprising the data item. In some instances, a label can include a classification label indicative of one or more classes to which a data item belongs. In some instances, a label can include a classification label indicative of one or more properties or characteristics the data item has or does not have. As a non-limiting illustrative example, an entailment classification label caninclude a classification label (e.g., binary yes / no or entailed / not-entailed classification label) indicative of whether or not a first portion (e.g., first statement, summary, etc.) of the data item is entailed by a second portion (e.g., statement, paragraph, document, etc.) of the data item. In some instances, example classification labels indicative of example properties or characteristics of machine-learned inference outputs can include entailment classification labels, relevance classification labels, risk classification labels (e.g., user satisfaction risk labels, health and safety risk labels, business risk labels, legal risk labels, etc.), conciseness classification labels, creativity classification labels, style classification labels (e.g., labels indicative of a degree of compliance with one or more input style goals or preferences, etc.), genre classification labels (e.g., musical genre, movie genre, etc.), consistency labels, instruction compliance labels indicative of a degree to which a data item complies with one or more instructions, or other labels. In some instances, labeled data can include data that has been used to train one or more machine-learned models of interest (e.g., machine-learned classifier being evaluated; machine-learned generative model generating outputs to be classified; etc.), or data that has not been used to train any machine-learned models of interest.

[0061] In some instances, obtaining labeled data 104 can include one or more of: retrieving labeled data 104 from one or more non-transitory computer-readable media (e.g., hard disk or solid state storage device; random access memory (RAM); etc ), receiving labeled data 104 via a communication channel (e.g., from another computing device; from a user; etc.), generating labeled data 104 (e.g., according to a synthetic data generation process; according to a machine-learned labeling process; etc.), collecting or curating labeled data 104, or other means of obtaining. More generally, as used herein, “obtaining"’ data can refer to any method of obtaining data, including but not limited to retrieving, receiving, generating, or collecting.

[0062] A machine-learned confidence value 106 can generally include or otherwise represent various types of data. A machine-learned confidence value 106 can include one type or many different types of data. Example data types for a machine-learned confidence value 106 can include, for example, numerical data (e.g., real -numbered or floating-point numerical values, percentage values, etc.).

[0063] In some instances, a machine-learned confidence value 106 can include an output of a final layer or intermediate layer of a machine-learned model (e.g., machine-learned classifier model, etc.) to be evaluated. For example, in some instances, a machine-learned confidence value 106 can include an intermediate layer output (e.g., softmax output,logit output, etc.) indicative of a probability associated with a classification, such as an estimated probability that an input belongs to a particular class, or another probability (e.g., temperature-based sampling probability, etc.). In some instances, each data item of the labeled data can be associated with a plurality of machine-learned confidence values 106, such as a plurality of machine-learned confidence values 106 respectively associated with a plurality of possible classes (e.g., two classes for a binary classifier, more than two classes, etc.). In some instances, a plurality of confidence values 106 associated with a data item can include a plurality of values that sum to 1.0 or 100 percent (e.g., X and (1 - X) confidence values of a binary classifier, where X is a real number between zero and 1; X, Y, and (1 - X -Y) confidence values of a ternary classifier, where Y is a real number between zero and (1 -X); etc.). Further details of an example system for generating machine-learned confidence values 106 are provided below with respect to Figure 2.

[0064] A probability matrix 108 can include, for example, a matrix (e.g., onedimensional vector matrix, two-dimensional matrix, multi-dimensional tensor matrix, etc.; intensity matrix, etc.) correlating machine-learned confidence values 106 with corresponding label probabilities (e.g., classification probabilities, etc.). For example, in some instances, a probability matrix 108 can include a plurality of data entries each correlating a range of confidence values to one or more corresponding classification probabilities for that confidence value range. As a non-limiting illustrative example, in the case of a binary-classification, an example probability matrix 108 can include a plurality of data entries each correlating a range of confidence values to one or more of: a corresponding probability that a data item having a machine-learned confidence value 106 within the range of confidence values will belong to a first class of two binary classes; and a corresponding probability that a data item having a machine-learned confidence value 106 within the range of confidence values will belong to a second class of two binary classes. As another non-limiting example, in the case of a multi-class classification with more than two possible classes, a probability matrix 108 can include a multi-dimensional matrix correlating sets of confidence value ranges (e.g., X-Y pairs of confidence value ranges for a ternary classifier, where X is a first confidence value range associated with a first class and Y is a second confidence value range associated with a second class of the ternary classifier, etc.), or a one-dimensional matrix correlating ranges of a single-class confidence value or combined confidence value (e.g., sum of a set of confidence values, etc.) to corresponding class probabilities. Other examples are possible.

[0065] In some instances, generating a probability matrix 108 can include parametrizing the probability matrix 108 and determining one or more parameters of the parametrized probability matrix. In some instances, determining one or more parameters of a parametrized probability matrix can include determining based at least in part on a relationship between the one or more parameters and a probability of randomly sampling a set of labels of the labeled data 104 given the probability matrix. For example, in some instances, determining one or more parameters of a parametrized probability matrix can include determining, given a parametrized model (e g., linear model, sigmoidal or logistic model, polynomial model, Poisson model, etc.) of a relationship between one or more confidence values and one or more true classification probabilities, a set of parameters that increases (e.g.. maximizes or nearly maximizes; increases relative to a random or median set of parameters, etc.) a likelihood of randomly sampling the labeled data 104 labels given the probability matrix. In some instances, determining one or more parameters of the parametrized probability matrix can include determining based at least in part on an obj ective function (e g., maximum likelihood objective, expectation maximization objective, test function described herein with respect to Figure 2, loss function, reward function, etc.).

[0066] In some instances, generating a probability matrix 108 can include sorting the labeled data 104 into confidence bins, and generating a probability matrix 108 based at least in part on the confidence bins. Further details of an example system for generating a probability matrix 108 based on confidence bins are provided below with respect to Figure 2.

[0067] Figure 2 is a block diagram of an example system for generating an intensity matrix according to example implementations of aspects of the present disclosure. A computing device can obtain labeled data 204, which can include one or more data items 204a and one or more ground truth labels 204b. A machine-learned classifier 210 can generate, based on the data item(s) 204a, one or more machine-learned confidence values 106 indicative of a confidence that each data item 204a belongs to a corresponding class associated with the confidence value. Based on the machine-learned confidence values 106 and the ground truth labels 204b, a binning component 212 can assign each data item 204a to a corresponding confidence bin 214 of a plurality of confidence bins 214. Based on the confidence bins 214, a counting system 216 can generate a count matrix 218 comprising a plurality of count values, each count value being indicative of a number of ground truth labels 204b of a corresponding class within a corresponding confidence bin 214 associated with the count value. Based on the count matrix 218. a Poisson factorization system 220 can generate an intensity’ matrix 208.

[0068] In some instances, labelled data 204 can be, comprise, be comprised by, or otherwise share one or more properties with labelled data 104. For example, in some instances, labelled data 204 can have any property described herein with respect to labelled data 104, and vice versa. In some instances, labelled data can include a plurality of data entries, each data entry comprising one or more data items 204a and one or more ground truth labels 204b.

[0069] In some instances, a data item 204a can include a data item to be classified. A data item 204a can include one type or many types of data. Example data types for a data item 204a can include, for example, any data type described below with respect to Figures 9-10 and inputs 2, such as language data (e.g., natural language data such as text or speech data, programming language data. etc.), sequence data (e.g., language sequence, time series, etc.), image data, audio data, video data, or another data type. In some instances, a data item 204a can include or not include tuples (e.g., pairs, sets, groups, etc.) of related data. As a nonlimiting illustrative example, in some instances, a data item 204a for evaluating an entailment machine-learned classifier 210 can include a pair comprising a first entry (e.g., document, statement, paragraph, etc.) and a second entry (e.g., document summary, statement, etc.), wherein the entailment classifier is configured to classify whether the second entry is entailed by the first entry. As another example, in some instances, a data item 204a for evaluating a machine-learned classifier 210 configured to classify an output value’s compliance with input instructions can include a pair comprising a first entry and a second entry, and a corresponding ground truth label 204b can indicate whether the second entry complies with instructions included in the first entry. Other examples are possible.

[0070] In some instances, a ground truth label 204b can include a human-generated label (e.g., expert-generated label for scientific data, crowdsourced label, etc.) or a machinegenerated label. Example machine-generated labels can include, for example, a label generated by a machine learning system (e.g., machine-generated label from a trusted classifier, such as a machine-learned classifier having a higher computational cost than a machine-learned classifier 210 being evaluated, etc.), a label generated by a synthetic data generation system, or other machine-generated label. For example, in some instances, labeled data 204 can include or not include synthetic data generated by a computing system (e.g., non-machine-leaming system, etc.) in a manner that allows the computing system to know the label with high confidence (e.g., with certainty, etc.).

[0071] In some instances, an intensity matrix 208 can be, comprise, be comprised by, or otherwise share one or more properties with a probability matrix 108. For example, insome instances, an intensity matrix 208 can have any property described herein with respect to a probability matrix 108, and vice versa. In some instances, an intensity matrix 208 can include a plurality of data entries correlating a plurality of confidence bins 214 with one or more class probabilities associated with each respective confidence bin 214.

[0072] The machine-learned classifier(s) 210 can include one or more machine-learned models. The machine-learned classifier(s) 210 can include binary classifier(s) or multi-class classifier(s) having more than two available classes. The machine-learned classifier(s) 210 can include various model architectures, such as various neural network model architectures or other classifier architectures. An example model architecture for a machine-learned classifier(s) 210 can include an architecture configured to receive one or more inputs (e.g., natural language input, text input, image input, audio input, video input, multimodal input, etc.) and generate a classification output based on the input(s). In some instances, a classification output can include a class label; one or more confidence values 106 indicative of a confidence associated with one or more corresponding class labels or indicative of an estimated likelihood that the input(s) belong to one or more corresponding classes; or other data indicative of a class label (e.g., natural language output indicative of a class label, etc.) or confidence value. In some instances, a machine-learned classifier 210 can include a sequence processing model architecture, such as a configured to receive an input sequence and generate a classification label based on the input sequence. In some instances, a machine-learned classifier 210 can include a model architecture having an attention mechanism (e g., self-attention). In some instances, a machine-learned classifier 210 can be a pre-trained model (e.g., pretrained using large-scale unsupervised learning). In some instances, a machine-learned classifier 210 can be fine-tuned over one or more fine-tuning datasets, such as a fine-tuning dataset associated with one or more specialized classification tasks.

[0073] Generating a machine-learned confidence value 106 can include, for example, providing a data item 204a to a machine-learned classifier 210; and receiving, from the machine-learned classifier 210, one or more machine-learned confidence values 106 indicative of a confidence associated with one or more classes.

[0074] A binning component 212 can be or include one or more software, firmware, or hardware components configured to sort a plurality of machine-learned confidence values 106 or data items 204 into confidence bins 214. In some instances, the binning component 212 can be, comprise, be comprised by, or share one or more properties with a computingdevice or system described below with respect to Figures 15-17 (e.g., computing device 50, third-party system 80, computing device 98. computing device 99, etc.).

[0075] Confidence bins 214 can include, for example, one or more data structures (e.g., database, table, collection, file, folder, object, struct, etc.) correlating a plurality of data items 204a to a plurality of confidence bins 214, each confidence bin 214 associated with a range of confidence values 106 generated by the machine-learned classifier(s) 210 based on the data items 204a. In some instances, generating confidence bins 214 can include, for example, selecting a number of bins to generate, such as about 1000, about 500, about 200, about 100, about 50, about 20, about 10, or another number of bins. As used herein, the terms “about"’ or “approximately” in conjunction with a numerical value refer to within 20 percent of the numerical value. In some instances, a number of bins to generate can be selected based at least in part on data indicative of an amount of variance associated with a given number of bins, such as data indicative of a number of data items 204a for which ground truth labels 204b are available; data indicative of a variance of confidence values 106; or other relevant data.

[0076] In some instances, generating a plurality of confidence bins 214 can further include defining, based on the number of bins selected and based on one or more minimum and maximum confidence values, a plurality of confidence value 106 ranges respectively associated with the plurality of confidence bins 214. In some instances, a minimum confidence value can be zero and a maximum confidence value can be one. In some instances, generating a plurality of confidence bins 214 can include defining a plurality of equal-width (i.e., evenly spaced, etc.) bins whose union forms the entire range between a minimum confidence value 106 and a maximum confidence value 106. However, other spacings are possible (e.g.. equal width on a log scale; unequal width based on a metric of sparsity, such as equal number of data items 204a per bin; or any other spacing). In some instances, a machine-learned classifier 210 can be a binary classifier and a set of confidence value ranges can be one-dimensional (e.g., based on confidence values 106 associated with only a first class, wherein confidence values associated with a second class are inherently equal to one minus the first-class confidence values, etc.). In some instances, a machine-learned classifier 210 can be associated with more than two classes and a set of confidence value ranges can be multidimensional (e.g., first-class confidence values ranging from 0 to 1; second-class confidence values ranging from 0 to one minus first-class confidence value; and so on) if sufficient data is available, or can be one-dimensional (e.g., based only on confidence values associated with a first class, based only on a sum of a first set ofconfidence values, etc.), which can be equivalent to treating a multi-class classifier as a binary classifier.

[0077] In some instances, generating a plurality of confidence bins 214 can include, for each machine-learned confidence value 106 of a plurality of confidence values 106, assigning a data item 204a from which the machine-learned confidence value 106 was generated to a confidence bin 214 having a confidence value range that includes the machine-learned confidence value.

[0078] A counting system 216 can be or include one or more software, firmware, or hardware components configured to generate a count matrix 218 based on confidence bins 214 and labels 204b associated with the confidence bins. In some instances, the counting system 216 can be. comprise, be comprised by, or share one or more properties with a computing device or system described below with respect to Figures 15-17 (e.g., computing device 50, third-party system 80, computing device 98, computing device 99, etc.). In some instances, a computing device associated with the counting sy stem 216 can be the same as or different from a computing device associated with a binning system 212.

[0079] A count matrix 218 can include, for example, a data structure (e.g., vector, matrix, tensor, etc.) correlating a plurality of confidence bins to a plurality of count values indicative of the number of ground truth labels 204b that are associated with each class. As a non-limiting illustrative example, a count matrix 218 associated with a binary classifier and a set of 100 confidence bins 214 can include a 2 x 100 matrix comprising, for each confidence bin 214, a first count value indicative of a number of data items 204a in the confidence bin 214 belonging to a first class according to a ground truth label 204b, and a second count value indicative of a number of data items 204a in the confidence bin 214 that belong to a second class according to a corresponding ground truth label 204b. Similar (e.g., same, etc.) count matrices 218 can be generated for any number of classes or any number of confidence bins 214.

[0080] A Poisson factorization system 220 can be or include one or more software, firmware, or hardware components configured to generate an intensity matrix 208 based at least in part on a count matrix 218. In some instances, a Poisson factorization system 220 can be, comprise, be comprised by, or share one or more properties with a computing device or system described below with respect to Figures 15-17 (e.g., computing device 50, third-party system 80, computing device 98, computing device 99, etc.). In some instances, a computing device associated with a Poisson factorization system 220 can be the same as or different from a computing device associated with a binning system 212 or counting system 216.

[0081] In some instances, determining an intensity matrix 208 based on a count matrix 218 can include Poisson factorization. Poisson factorization can include, for example, determining one or more parameters of a parametrized intensity matrix 208 based at least in part on a relationship between the one or more parameters and a probability of randomly sampling the count matrix 218 given the intensity matrix 208. In some instances, parametrizing an intensity matrix 208 and determining one or more parameters of the intensity matrix 208 can include one or more methods described below with respect to the Example Discussion of Model-Based Predictive Performance Estimation (e.g., methods described below with respect to “Model-based distribution estimation using Poisson factorization,” such as the non-limiting illustrative example “Algorithm 1 MBAPPE” method, etc.).Example Discussion of Model-Based Predictive Performance Estimation

[0082] In some instances, evaluating a performance of a classifier can include evaluating the performance of a classifier < >(y|%) where a confidence value ŷ ∈ {0, 1} can include a confidence value 106 generated by the classifier < > based on an input x, which can include a data item 204a input to a machine-learned classifier 210. In some instances, evaluating the performance can be achieved by comparing <p against a “true” rating distribution p(y\x). where y can include a ground truth label 204b, under the true input distribution p(x). such as by computing the following expectation:

[0084] where t = t( y, y) is a test function determined by or otherwise associated with a metric of interest, such as accuracy, false positive rate, F-score, or other relevant metric.

[0085] For example, the accuracy a of a classifier cp can be written

[0087] wherein t = ta( y, y) = [y = y] and [•] denotes an indicator function, that is 1 when its argument is true and 0 otherwise.

[0088] Beyond accuracy, for binary classification one can also report other metrics such as false positive rate (FPR),

[0090] Other metrics, such as specificity (i.e. Pr{P = 1|F = 1}) or sensitivity (i.e. Pr[P = 0|Y = 0}) can be defined analogously. For many such metrics, one can estimate theentries of the expected confusion matrix that can be expressed as expectations of form Ep,φ{ Ŷ = a [Y = b] }. for a, b E {0, 1 }, with true positives (TP) are a = 1, b = 1, false positives (FP) are a = 1, b = 0.

[0091] Performance evaluation by computing expectations of form in (1) can include computing integrals of form

[0093] In some instances, the test function t can be given by:

[0094] Φt(x,y) = t(y,ŷ) φ(ŷ|x) (5)

[0095] wherein < >(y|x) can correspond to a classifier.

[0096] An example discussion of model-free estimation that reduces variance compared to a standard Monte Carlo estimate is included below. However, some example model-free estimations can require an input x to be discrete valued and have enough data to be able to estimate the conditional distributions p(Y = y\X = x). which may not be the case for some (e.g., most) practical applications. In contrast, some example model-based methods, such as example models described herein with respect to Figures 1-6, can provide a practical estimator applicable to continuous-valued x and low-data regimes.

[0097] Low-dimensional representation:

[0098] If there exists a low-dimensional representation z such that:

[0100] then one can rewrite the expectation of interest as:)

[0104] where p(z, y) can denote

[0105] p(z, y) = p(z|x)p(x, y)dx (11)

[0106] When z is a deterministic function of x, then p(z\x) = p(z\x. y) and thus (11) corresponds to the joint distribution of z, y.

[0107] This result shows that, in some instances, in order to evaluate the expectation, it can be sufficient to estimate p(y, z), rather than the higher dimensional p( y, x). For neural network classifiers, an example low-dimensional z representation can include an intermediate layer activation. For example, in some instances, such as when <p is a binary classifier, z can include a confidence output (e.g., softmax output, logit output, etc.) of the classifier model (e.g., machine-learned classifier 210, etc.). In some instances, this can turn performanceevaluation to calculating a one dimensional integral. In some examples, a histogram approximation to p(y, z) can be performed (e.g., as described below and in Appendix, which is incorporated by reference herein in its entirety and forms a part of this disclosure).

[0108] Dimension reduction. Consider the case where cp(y\x) is associated with a Bernoulli distribution Ber(y; fe (x)), where fe (x) is a soft-max output of a neural network. We have a binary classifier and we can write

[0109] φ(ŷ|x)= fθ(x)y(1 -fθ(x)1-y).(12)

[0110] In that case, a sufficient statistic for φ can be given by z = fθ(x), and

[0114] In such instances, one can write (10) as

[0116] In some instances, for obtaining a crisp decision, a threshold parameter can be used to truncate the confidence outputs z of the classifier cp. In such instances, one can redefine

[0117] φ̄(ŷ|z) = φ̄τ(ŷ|z) = [z ≥ τ]ŷ(1 - [z ≥ τ])1-ŷ(16)

[0118] Binning and matrix estimation. In some instances, p(y, z) can be a probability distribution on {0, 1} x [0. 1], In some instances, estimating a performance of a classifier (e.g., generating a probability matrix 108, etc.) can include performing a count matrix estimation problem by discretizing [0, 1], For example, in some instances, discretizing [0, 1] can include defining C non-overlapping bins, { / c}c=o and [0, 1] = Uc=o lc- In some instances, z(c) e lccan denote a center of a corresponding bin lc. In such instances. (15) can be rewritten as

[0122] Where p(y, c) = ∫Ip(y, z)dz. Similarly, for a threshold classifier < >Tone would have

[0124] In some instances, a number of confidence bins 214 can be selected based at least in part on a target precision, a number of ground truth labels 204b available, or other relevant data. For example, in some instances, approximations (19) or (20) can become more precise as a number of confidence bins 214 increases and a width of each confidence bin 214 decreases, at least if a large enough number of ground truth labels 204b are available. This can have the effect of discretizing a continuous variable x, thereby solving one difficulty associated with using an example model-free estimation strategy. However, in some instances, approximations according to (19) or (20) may not solve difficulties associated with limited availability of ground truth labels 204b or other data (e.g., label data). In some instances, an example model-based approach according to example aspects of the present disclosure (e.g., one or more example model-based approaches described herein with respect to Figures 1-6, etc.) can be used to address this issue.

[0125] Model-based distribution estimation using Poisson factorization

[0126] As p(y, c) is a discrete distribution, the sufficient statistics can simply be the bin count matrices 218, which can be denoted as S( y, c). In some instances, Poisson factorization or another statistical model for count matrices can be applied to one or more count matrices 218 to generate one or more intensity matrices 208. For example, in some instances, a Poisson factorization model for a table of counts can be configured to estimate a latent intensity matrix Λθ.

[0128] where As can be a function of some parameters 6 and PO(A) can denote a Poisson distribution with intensity parameter A. In some instances, such parameters 6 can be determined based at least in part on a Poisson likelihood associated with the parameters 9 given an observed count matrix 218. For example, in some instances, aPoisson likelihood can be maximized (or nearly maximized, etc.). For example, by maximizing the Poisson likelihood, given an observed count matrix S, one can find the optimal parameters as 9* and set the associated model-based estimate as

[0130] where A* = Ae.

[0131] In some instances, instead of using directly the joint distribution estimateone can usethe following estimate:

[0133] = ΣyΣŷΣct(y, ŷ) φ̄(c, ŷ)pθ*(y|c)(22)

[0134] where < >(c, y) = p(c) (y|z(c)). In some instances, a similar model-free estimation strategy can be performed by using a non-parametric conditional distribution estimate instead of the parametric estimate shown here. An example discussion of some example model-free estimation strategies is provided below, just before the discussion of Figure 7 below.

[0135] In some instances, various parametric models can be used, which may differ in how a latent intensity matrix can be approximated.

[0136] In a first example model, a latent intensity matrix A can be modeled as a low rank matrix. Such techniques can be useful, for example, in prediction when a total number of ground truth labels 204b are low and a count matrix 218 is consequently sparse:

[0138] where B can be a fixed nonnegative matrix and 9 can be a parameter matrix to be estimated. In some instances, B(:, j) can include a collection of fixed kernels, and 0(:, y) can be the associated parameters for each class y. In some instances, the parameters 0(:, y) can be or correspond to mixture weights conditional on each individual class.

[0139] For example, in some instances, an intensity matrix Aecan include a matrix product of a fixed basis matrix B and a matrix of coefficient values 9.

[0140] In some instances, one or more parameters 9 can be determined based at least in part on a likelihood of a count matrix 218 (e.g., matrix 218 of observed counts S, etc.). For example, in some example implementations, a maximum likelihood estimate of 9 can beobtained. For example, in some instances, given observed counts S and a fixed basis matrix B, a maximum likelihood estimate of 9 can be obtained by an Expectation-Maximization algorithm that can be simplified as the following fixed point iteration:

[0141] θ(τ)(j,y) = θ(τ-1)(j,y)(24)

[0142] given a suitable initialization θ(0).

[0143] For example, in some instances, fixed point iteration can include initializing, by a computing system, an initial value 0 °) for a matrix of coefficient values; and. for each of a plurality of iterations, determining, by the computing system, an updated value θ(τ)(j, y) for the matrix of coefficient values, wherein the updated value (j, y) is determined based at least in part on a current value θ(τ-1)(j, y) of the matrix of coefficient values.

[0144] In some instances, B can be fixed, such that the resulting objective is log concave (in 0) and the EM iterations will converge to a global maximum 0*. In some instances, a model-based estimation can be performed using one or more machine-generated labels (e.g., in addition to human-generated labels, etc.), or can be performed without using any machine-generated labels. For example, in some instances, a count matrix 218 (e.g., count matrix 218 having a plurality of count values S(y, c), etc.) can include counts of only human-generated labels, and a model that uses such a count matrix 218 without reference to any additional machine-generated labels can be used (e.g., a model according to (24), etc.). As another example, in some instances, a count matrix 218 can include a count matrix 218 (e.g., weighted count matrix, etc.) that includes one or more count values associated with one or more machine-generated labels, and a probability matrix 208 can be generated from the count matrix 218 (e.g., according to (24), etc.). Further details of an example system for generating a weighted count matrix comprising human-generated and machine-generated labels are provided below with respect to Figure 3.

[0145] In some instances, an example method for model-based performance estimation without using any machine-generated labels can include Algorithm 1 below, wherein no optional "pseudo-labels'’ are obtained, and wherein p is zero. In some instances, an example method for model-based performance estimation with some machine-generated labels can include Algorithm 1 below, wherein at least some machine-generated “pseudolabels” are obtained, and wherein p > 0 (e.g., as described below with respect to Figure 3, etc.).

[0146] Algorithm 1 MBAPPE

[0147] Require:

[0148] <p Classifier to be evaluated, p: Trusted rater, q: Autorater

[0149] Inputs:

[0150] • Ŷi~ p(y|x = xi) for i = 1,..., n, (Trusted labels)

[0151] • Ŷi~ φ(y|x = xi) for i = 1,..., n + m, (Labels from the classifier under evaluation)

[0152] • Zi= φ(y = 1|x = xi) for i = 1,..., n + m, (Confidence outputs of φ)

[0153] • Ŷi~ q(y|x = xi) for i = 1,..., n + m, (Optional - Pseudo-labels obtained from the autorater q)

[0154] Algorithm parameters:

[0155] • t: test function

[0156] • C: number of bins (e.g., C = 100, etc.)

[0157] • B: C x j Kernel matrix (e.g., / = 30, etc.)

[0158] • p: fraction factor for autorater counts (e.g., p E {0, 0.0005}, etc.)

[0159] 1. Construct C bins { / c}c=o such that [0, 1]

[0160] by forming empirical (c + l) / C-quantiles of all available inputs {Zi } for i = 1... n + m

[0161] 2. Form a C X 2 count matrix S from n trusted labels

[0162] S(c,y) = Σni=1[Ŷi= y][Zi∈ Ic]

[0163] 3. Form a C * 2 count matrix A from n + m pseudo-labels

[0164] A(c,y) = Σi[Ŷi= y][Zi∈ Ic]

[0165] 4. Optimize an intensity matrixgiven S and A using a fixed point iteration

[0166] Λθ= Bθ, θ* = argmaxθlog PO(S + ρA; Λθ).

[0167] 5. Estimate the true labeling distribution p(y|c), using A0»,

[0168] 6. Form a C x 2 count matrix < D from classifier labels

[0170] Return Estimate En=.c'.y' yt(tyy)‘i’(c<y)p(y|c)

[0171] As another example, in some instances, another example model for using machine-generated labels (e.g., instead of or in addition to human-generated labels, such as in instances when a number of available human-generated labels are limited) can include “correcting the autorater.” For example, in some instances, a plurality- of data items 204a can be provided to a machine-learned model (e.g., critic, autorater, etc.), and the machine-learnedmodel can output a plurality of corresponding classification labels (e.g., “pseudo-labels” of a machine-learned model that may not be completely trusted; ground truth labels 204b; etc.) respectively associated with the plurality of data items 204a. In some instances, the empirical j oint probability of an example appearing in bin c and being classified as y by a machine-learned model (e.g., critic, autorater, classifier model different from the machine-learned classifier 210 being evaluated, etc.) can be denoted A(y. c). In some instances, a correction term can be learned, such as according to:

[0173] This model can learn, for example, a rank-1 multiplicative correction 0102Tto the observed autorater counts. In some instances, the free parameters of this model can be learned based at least in part on a likelihood associated with the parameters and a given count matrix 218, such as using an expectation maximization (EM) algorithm.

[0174] Such a model is likely to give good results especially when the autorater distribution is close to the to the expert labeling distribution such that a rank-1 correction can be accurate. However, in practice, it may be hard to verify this condition in some instances. Thus, in some instances, other methods can be used that take advantage of machine-learned labels and do not require a rank-1 correction to be sufficient. Further details of an example implementation of model-based performance estimation that uses machine-learned labels without necessarily requiring a rank-1 correction to be sufficient are provided below with respect to Figure 3.

[0175] Figure 3 is a block diagram of an example system for generating a weighted count matrix 318 from partially labeled data 304 according to example implementations of some aspects of the present disclosure. Partially labeled data 304 can include a plurality of data items 304a and a plurality of ground truth labels 304b associated with a subset of the plurality of data items 304a. The plurality of data items 304a can be provided to one or more machine-learned classifiers 210, which can generate machine-learned confidence values 106 associated with the data items 304a. One or more unlabeled data items of the plurality of data items 304a can be provided to one or more machine-learned labelers 322, which can generate one or more machine-learned labels 324 for each unlabeled data item 304a. Based on the machine-learned confidence values 106, the plurality of data items 304a can be assigned to a plurality of confidence bins 214. Based on the plurality of confidence bins 214, the plurality of ground truth labels 204b, and the plurality of machine-learned labels 324, a counting system 216 can generate a weighted count matrix 318.

[0176] In some instances, partially labelled data 304 can be, comprise, be comprised by, or otherwise share one or more properties with labelled data 104, 204. For example, in some instances, partially labelled data 304 can have any property described herein with respect to labelled data 104, 204, and vice versa. In some instances, partially labelled data 304 can include a plurality of labeled data entries and a plurality of unlabeled data entries, each labeled data entry comprising one or more data items 204a and one or more ground truth labels 204b, and each unlabeled data entry comprising one or more data items 304a.

[0177] In some instances, a weighted count matrix 318 can be, comprise, be comprised by, or otherwise share one or more properties with a count matrix 218. For example, in some instances, a weighted count matrix 318 can have any property described herein with respect to a count matrix 218, and vice versa. In some instances, a weighted count matrix 218 can include a plurality of weighted counts associated with a plurality of confidence bins. For example, in some instances, each label 204b, 324 of a plurality of labels can be associated with a weighting value, and a weighted count can be equal to a weighted sum of the weighting values. As a non-limiting illustrative example, in instances where a machine-learned classifier 210 is a binary classifier and the confidence bins 214 comprise 100 evenly spaced bins, a weighed count matrix 318 can include 200 weighted count values associated respectively with a corresponding confidence bin and corresponding class of the binary classifier. Continuing the example, each weighted count value can be proportional to (e.g.. equal to, etc.) a combination (e.g., sum, etc.) of weighting values of all labels 324, 204b associated with the corresponding confidence bin 214 and class associated with the weighted count value.

[0178] The machine-learned label er(s) 322 can include one or more machine-learned models. The machine-learned labeler(s) 322 can include binary classifier(s) or multi-class classifier(s) having more than two available classes. The machine-learned labeler(s) 322 can include various model architectures, such as various neural network model architectures or other classifier architectures. An example model architecture for a machine-learned labeler(s) 322 can include an architecture configured to receive one or more inputs (e.g., natural language input, text input, image input, audio input, video input, multimodal input, etc.) and generate a classification label output based on the input(s). In some instances, a machine-learned labeler 322 can include a sequence processing model architecture, such as a configured to receive an input sequence and generate a classification label based on the input sequence. In some instances, a machine-learned labeler 322 can include a model architecture having an attention mechanism (e.g., self-attention). In some instances, a machine-learnedlabeler 322 can be a pre-trained model (e.g., pretrained using large-scale unsupervised learning). In some instances, a machine-learned labeler 322 can be fine-tuned over one or more fine-tuning datasets, such as a fine-tuning dataset associated with one or more specialized classification tasks.

[0179] In some instances, a machine-learned labeler 322 can include a machine-learned model that is different from a machine-learned classifier 210. In some instances, a machine-learned labeler 322 can include a model having a higher computational cost than the machine-learned classifier 210 being evaluated. In some instances, a machine-learned labeler 322 can include a model having one or more known performance characteristics, such as a model known to generate labels with high accuracy compared to one or more other models; a model known to generate labels or confidence values that are not strongly correlated with confidence values 106 of a machine-learned classifier 210 being evaluated; or other known performance characteristics.

[0180] In some instances, a machine-learned label 324 can be, comprise, be comprised by, or otherwise share one or more properties with a ground truth label 204b. For example, in some instances, a machine-learned label 322 can have any property described herein with respect to a ground truth label 204b, and vice versa.

[0181] In some instances, performance estimation using a weighted count matrix 318 can be performed in any manner described herein with respect to Figure 1. In some instances, performance estimation using a weighted count matrix 318 can include generating a modified observed count matrix according to

[0182] Sρ(y, c) = S(y, c) + ρA(y, c),

[0183] where S y, c) is a count matrix 218 comprising counts of human-generated labels; A(y, c) is a count matrix 218 comprising counts of machine-generated labels (e g., labels generated by a machine-learned labeler 322 different from a machine-learned classifier 210 being evaluated, etc.); and p is a real number, such as a real number less than or equal to 0.0005 or the like. As a non-limiting illustrative example, in some instances, performance estimation using a weighted count matrix 318 can include performing Algorithm 1 described above with respect to Figure 1, wherein p is a positive real number less than 1 (e.g., much less than 1, such as 0.0005, etc.).

[0184] Figure 4 is a block diagram of an example system for selecting an action based on an estimated performance metric according to example implementations of aspects of the present disclosure. A computing system 102 can obtain a probability matrix 108 and candetermine, based on the probability matrix 108, a performance metric 426. Based on the performance metric 426. an action selector 428 can generate an action selection 430 output. In some instances, a computing system can cause an action corresponding to the action selection 430 to be performed.

[0185] A performance metric 426 can generally include or otherwise represent various types of data. A performance metric 426 can include one type or many different types of data. Example data types for a performance metric 426 can include, for example, numerical data such as a numerical rating, score, metric or the like. In some instances, an example performance metric can be a numerical value indicative of an estimated statistical value (e.g., accuracy, false positive rate, false negative rate, F score, sensitivity, specificity, likelihood, etc.). In some instances, an example performance metric can include a performance metric defined by an objective function (e.g., loss function, reward function, test function such as test function described above with respect to Figure 2, etc.).

[0186] In some instances, a performance metric 426 can include a metric indicative of a performance of a machine-learned classifier 210, or a metric indicative of a performance of another machine-learned model evaluated by the machine-learned classifier 210. For example, in some instances, a performance metric 426 indicative of a performance of a machine-learned classifier 210 can include one or more IE{t} metrics (e.g., IE{t} metrics described above with respect to Figure 2, etc.) indicative of an expected value of a test function (e.g., accuracy, false positive rate, etc.) across a plurality of data items 204a (e.g.. across a probability distribution of data items 204a, etc.). In some instances, an action selection 430 based on a performance of a machine-learned classifier 210 (e.g., machine-learned critic, autorater, etc.) can include selecting whether or not to use the machine-learned classifier 210 for one or more actions, such as for evaluating one or more aspects of another machine-learned model’s output(s) (e.g., according to methods described below with respect to Figure 5, etc.). For example, in some instances, a computing system can generate, for each of a plurality of machine-learned classifiers 210, a respective probability matrix 108; generate, for each of the plurality of machine-learned classifiers 210 based on the respective probability matrix 108, a respective performance metric 426 indicative of an expected performance (e.g., accuracy, false positive rate, F score, or other objective function, etc.) of the machine-learned classifier 210; and select, based on the performance metrics 426, a machine-learned classifier 210 for use in one or more systems or methods (e.g., for use in evaluating outputs of another machine-learned model, such as according to methods described herein with respect to Figure 5, etc ). As another example, in some instances, acomputing system can generate, for each of a plurality of data subsets (e.g., domain-specific subsets, subsets described herein with respect to Figure 6. etc.), a respective performance metric 426 for a machine-learned classifier 210 with respect to the data subset; and, responsive to determining that one or more of the performance metrics is below a performance metric threshold, the computing system can fine-tune one or more machine-learned classifiers 210 based at least in part on the data subset, or perform another action (e.g., omitting the data subset or a data domain associated with the data subset from a set of domains for which a system comprising the machine-learned classifier 210 is used, etc ).

[0187] In some instances, a performance metric 426 can include a metric indicative of an expected performance (e.g., likelihood of belonging to a target class, likelihood of conforming to a target property or characteristic, etc.) of a machine-learned model other than a machine-learned classifier 210 used to generate the probability matrix 108. For example, in some instances, a machine-learned classifier 210 can generate a confidence value 106 based on one or more input values (e.g., inference outputs of another machine-learned model such as a generative language model or generative multimodal model, etc.; or another value of interest), and a computing system can generate, based on the confidence value 106 and a probability matrix 108 associated with the machine-learned classifier 210, a performance metric 426 indicative of a likelihood that the one or more input values belong to a target classification. Based on the performance metric 426 indicative of the likelihood, an action selector 428 can select an action 430, such as an action to accept (e.g., output, act on, etc.) or reject (e g., edit, regenerate, etc.) the one or more input values; an action to execute one or more instructions (e.g., application programming interface (API) calls generated by a machine-learned agent, etc.) associated with the one or more input values; a machine-learned inference operation; or another action.

[0188] An action selector 428 can be or include one or more software, firmware, or hardware components configured to select one or more actions 430 based at least in part on a performance metric 426. In some instances, the action selector 428 can be, comprise, be comprised by, or share one or more properties with a computing device or system described below with respect to Figures 15-17 (e.g., computing device 50, third-party system 80, computing device 98, computing device 99, etc.). In some instances, a computing device associated with the action selector 428 can be the same as or different from a computing device associated with the computing system 102.

[0189] An action selection 430 can generally include or otherwise represent various types of data. An action selection 430 can include one type or many different types of data.Example data types for an action selection 430 can include, for example, text data (e.g., computer code data, natural language data, etc.), numerical data (e.g., numerical action selection identifier, etc.), or other datatype. In some instances, an action selection 430 can include computer code that, when executed, causes one or more software, firmware, or hardware components to perform the selected actions, such as one or more application programming interface (API) calls associated with an API; computer-executable network requests (e.g., hypertext transfer protocol requests, etc.); or other computer-executable instructions. In some instances, an action selection 430 can include other data indicative of a selected action, such as an action name, action identifier (e.g., numerical identifier, token, etc.), action description, action parameters, or the like. For example, in some instances, a computing system 102 can obtain (e.g., generate, receive, etc.) an action selection 430 that does not comprise executable computer code, and can execute one or more operations based on the action selection 430. As a non-limiting illustrative example, a computing system 102 can include glue code configured to receive an action selection 430 comprising an action name; and perform, based on the action selection 430. an action identified by the action name.

[0190] In some instances, causing an action to be performed can include mapping an action selection 430 to corresponding executable code (e.g., corresponding application programming interface (API) call, etc.) and executing the executable code. In some instances, mapping an action selection 430 to executable code can include retrieving corresponding executable code (e g., corresponding API call, etc.) from a data structure (e.g., database, table, row, column, file, object, etc.) based at least in part in part on the action selection 430 (e.g., based on an action identifier, etc.). In some instances, mapping the action selection 430 to executable code can include passing the action selection 430 to glue code (e.g., glue code comprising one or more compiler, interpreter, or parser functions, etc.) configured to map action selections 430 to executable actions.

[0191] In some instances, an action selection 430 can include or be based on an action selection output generated by a machine-learned agent. For example, in some instances, a machine-learned agent (e.g., machine-learned model different from the machine-learned classifier 210, etc.) can generate an action selection output; a computing system 102 can generate a performance metric 426 indicative of a likelihood that the action selection output meets one or more goals (e.g., correct syntax or other grammar; compliance with one or more input instructions on which the action selection output was based; etc.): and an action selector 428 can select, based at least in part on the performance metric, an action selection 430. Insome instances, an example action selection 430 can include executing, responsive to determining that the performance metric 426 satisfies a performance metric threshold, an action corresponding to the agent-generated action selection output (e.g., providing an agentgenerated API call to a corresponding API, etc.). In some instances, an example action selection 430 can include performing, responsive to determining that a performance metric 426 does not satisfy the performance metric threshold, one or more corrective actions.Example corrective actions can include editing or regenerating an action selection output (e.g., using a machine-learned model that is the same as or different from a model used to generate the action selection output being evaluated); outputting an error message or alert (e.g., to a user; to a maintenance technician, etc.); selecting a new machine-learned agent to generate a new action selection output; or other corrective action.

[0192] In some instances, performing an action associated with an action selection 430 can include using one or more tools, such as one or more API tools having an application programming interface (API) that can be called to invoke the API tool. An API tool can include, for example, any tool (e.g., hardware tool, software tool, firmware tool, etc.) that can be accessed via an API. For example, in some instances, an API tool can include software (e g., application, operating system, etc.) installed on a device (e.g., server device, client device, etc.) running the action selector 428; software available via a network (e.g., internet); hardware devices (e.g., internet-connected hardware devices, Bluetooth-connected hardware devices, etc.); etc. In some instances, a hardware API tool can include a hardware tool connected (e.g., via a network, via a wireless or wired connection, etc.) to a device running the action selector 428. In some instances, an API tool can include a navigation API (e.g., map-related API, global positioning system-related API, etc.); a communication API (e.g., API associated with making phone calls, emails, or text messages such as SMS or MMS; APIs associated with communication applications, such as messaging applications, social media communication applications, etc.); a scheduling API (e.g., calendar, alarm, automated task scheduling, etc.); media player API (e.g., video player such as YouTube, audio player, etc.); shopping API; payment API such as Google Pay; mobile banking API; travel-related API (e.g., flight booking, hotel booking, etc.); or API of an application installed on a mobile device. In some instances, an API tool can include a hardware device such as a Bluetooth-connected lock, gate opener, garage door opener, etc.; an internet-connected doorbell or surveillance camera device; smart home device such as smart TV, smart appliance, lighting devices, thermostats, etc.; or any other API-accessible hardware tool.

[0193] In some instances, a tool can include a tool configured to perform one or more operations that may include a machine learning component, such as image processing (e.g., machine-learned image processing; image generation, visual question answering, visual identification such as facial recognition, image captioning, etc.), audio processing (e.g., machine-learned image processing; audio generation such as speech or music generation, speech-to-text or text-to-speech processing, audio identification such as voice identification, etc.), video processing (e.g., video generation, etc.), sequence processing (e.g., natural language sequence generation, natural language translation, question answering, computer code generation, etc.), robotics (e.g., robotic systems comprising machine-learned agent(s) configured to control physical manipulation tools, etc ), mobile digital assistants (e.g., smart phone assistants configured to perform communication actions, navigation actions, calendar actions, smart appliance control actions, etc.) or other action type. Other tool types and other action types are possible.

[0194] Figure 5 is a block diagram of an example system for modifying one or more machine-learned outputs based on a performance metric according to example implementations of aspects of the present disclosure. One or more generative machine-learned models 532 can generate one or more inference outputs 534. One or more machine-learned classifiers 510 can generate, based on the inference output(s) 534, one or more machine-learned confidence values 106. Based on the machine-learned confidence value(s) 106 and a probability matrix 108, a computing system 102 can generate one or more performance metrics 426. A threshold comparison system 528 can compare the performance metric 426 to one or more performance metric thresholds. If the performance metric 426 satisfies the performance metric threshold(s) (indicated in Figure 5 as a Success 530a), the inference output(s) 534 can be output (e.g., to a user 536). If the performance metric 426 does not satisfy a performance metric threshold (indicated in Figure 5 as a Failure 530b), the threshold comparison system 528 can select one or more editing or regeneration actions 538, and one or more machine-learned models 540 can generate one or more edited or regenerated outputs 542 responsive to the Failure 530b determination. In some instances, the process can be repeated, providing the edited or regenerated outputs 542 to the one or more machine-learned classifiers 510; generating one or more performance metrics 426 for the edited / regenerated outputs 542; and selecting, based on the performance metric 426, an action to be performed.

[0195] In some instances, a machine-learned classifier 510 can be, comprise, be comprised by, or otherwise share one or more properties with a machine-learned classifier210. For example, in some instances, a machine-learned classifier 510 can have any property- described herein with respect to a machine-learned classifier 210. and vice versa.

[0196] In some instances, a threshold comparison system 528 can be, comprise, be comprised by, or otherwise share one or more properties with an action selector 428. For example, in some instances, a threshold comparison 528 can have any property described herein with respect to an action selector 428, and vice versa.

[0197] In some instances, a threshold comparison 528 can include a comparison between one or more performance metrics 426 and one or more corresponding performance metric thresholds. For example, in some instances, an example performance metric 426 can include an estimated probability that an inference output 534 belongs to one or more target classes or satisfies one or more goals (e.g., entailment, consistency, etc.). Continuing the example, in some instances, the estimated probability can be compared to one or more probability thresholds. If an estimated probability of satisfying one or more goals is greater than a probability- threshold, then the inference output 534 can be identified as successful, and an appropriate success 530a action can be performed (e g., outputting the inference output 534 to a user 536, etc.). If an estimated probability of satisfying one or more goals is less than a probability threshold, then the inference output 534 can be identified as unsuccessful, and an appropriate failure 530b action can be performed (e.g., edit / regeneration action 538, etc.).

[0198] In some instances, a success 530a or failure 530b action selection can be, comprise, be comprised by, or otherwise share one or more properties with an action selection 430. For example, in some instances, a success 530a or failure 530b action selection can have any property described herein with respect to an action selection 430, and vice versa.

[0199] The generative machine-learned model(s) 532 can include one or more machine-learned models. The generative machine-learned model(s) 532 can include various model architectures, such as various neural network model architectures. An example model architecture for a generative machine-learned model 532 can include a sequence processing model architecture (e.g., a transformer model). For example, the generative machine-learned model(s) 532 can be configured to receive an input sequence and generate an output sequence. For instance, generative machine-learned model(s) 532 can be configured to generate an output sequence where elements of the output sequence are predicted based on the elements of the input sequence. In some instances, a generative machine-learned model 532 can include a model architecture having an attention mechanism (e.g., self-attention). In some instances, a generative machine-learned model 532 can be a pre-trained model (e.g.,pretrained using large-scale unsupervised learning). In some instances, the generative machine-learned model 532 can be fine-tuned over one or more fine-tuning datasets, such as a fine-tuning dataset associated with one or more specialized generation tasks.

[0200] An inference output 534 can generally include or otherwise represent various types of data. An inference output 534 can include one type or many different types of data. Example data types for an inference output 534 an include, for example, any data type described below with respect to Figures 9-10 and inputs 2, such as language data (e.g., natural language data such as text or speech data, programming language data, etc.), sequence data (e.g., language sequence, time series, etc.), image data, audio data, video data, or another data type.

[0201] In some instances, generating an inference output 534 can include providing, by a computing system, an input context (e.g., natural language, text, audio, video, image, or multimodal input context, etc.) to a generative machine-learned model 532; generating, by the generative machine-learned model 532, the inference output 534 based on the input context; and receiving, by the computing system from the machine-learned model 532. the inference output 534.

[0202] A user 536 can include, for example, a person, user account, organization, or other entity to which inference outputs 534 are provided. In some instances, providing inference output(s) 534 to a user 536 can include transmitting the inference output(s) 534 over a communication channel (e.g.. network such as the internet, etc.) to a computing device associated with the user; displaying the inference output(s) 534 to a user via a display device or graphical user interface; or other method for providing inference output(s) 534 to a user.

[0203] In some instances, a computing system 202 can be, comprise, be comprised by, or otherwise share one or more properties with a computing system 102. For example, in some instances, a computing system 202 can have any property described herein with respect to a computing system 102, and vice versa.

[0204] In some instances, an edit / regeneration action 538 can be, comprise, be comprised by, or otherwise share one or more properties with an action 430. For example, in some instances, an edit / regeneration action 538 can have any property described herein with respect to an action 430, and vice versa. In some instances, an edit / regeneration action 538 can include causing, by a computing system, a machine-learned model 540 to edit or regenerate an inference output 534. Regenerating an inference output 534 can include, for example, generating a new inference output 542 based on an input context that is similar to (e.g., same as, etc.) an input context used to generate the inference output 534. Editing aninference output 534 can include, for example, generating an edited inference output based at least in part on an input context that is similar to (e.g.. same as, etc.) an input context used to generate the inference output 534. In some instances, generating an edited inference output can include generating a new standalone output (e.g., based on an input context comprising the inference output 534 to be edited, etc.), generating one or more edit values (e.g., deletions, additions, modifications, etc.), selecting one or more edit actions, interacting with an editing interface (e.g., application programming interface, graphical user interface, etc.), or other editing method. In some instances, an input context used to generate an edited / regenerated output 542 can include a first input context that was used to generate an inference output 534 to be edited or regenerated. In some instances, an input context used to generate an edited / regenerated output 542 can further comprise additional data, such as one or more instructions to edit or regenerate the inference output 534; an inference output 534 to be edited; data indicative of a failure 530b, such as data (e.g., natural language data, binary data, embedding data, etc.) indicating a type of the failure (e.g., lack of entailment, improper length, improper style, failure to follow instructions, etc.).

[0205] In some instances, a machine-learned model 540 can be, comprise, be comprised by, or otherwise share one or more properties with a generative machine-learned model 532. For example, in some instances, a machine-learned model 540 can have any property described herein with respect to a generative machine-learned model 532, and vice versa. In some instances, a machine-learned model 540 can be the same machine-learned model or a different machine-learned model compared to the generative machine-learned model 532. For example, in some instances, a computing system can, responsive to detecting a failure 530b, select an edit / regeneration action 538 to cause the generative machine-learned model 532 to edit its own output or to generate a new output based on a similar (e.g.. same) input context used to generate an inference output 534 associated with the failure 530b. As another example, in some instances, a computing system can generate a first inference output 534 using a lightweight (e.g., low computational cost, low parameter count, low architectural complexity, low latency, etc.) machine-learned model and causing, responsive to determining that a performance metric 426 associated with the first inference output 534 does not satisfy a performance threshold 528, a higher-powered (e.g., more accurate or otherwise higher-quality; higher computational cost, parameter count, architectural complexity, or latency; etc.) machine-learned model 540 to edit the inference output 534 or to generate a new inference output 542 based on an input context similar to (e.g., same as, etc.) an input context used to generate the inference output 534.

[0206] In some instances, an edited / regenerated output 542 can be, comprise, be comprised by, or otherwise share one or more properties with an inference output 534. For example, in some instances, an edited / regenerated output 542 can have any property described herein with respect to an inference output 534, and vice versa. In some instances, an edited / regenerated output 542 produced in a first iteration can evaluated in a second iteration according to any method described herein with respect to an inference output 534. In some instances, a plurality of iterations can be performed until a performance metric 426 indicative of success 530a is obtained.

[0207] Figure 6 is a block diagram of an example system for generating a plurality of probability matrices 608 associated with a plurality of subsets of a dataset 604. A data splitter 644 can obtain labelled data 604a and can generate, based on the labelled data 604a, a plurality of labelled subsets 604b, 604c, 604d. Based on the plurality of labelled subsets 604 and a plurality of machine-learned confidence values 106 associated with the labelled subsets, a computing system 102 can generate a plurality of probability matrices 608b, 608c, 608d.

[0208] In some instances, labeled data 604 can be, comprise, be comprised by, or otherwise share one or more properties with labeled data 104. For example, in some instances, labeled data 604 can have any property described herein with respect to labeled data 104 and vice versa.

[0209] In some instances, a set 604a of labeled data can include all data of a dataset, such as all data of a particular document corpus; all data available for training or evaluating a machine-learned classifier 210; or other data set. In some instances, labeled subsets 604b, 604c, 604d can include strict subsets of the set of labeled data 604a. In some instances, labeled subsets 604b, 604c, 604d can include overlapping or non-overlapping subsets.

[0210] In some instances, determining subsets 604b, c, d of a dataset 604a can include splitting the dataset 604a according to one or more categories or properties, such as splitting a dataset by domain (e.g., scientific domain, etc.), topic, genre, style, datatype (e.g., video, audio, etc.), or other category or property. As a non-limiting illustrative example, in some instances, a dataset 604a can be split by knowledge domain, and a machine-learned entailment classifier 210 can be evaluated using each of a plurality of data subsets 604b, c, d associated with each of a plurality of knowledge domains. Continuing the example, in some instances, the entailment classifier can be selected to be used for one or more actions 430 (e.g., for evaluating inference outputs 534 according to methods described with respect to Figure 5, etc.) associated with knowledge domains for which the entailment classifierperforms well, and a different entailment classifier can be used for one or more actions 430 associated with knowledge domains for which the entailment classifier performs poorly. For example, in some instances, one or more fine-tuned entailment classifiers can be trained for specific knowledge domains in which a general-purpose entailment classifier performs poorly (e.g., below a threshold performance metric such as below a threshold IE {t} value, etc.).

[0211] A data splitter 644 can be or include one or more software, firmware, or hardware components configured to divide a labeled dataset 604a into subsets 604b, c, d. In some instances, the data splitter 644 can be, comprise, be comprised by, or share one or more properties with a computing device or system described below with respect to Figures 15-17 (e.g., computing device 50, third-party system 80, computing device 98, computing device 99. etc.). In some instances, a computing device associated with the data splitter 644 can be the same as or different from a computing device associated with the computing system 102.

[0212] In some instances, a probability matrix 608b, c, d can be, comprise, be comprised by, or otherwise share one or more properties with a probability matrix 108. For example, in some instances, a probability matrix 608b, c, d can have any property described herein with respect to a probability matrix 108, and vice versa. In some instances, each probability matrix 608 can include a subset-specific probability matrix 608 generated based on a corresponding labeled subset 604b, c, d. For example, in some instances, a subsetspecific probability matrix 608 can be generated in any manner described herein with respect to Figures 1-3, wherein labeled data 104, 204 used to generate the subset-specific probability matrix 608 consists of a corresponding subset 604b, c, d. In some instances, a subset-specific probability matrix 608 can be indicative of a subset-specific (e.g., domain-specific, etc.) performance of a machine-learned classifier 210. In some instances, a subset-specific (e.g., domain-specific, etc.) performance metric 426 can be generated (e.g., according to any method described herein with respect to Figure 2 or Figures 4-5, etc.) based on a corresponding probability matrix 608, and an action selection 430 can be determined based on the performance metric 426 (e.g., according to methods described herein with respect to Figures 4-5, etc.). For example, in some instances, a computing system can receive an input directed to a generative machine-learned model; identify a category or property (e.g., domain, topic, genre, style, datatype (e.g., video, audio, etc.), or other category or property) associated with the input; and determine a performance metric based at least in part on a subset-specific probability matrix 608 determined from a labeled data subset 604b, c, d associated with the identified category or property.

[0213] In some instances, a subset-specific probability' matrix 608b can include a weighted or unweighted evaluation. As a non-limiting illustrative example, evaluating a classifier (e.g., entailment classifier, etc.) with respect to a subdomain for which labeled data is sparse can include performing a weighted evaluation, wherein count data for the subdomain being evaluated may be combined with count data from related or similar subdomains, and wherein the evaluation is weighted based at least in part on a metric of similarity between the subdomain being evaluated and the other subdomains. Continuing the example, in some instances, a weighted evaluation can include an evaluation based on a weighted count matrix 318, wherein a weight of each count value can be based at least in part on a metric of similarity between the subdomain being evaluated and the other subdomains. Continuing the example, in some instances, a weighted evaluation can include another form of weighted evaluation, such as determining a weighted combination (e.g., sum, etc.) of subdomain-specific probability matrices 608, or other weighted evaluation.Example Discussion of Example Model-Free Evaluation of Classifier Performance

[0214] Although some examples described herein are directed to model-based evaluation of classifier performance, in some instances, model-free evaluation can be performed without deviating from the scope of the present disclosure. In some instances, an example model-free estimate can leverage the ability to collect unlabeled data from the data distribution p(x). In some instances, such an estimate can be shown to admit asymptotically a lower RMSE than the Monte Carlo estimate. This can show that there is significant theoretical headroom in improving such performance estimates. This model-free estimator can work w ell when x is finite and a large amount of data is available, whereas model-based estimates provided herein (e.g., with respect to FIGS. 1-6, etc.) can work well in a low-data regime.

[0215] Setup:

[0216] We want to evaluate the performance of a classifier <p(y\x) where y E {0, 1}. In some instances, this can be achieved by comparing < > against a '“true7’ rating distribution p(y|x) under the true input distribution p(x), by computing the following expectation.

[0218] where t = t( y, y) is a test function determined by the metric we are interested in.

[0219] For example, the accuracy a of the classifier cp is

[0221] where we have chosen t = ta( y, y) = [y = y] and [ ] denotes an indicator function, that is 1 when its argument is true and 0 otherwise.

[0222] Beyond accuracy, for binary classification we can also report other metrics such as false positive rate (FPR),

[0224] Other metrics, such as specificity (i.e. Pr{P = 1|T = 1}) or sensitivity (i.e. Pr{P = 0 |T = 0}) can be defined analogously. For many such metrics, one can estimate the entries of the expected confusion matrix that can be expressed as expectations of formY = a [K = b]}, for a, b E {0, 1}, with true positives (TP) are a = 1, b = 1, false positives (FP) are a = 1, b = 0.

[0225] Performance evaluation by computing expectations of form in (1) can include computing integrals of form

[0227] where < I»t (x, y) = y t(y, y)< >( y|x) can be computed exactly. The distribution p(x, y) = p(x) p(y|x) may not be available in closed-form and one may only have access to i.i.d. samples (Xz, Yi) ~ p(x, y) where i = 1,..., n, i.e. human labelled data. One can use these samples to consider the following Monte Carlo estimate

[0229] which is unbiased and has variance

[0230] var[

[0231] The key problem with evaluating such performance metrics is that the variance of the Monte Carlo estimate (6) can be high when a limited number of samples n is available.

[0232] One approach can be to augment the number of labelled data by sampling m synthetic data using Xi ~ p(x) and then Yi ~ q y|x) for i = n + 1,..., n + m, where q( y|x) is an autorater model. However, if q(y\x) is significantly different from the human label distribution p(y|x) then this approach introduces very substantial bias as illustrated in Fig. 3 of Appendix A.

[0233] Model-Free Variance Reduction

[0234] Consider here a simple scenario where X takes values in a finite space and Y also takes values in a finite space. Assume we can obtain as many unlabeled data as we want, i.e. samples from p(x). This can be the case in some practical application of interest. Weshow here how we can leverage this information to provide a consistent estimate that admits a lower asymptotic variance and mean square error compared to the Monte Carlo estimate (6).

[0235] In some instances, the Monte Carlo estimate (5) can be rewritten as

[0237] By sampling an additional number m of unlabeled data Xi ~ p(x) for i = n + 1, n + m, we can propose instead the following estimate

[0239] In practice, one can select m » n so that practically the fluctuations of lEji ^ft] are essentially identical to those of En[t] where

[0241] Note that (7), (8) and (9) are only well defined if for all x there exists an Xi = x for i = 1,..., n. As n — > oo, these estimates will always be defined from the strong law of large numbers. The model-free estimate E„[t] has a lower asymptotic variance than En[t], Further details of an example comparison between E„[t] and En[t] are provided in Appendix A, which is incorporated by reference herein in its entirety and forms a part of this disclosure. Additionally, in some instances, the model-free estimate E.„ [t] can in some instances have a lower root mean square error (RMSE) for a given number of labeled data 104 datapoints compared to En[t], or can achieve a similar (e.g., same, etc.) RMSE using a reduced number of labeled data 104 data points (e.g., reduced number of human-generated labels, etc.) compared to [t]. Further details of some example experiments comparing En[t] and Efl[t] are provided in Appendix A.Example Methods

[0242] Figure 7 depicts a flowchart diagram of an example method for generating a probability matrix according to example embodiments of the present disclosure. Although Figure 7 depicts steps performed in a particular order for purposes of illustration and discussion, the methods of the present disclosure are not limited to the particularly illustrated order or arrangement. The various steps of example method 700 can be omitted, rearranged, combined, and / or adapted in various ways without deviating from the scope of the present disclosure.

[0243] At 702, example method 700 can include obtaining, by a computing system comprising one or more computing devices, a dataset (e.g., labeled data 104, 204; labeled subset 604; etc.) comprising a plurality of data entries, each data entry of the plurality of data entries correlating a first data item (e.g., data item 204a, etc.) of the data entry to one or more ground truth classification labels (e.g., ground truth labels 204b, etc.) associated with the first data item. In some instances, example method 700 at 702 can include using one or more systems or performing one or more activities described with respect to Figures 1-2 or 6.

[0244] At 704, example method 700 can include obtaining, by the computing system, a plurality7of machine-learned confidence values (e.g., confidence values 106, etc.) comprising, for each data entry of the plurality of data entries, one or more respective machine-learned confidence values indicative of a confidence that the first data item of the data entry belongs to one or more respective classes associated with the one or more ground truth classification labels. In some instances, example method 700 at 704 can include using one or more systems or performing one or more activities described with respect to Figure 1-2.

[0245] At 706, example method 700 can include assigning, by the computing system based at least in part on the plurality of machine-learned confidence values, each data entry of the plurality of data entries to a corresponding confidence bin of a plurality of confidence bins (e.g., confidence bins 214, etc.). In some instances, example method 700 at 706 can include using one or more systems or performing one or more activities described with respect to Figure 2.

[0246] At 708, example method 700 can include determining, by the computing system based at least in part on the plurality of confidence bins and the one or more ground truth classification labels of the plurality of data entries, a count matrix (e.g., count matrix 218, weighted count matrix 318, etc.) indicative of a plurality of ground truth classification label counts, each count of the plurality of ground truth classification label counts being associated respectively with a corresponding confidence bin of the plurality of confidence bins and a corresponding class of a plurality of possible classes. In some instances, example method 700 at 708 can include using one or more systems or performing one or more activities described with respect to Figures 2-3.

[0247] At 710, example method 700 can include determining, by the computing system based at least in part on the count matrix, a probability matrix (e.g., probability matrix 108, intensity7matrix 208, etc.) correlating each respective confidence bin of the plurality of confidence bins with one or more corresponding classification probabilities associated withthe respective confidence bin. In some instances, example method 700 at 710 can include using one or more systems or performing one or more activities described with respect to Figures 1-3 or 6.

[0248] Figure 8 depicts a flowchart of a method 800 for training one or more machine-learned models according to aspects of the present disclosure. For instance, an example machine-learned model can include a machine-learned classifier 210, 510 or other machine-learned model 532, 540.

[0249] One or more portion(s) of example method 800 can be implemented by a computing system that includes one or more computing devices such as, for example, computing systems described with reference to the other figures. Each respective portion of example method 800 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 800 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 8 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. Figure 8 is described with reference to elements / terms described with respect to other systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of example method 800 can be performed additionally, or alternatively, by other systems.

[0250] At 802, example method 800 can include obtaining a training instance. A set of training data can include a plurality of training instances divided between multiple datasets (e.g., a training dataset, a validation dataset, or testing dataset). A training instance can be labeled or unlabeled. Although referred to in example method 800 as a '‘training’’ instance, it is to be understood that runtime inferences can form training instances when a model is trained using an evaluation of the model’s performance on that runtime instance (e.g., online training / leaming). Example data types for the training instance and various tasks associated therewith are described throughout the present disclosure.

[0251] At 804, example method 800 can include processing, using one or more machine-learned models, the training instance to generate an output. The output can be directly obtained from the one or more machine-learned models or can be a downstream result of a chain of processing operations that includes an output of the one or more machine-learned models.

[0252] At 806, example method 800 can include receiving an evaluation signal associated with the output. The evaluation signal can be obtained using a loss function.Various determinations of loss can be used, such as mean squared error, likelihood loss, cross entropy loss, hinge loss, contrastive loss, or various other loss functions. The evaluation signal can be computed using known ground-truth labels (e.g., supervised learning), predicted or estimated labels (e.g., semi- or self-supervised learning), or without labels (e.g., unsupervised learning). The evaluation signal can be a reward (e.g., for reinforcement learning). The reward can be computed using a machine-learned reward model configured to generate rewards based on output(s) received. The reward can be computed using feedback data describing human feedback on the output(s).

[0253] At 808, example method 800 can include updating the machine-learned model using the evaluation signal. For example, values for parameters of the machine-learned model(s) can be learned, in some embodiments, using various training or learning techniques, such as, for example, backwards propagation. For example, the evaluation signal can be backpropagated from the output (or another source of the evaluation signal) through the machine-learned model(s) to update one or more parameters of the model(s) (e.g., based on a gradient of the evaluation signal with respect to the parameter value(s)). For example, system(s) containing one or more machine-learned models can be trained in an end-to-end manner. Gradient descent techniques can be used to iteratively update the parameters over a number of training iterations. In some implementations, performing backwards propagation of errors can include performing truncated backpropagation through time. Example method 800 can include implementing a number of generalization techniques (e.g., weight decays, dropouts, etc.) to improve the generalization capability of the models being trained.

[0254] In some implementations, example method 800 can be implemented for training a machine-learned model from an initialized state to a fully trained state (e.g., when the model exhibits a desired performance profde, such as based on accuracy, precision, recall, etc.).

[0255] In some implementations, example method 800 can be implemented for particular stages of a training procedure. For instance, in some implementations, example method 800 can be implemented for pre-training a machine-learned model. Pre-training can include, for instance, large-scale training over potentially noisy data to achieve a broad base of performance levels across a variety of tasks / data types. In some implementations, example method 800 can be implemented for fine-tuning a machine-learned model. Fine-tuning can include, for instance, smaller-scale training on higher-quality (e.g., labeled, curated, etc.)data. Fine-tuning can affect all or a portion of the parameters of a machine-learned model. For example, various portions of the machine-learned model can be "frozen" for certain training stages. For example, parameters associated with an embedding space can be '‘frozen” during fine-tuning (e.g., to retain information learned from a broader domain(s) than present in the fine-tuning dataset(s)). An example fine-tuning approach includes reinforcement learning. Reinforcement learning can be based on user feedback on model performance during use.Example Machine-Learned Models

[0256] Figure 9 is a block diagram of an example processing flow for using machine-learned model(s) 1 to process input(s) 2 to generate output(s) 3.

[0257] Machine-learned model(s) 1 can be or include one or multiple machine-learned models or model components. Example machine-learned models can include neural networks (e.g., deep neural networks). Example machine-learned models can include nonlinear models or linear models. Example machine-learned models can use other architectures in lieu of or in addition to neural networks. Example machine-learned models can include decision tree based models, support vector machines, hidden Markov models, Bayesian networks, linear regression models, k-means clustering models, etc.

[0258] Example neural networks can include feed-forward neural networks, recurrent neural networks (RNNs), including long short-term memory (LSTM) based recurrent neural networks, convolutional neural networks (CNNs), diffusion models, generative-adversarial networks, or other forms of neural netw orks. Example neural netw orks can be deep neural networks. Some example machine-learned models can leverage an attention mechanism such as self-attention. For example, some example machine-learned models can include multiheaded self-attention models.

[0259] Machine-learned model(s) 1 can include a single or multiple instances of the same model configured to operate on data from input(s) 2. Machine-learned model(s) 1 can include an ensemble of different models that can cooperatively interact to process data from input(s) 2. For example, machine-learned model(s) 1 can employ a mixture-of-experts structure. See, e.g., Zhou et al., Mixture-of-Experts with Expert Choice Routing, ARXIV:2202.09368V2 (Oct. 14, 2022).

[0260] Input(s) 2 can generally include or otherwise represent various types of data. Input(s) 2 can include one type or many different types of data. Output(s) 3 can be data of thesame type(s) or of different types of data as compared to input(s) 2. Output(s) 3 can include one type or many different t pes of data.

[0261] Example data types for input(s) 2 or output(s) 3 include natural language text data, software code data (e.g., source code, object code, machine code, or any other form of computer-readable instructions or programming languages), machine code data (e.g., binary code, assembly code, or other forms of machine-readable instructions that can be executed directly by a computer's central processing unit), assembly code data (e.g., low-level programming languages that use symbolic representations of machine code instructions to program a processing unit), chemical or biochemical data, image data, audio data, audiovisual data, haptic data, statistical data, geographical data, astronomical data, historical data, sensor data generally (e.g., digital or analog values, such as voltage or other absolute or relative level measurement values from a real or artificial input, such as from an audio sensor, light sensor, displacement sensor, etc.), and the like. Data can be raw or processed and can be in any format or schema.

[0262] In multimodal inputs 2 or outputs 3, example combinations of data types include image data and audio data, image data and natural language data, natural language data and software code data, image data and astronomical data, sensor data and chemical data, etc. It is to be understood that any combination of data types in an input 2 or an output 3 can be present.

[0263] An example input 2 can include one or multiple data types, such as the example data types noted above. An example output 3 can include one or multiple data types, such as the example data types noted above. The data type(s) of input 2 can be the same as or different from the data type(s) of output 3. It is to be understood that the example data types noted above are provided for illustrative purposes only. Data types contemplated within the scope of the present disclosure are not limited to those examples noted above.Example Machine-Learned Sequence Processing Models

[0264] Figure 10 is a block diagram of an example implementation of an example machine-learned model configured to process sequences of information. For instance, an example implementation of machine-learned model(s) 1 can include machine-learned sequence processing model(s) 4. An example system can pass input(s) 2 to sequence processing model(s) 4. Sequence processing model(s) 4 can include one or more machine-learned components. Sequence processing model(s) 4 can process the data from input(s) 2 to obtain an input sequence 5. Input sequence 5 can include one or more input elements 5-1, 5-2,..., 5-M, etc. obtained from input(s) 2. Sequence processing model 4 can process input sequence 5 using prediction layer(s) 6 to generate an output sequence 7. Output sequence 7 can include one or more output elements 7-1, 7-2,..., 7-N, etc. generated based on input sequence 5. The system can generate output(s) 3 based on output sequence 7.

[0265] Sequence processing model(s) 4 can include one or multiple machine-learned model components configured to ingest, generate, or otherwise reason over sequences of information. For example, some example sequence processing models in the text domain are referred to as “Large Language Models,’’ or LLMs. See, e.g., PaLM 2 Technical Report, GOOG E, https: / / ai.google / static / documents / palm2techreport.pdf (n.d.). Other example sequence processing models can operate in other domains, such as image domains, see, e.g., Dosovitskiy et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, ARXIV:2010.11929v2 (Jun. 3, 2021), audio domains, see, e.g., Agostinelli et al., MusicLM: Generating Music From Text, ARXlV:2301.11325vl (Jan. 26, 2023), biochemical domains, see, e.g., Jumper et al., Highly accurate protein structure prediction with AlphaFold, 596 Nature 583 (Aug. 26, 2021), by way of example. Sequence processing model(s) 4 can process one or multiple types of data simultaneously. Sequence processing model(s) 4 can include relatively large models (e.g., more parameters, computationally expensive, etc.), relatively small models (e.g., fewer parameters, computationally lightweight, etc.), or both.

[0266] In general, sequence processing model(s) 4 can obtain input sequence 5 using data from input(s) 2. For instance, input sequence 5 can include a representation of data from input(s) 2 in a format understood by sequence processing model(s) 4. One or more machine-learned components of sequence processing model(s) 4 can ingest the data from input(s) 2, parse the data into pieces compatible with the processing architectures of sequence processing model(s) 4 (e.g., via “tokenization”). and project the pieces into an input space associated with prediction layer(s) 6 (e.g., via “embedding”).

[0267] Sequence processing model(s) 4 can ingest the data from input(s) 2 and parse the data into a sequence of elements to obtain input sequence 5. For example, a portion of input data from input(s) 2 can be broken down into pieces that collectively represent the content of the portion of the input data. The pieces can provide the elements of the sequence.

[0268] Elements 5-1, 5-2,..., 5-M can represent, in some cases, building blocks for capturing or expressing meaningful information in a particular data domain. For instance, the elements can describe “atomic units” across one or more domains. For example, for textual input source(s), the elements can correspond to groups of one or more words or sub-word components, such as sets of one or more characters.

[0269] For example, elements 5-1, 5-2,..., 5-M can represent tokens obtained using a tokenizer. For instance, a tokenizer can process a given portion of an input source and output a series of tokens (e.g., corresponding to input elements 5-1, 5-2,..., 5-M) that represent the portion of the input source. Various approaches to tokenization can be used. For instance, textual input source(s) can be tokenized using a byte-pair encoding (BPE) technique. See, e.g., Kudo et al., SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing, PROCEEDINGS OF THE 2018 CONFERENCE ON EMPIRICAL METHODS IN NATURAL LANGUAGE PROCESSING (System Demonstrations), pages 66-71 (October 31-November 4, 2018), https: / / aclanthology.org / D18-2012.pdf. Image-based input source(s) can be tokenized by extracting and serializing patches from an image.

[0270] In general, arbitrary data types can be serialized and processed into input sequence 5. It is to be understood that element(s) 5-1, 5-2,..., 5-M depicted in Figure 10 can be the tokens or can be the embedded representations thereof.

[0271] Prediction layer(s) 6 can predict one or more output elements 7-1. 7-2,..., 7-N based on the input elements. Prediction layer(s) 6 can include one or more machine-learned model architectures, such as one or more layers of learned parameters that manipulate and transform the input(s) to extract higher-order meaning from, and relationships between, input element(s) 5-1, 5-2,..., 5-M. In this manner, for instance, example prediction layer(s) 6 can predict new output element(s) in view of the context provided by input sequence 5.

[0272] Prediction layer(s) 6 can evaluate associations between portions of input sequence 5 and a particular output element. These associations can inform a prediction of the likelihood that a particular output follows the input context. For example, consider the textual snippet, “The carpenter’s toolbox was small and heavy. It was full of Example prediction layer(s) 6 can identify that “It” refers back to “toolbox” by determining a relationship between the respective embeddings. Example prediction layer(s) 6 can also link “It” to the attributes of the toolbox, such as “small” and “heavy.” Based on these associations, prediction layer(s) 6 can, for instance, assign a higher probability to the word “nails” than to the word “sawdust.”

[0273] A transformer is an example architecture that can be used in prediction layer(s) 4. See, e.g., Vaswani et al., Attention Is All You Need, ARXlV:1706.03762v7 (Aug. 2, 2023). A transformer is an example of a machine-learned model architecture that uses an attention mechanism to compute associations between items within a context window. The context window can include a sequence that contains input sequence 5 and potentially one ormore output element(s) 7-1, 7-2,..., 7-N. A transformer block can include one or more attention layer(s) and one or more post-attention layer(s) (e.g., feedforward layer(s), such as a multi-layer perceptron).

[0274] Prediction layer(s) 6 can include other machine-learned model architectures in addition to or in lieu of transformer-based architectures. For example, recurrent neural networks (RNNs) and long short-term memory (LSTM) models can also be used, as well as convolutional neural networks (CNNs). In general, prediction layer(s) 6 can leverage various kinds of artificial neural networks that can understand or generate sequences of information.

[0275] Output sequence 7 can include or otherwise represent the same or different data types as input sequence 5. For instance, input sequence 5 can represent textual data, and output sequence 7 can represent textual data. Input sequence 5 can represent image, audio, or audiovisual data, and output sequence 7 can represent textual data (e.g., describing the image, audio, or audiovisual data). It is to be understood that prediction layer(s) 6, and any other interstitial model components of sequence processing model(s) 4, can be configured to receive a variety of data types in input sequence(s) 5 and output a variety of data types in output sequence(s) 7.

[0276] Output sequence 7 can have various relationships to input sequence 5. Output sequence 7 can be a continuation of input sequence 5. Output sequence 7 can be complementary to input sequence 5. Output sequence 7 can translate, transform, augment, or otherwise modify input sequence 5. Output sequence 7 can answer, evaluate, confirm, or otherwise respond to input sequence 5. Output sequence 7 can implement (or describe instructions for implementing) an instruction provided via input sequence 5.

[0277] Output sequence 7 can be generated autoregressively. For instance, for some applications, an output of one or more prediction layer(s) 6 can be passed through one or more output layers (e.g., softmax layer) to obtain a probability distribution over an output vocabulary (e.g., a textual or symbolic vocabulary) conditioned on a set of input elements in a context window. In this manner, for instance, output sequence 7 can be autoregressively generated by sampling a likely next output element, adding that element to the context window, and re-generating the probability distribution based on the updated context window, and sampling a likely next output element, and so forth.

[0278] Output sequence 7 can also be generated non-autoregressively. For instance, multiple output elements of output sequence 7 can be predicted together without explicit sequential conditioning on each other. See, e.g.. Saharia et al., Non-Autoregressive Machine Translation with Latent Alignments, ARXIV:2004.07437V3 (NOV. 16, 2020).

[0279] Output sequence 7 can include one or multiple portions or elements. In an example content generation configuration, output sequence 7 can include multiple elements corresponding to multiple portions of a generated output sequence (e.g., a textual sentence, values of a discretized waveform, computer code, etc.). In an example classification configuration, output sequence 7 can include a single element associated with a classification output. For instance, an output “vocabulary” can include a set of classes into which an input sequence is to be classified. For instance, a vision transformer block can pass latent state information to a multilayer perceptron that outputs a likely class value associated with an input image.

[0280] Figure 11 is a block diagram of an example technique for populating an example input sequence 8. Input sequence 8 can include various functional elements that form part of the model infrastructure, such as an element 8-0 obtained from a task indicator 9 that signals to any model(s) that process input sequence 8 that a particular task is being performed (e.g., to help adapt a performance of the model(s) to that particular task). Input sequence 8 can include various data elements from different data modalities. For instance, an input modality 10-1 can include one modality of data. A data-to-sequence model 11-1 can process data from input modality 10-1 to project the data into a format compatible with input sequence 8 (e.g., one or more vectors dimensioned according to the dimensions of input sequence 8) to obtain elements 8-1, 8-2, 8-3. Another input modality 10-2 can include a different modality of data. A data-to-sequence model 11-2 can project data from input modality 10-2 into a format compatible with input sequence 8 to obtain elements 8-4, 8-5, 8-6. Another input modality 10-3 can include yet another different modality of data. A data-to-sequence model 11-3 can project data from input modality 10-3 into a format compatible with input sequence 8 to obtain elements 8-7. 8-8, 8-9.

[0281] Input sequence 8 can be the same as or different from input sequence 5. Input sequence 8 can be a multimodal input sequence that contains elements that represent data from different modalities using a common dimensional representation. For instance, an embedding space can have P dimensions. Input sequence 8 can be configured to contain a plurality of elements that have P dimensions. In this manner, for instance, example implementations can facilitate information extraction and reasoning across diverse data modalities by projecting data into elements in the same embedding space for comparison, combination, or other computations therebetween.

[0282] For example, elements 8-0...., 8-9 can indicate particular locations within a multidimensional embedding space. Some elements can map to a set of discrete locations inthe embedding space. For instance, elements that correspond to discrete members of a predetermined vocabulary of tokens can map to discrete locations in the embedding space that are associated with those tokens. Other elements can be continuously distributed across the embedding space. For instance, some datatypes can be broken down into continuously defined portions (e.g., image patches) that can be described using continuously distributed locations within the embedding space.

[0283] In some implementations, the expressive power of the embedding space may not be limited to meanings associated with any particular set of tokens or other building blocks. For example, a continuous embedding space can encode a spectrum of high-order information. An individual piece of information (e.g., a token) can map to a particular point in that space: for instance, a token for the word “dog” can be projected to an embedded value that points to a particular location in the embedding space associated with canine-related information. Similarly, an image patch of an image of a dog on grass can also be projected into the embedding space. In some implementations, the projection of the image of the dog can be similar to the projection of the word “dog” while also having similarity to a projection of the word “grass,” while potentially being different from both. In some implementations, the projection of the image patch may not exactly align with any single projection of a single word. In some implementations, the projection of the image patch can align with a combination of the projections of the words “dog” and “grass.” In this manner, for instance, a high-order embedding space can encode information that can be independent of data modalities in which the information is expressed.

[0284] Task indicator 9 can include a model or model component configured to identify a task being performed and inject, into input sequence 8, an input value represented by element 8-0 that signals which task is being performed. For instance, the input value can be provided as a data type associated with an input modality and projected along with that input modality (e.g., the input value can be a textual task label that is embedded along with other textual data in the input; the input value can be a pixel-based representation of a task that is embedded along with other image data in the input; etc.). The input value can be provided as a data type that differs from or is at least independent from other input(s). For instance, the input value represented by element 8-0 can be learned within a continuous embedding space.

[0285] Input modalities 10-1, 10-2, and 10-3 can be associated with various different data types (e.g., as described above with respect to input(s) 2 and output(s) 3).

[0286] Data-to-sequence models 11-1, 11-2, and 11-3 can be the same or different from each other. Data-to-sequence models 11-1, 11-2, and 11-3 can be adapted to each respective input modality 10-1, 10-2, and 10-3. For example, a textual data-to-sequence model can subdivide a portion of input text and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-1, 8-2, 8-3, etc.). An image data-to-sequence model can subdivide an input image and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-4, 8-5, 8-6, etc.). An arbitrary datatype data-to-sequence model can subdivide an input of that arbitrary datatype and project the subdivisions into element(s) in input sequence 8 (e.g., elements 8-7, 8-8, 8-9, etc.).

[0287] Data-to-sequence models 11-1, 11-2, and 11-3 can form part of machine-learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be jointly trained with or trained independently from machine-learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be trained end-to-end with machine-learned sequence processing model(s) 4.Example Machine-Learned Model Development Platform

[0288] Figure 12 is a block diagram of an example model development platform 12 that can facilitate creation, adaptation, and refinement of example machine-learned models (e.g., machine-learned model(s) 1, sequence processing model(s) 4, etc.). Model development platform 12 can provide a number of different toolkits that developer systems can employ in the development of new or adapted machine-learned models.

[0289] Model development platform 12 can provide one or more model libraries 13 containing building blocks for new models. Model libraries 13 can include one or more pretrained foundational models 13-1, which can provide a backbone of processing power across various tasks. Model libraries 13 can include one or more pre-trained expert models 13-2, which can be focused on performance in particular domains of expertise. Model libraries 13 can include various model primitives 13-3, which can provide low-level architectures or components (optionally pre-trained), which can be assembled in various arrangements as desired.

[0290] Model development platform 12 can receive selections of various model components 14. Model development platform 12 can pass selected model components 14 to a workbench 15 that combines selected model components 14 into a development model 16.

[0291] Workbench 15 can facilitate further refinement and adaptation of development model 16 by leveraging a number of different toolkits integrated with model developmentplatform 12. For example, workbench 15 can facilitate alignment of the development model 16 with a desired performance profile on various tasks using a model alignment toolkit 17.

[0292] Model alignment toolkit 17 can provide a number of tools for causing development model 16 to generate outputs aligned with desired behavioral characteristics. Alignment can include increasing an accuracy, precision, recall, etc. of model outputs.Alignment can include enforcing output styles, schema, or other preferential characteristics of model outputs. Alignment can be general or domain-specific. For instance, a pre- trained foundational model 13-1 can begin with an initial level of performance across multiple domains. Alignment of the pre-trained foundational model 13-1 can include improving a performance in a particular domain of information or tasks (e g., even at the expense of performance in another domain of information or tasks).

[0293] Model alignment toolkit 17 can integrate one or more dataset(s) 17-1 for aligning development model 16. Curated dataset(s) 17-1 can include labeled or unlabeled training data. Dataset(s) 17-1 can be obtained from public domain datasets. Dataset(s) 17-1 can be obtained from private datasets associated with one or more developer system(s) for the alignment of bespoke machine-learned model(s) customized for private use-cases.

[0294] Pre-training pipelines 17-2 can include a machine-learned model training workflow configured to update development model 16 over large-scale, potentially noisy datasets. For example, pre-training can leverage unsupervised learning techniques (e g., denoising, etc.) to process large numbers of training instances to update model parameters from an initialized state and achieve a desired baseline performance. Pre-training pipelines 17-2 can leverage unlabeled datasets in dataset(s) 17-1 to perform pre-training. Workbench 15 can implement a pre-training pipeline 17-2 to pre-train development model 16.

[0295] Fine-tuning pipelines 17-3 can include a machine-learned model training workflow configured to refine the model parameters of development model 16 with higher-quality data. Fine-tuning pipelines 17-3 can update development model 16 by conducting supervised training with labeled dataset(s) in dataset(s) 17-1. Fine-tuning pipelines 17-3 can update development model 16 by conducting reinforcement learning using reward signals from user feedback signals. Workbench 15 can implement a fine-tuning pipeline 17-3 to finetune development model 16.

[0296] Prompt libraries 17-4 can include sets of inputs configured to induce behavior aligned with desired performance criteria. Prompt libraries 17-4 can include few-shot prompts (e.g., inputs providing examples of desired model outputs for prepending to a desiredruntime query ), chain-of-thought prompts (e.g., inputs providing step-by-step reasoning within the exemplars to facilitate thorough reasoning by the model), and the like.

[0297] Example prompts can be retrieved from an available repository of prompt libraries 17-4. Example prompts can be contributed by one or more developer systems using workbench 15.

[0298] In some implementations, pre-trained or fine-tuned models can achieve satisfactory performance without exemplars in the inputs. For instance, zero-shot prompts can include inputs that lack exemplars. Zero-shot prompts can be within a domain within a training dataset or outside of the training domain(s).

[0299] Prompt libraries 17-4 can include one or more prompt engineering tools. Prompt engineering tools can provide workflows for retrieving or learning optimized prompt values. Prompt engineering tools can facilitate directly learning prompt values (e.g., input element values) based on one or more training iterations. Workbench 15 can implement prompt engineering tools in development model 16.

[0300] Prompt libraries 17-4 can include pipelines for prompt generation. For example, inputs can be generated using development model 16 itself or other machine-learned models. In this manner, for instance, a first model can process information about a task and output a input for a second model to process in order to perform a step of the task. The second model can be the same as or different from the first model. Workbench 15 can implement prompt generation pipelines in development model 16.

[0301] Prompt libraries 17-4 can include pipelines for context injection. For instance, a performance of development model 16 on a particular task can improve if provided with additional context for performing the task. Prompt libraries 17-4 can include software components configured to identify desired context, retrieve the context from an external source (e.g., a database, a sensor, etc.), and add the context to the input prompt. Workbench 15 can implement context injection pipelines in development model 16.

[0302] Although various training examples described herein with respect to model development platform 12 refer to “‘pre-training” and “fine-tuning.” it is to be understood that model alignment toolkit 17 can generally support a wide variety of training techniques adapted for training a wide variety of machine-learned models. Example training techniques can correspond to the example training method 800 described above.

[0303] Model development platform 12 can include a model plugin toolkit 18. Model plugin toolkit 18 can include a variety of tools configured for augmenting the functionality of a machine-learned model by integrating the machine-learned model with other systems.devices, and software components. For instance, a machine-learned model can use tools to increase performance quality where appropriate. For instance, deterministic tasks can be offloaded to dedicated tools in lieu of probabilistically performing the task with an increased risk of error. For instance, instead of autoregressively predicting the solution to a system of equations, a machine-learned model can recognize a tool to call for obtaining the solution and pass the system of equations to the appropriate tool. The tool can be a traditional system of equations solver that can operate deterministically to resolve the system of equations. The output of the tool can be returned in response to the original query. In this manner, tool use can allow some example models to focus on the strengths of machine-learned models — e.g., understanding an intent in an unstructured request for a task — while augmenting the performance of the model by offloading certain tasks to a more focused tool for rote application of deterministic algorithms to a well-defined problem.

[0304] Model plugin toolkit 18 can include validation tools 18-1. Validation tools 18-1 can include tools that can parse and confirm output(s) of a machine-learned model.Validation tools 18-1 can include engineered heuristics that establish certain thresholds applied to model outputs. For example, validation tools 18-1 can ground the outputs of machine-learned models to structured data sources (e.g., to mitigate “hallucinations”).

[0305] Model plugin toolkit 18 can include tooling packages 18-2 for implementing one or more tools that can include scripts or other executable code that can be executed alongside development model 16. Tooling packages 18-2 can include one or more inputs configured to cause machine-learned model(s) to implement the tools (e g., few-shot prompts that induce a model to output tool calls in the proper syntax, etc.). Tooling packages 18-2 can include, for instance, fine-tuning training data for training a model to use a tool.

[0306] Model plugin toolkit 18 can include interfaces for calling external application programming interfaces (APIs) 18-3. For instance, in addition to or in lieu of implementing tool calls or tool code directly with development model 16, development model 16 can be aligned to output instructions that initiate API calls to send or obtain data via external systems.

[0307] Model plugin toolkit 18 can integrate with prompt libraries 17-4 to build a catalog of available tools for use with development model 16. For instance, a model can receive, in an input, a catalog of available tools, and the model can generate an output that selects a tool from the available tools and initiates a tool call for using the tool.

[0308] Model development platform 12 can include a computational optimization toolkit 19 for optimizing a computational performance of development model 16. Forinstance, tools for model compression 19-1 can allow development model 16 to be reduced in size while maintaining a desired level of performance. For instance, model compression 19-1 can include quantization workflows, weight pruning and sparsification techniques, etc. Tools for hardware acceleration 19-2 can facilitate the configuration of the model storage and execution formats to operate optimally on different hardware resources. For instance, hardware acceleration 19-2 can include tools for optimally sharding models for distributed processing over multiple processing units for increased bandwidth, lower unified memory requirements, etc. Tools for distillation 19-3 can provide for the training of lighter-weight models based on the knowledge encoded in development model 16. For instance, development model 16 can be a highly performant, large machine-learned model optimized using model development platform 12. To obtain a lightweight model for running in resource-constrained environments, a smaller model can be a '‘student model’’ that learns to imitate development model 16 as a “teacher model.” In this manner, for instance, the investment in learning the parameters and configurations of development model 16 can be efficiently transferred to a smaller model for more efficient inference.

[0309] Workbench 15 can implement one, multiple, or none of the toolkits implemented in model development platform 12. Workbench 15 can output an output model 20 based on development model 16. Output model 20 can be a deployment version of development model 16. Output model 20 can be a development or training checkpoint of development model 16. Output model 20 can be a distilled, compressed, or otherwise optimized version of development model 16.

[0310] Figure 13 is a block diagram of an example training flow for training a machine-learned development model 16. One or more portion(s) of the example training flow¬ can be implemented by a computing system that includes one or more computing devices such as, for example, computing systems described with reference to the other figures. Each respective portion of the example training flow can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of the example training flow can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. FIG. 13 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. FIG. 13 is described with reference to elements / terms described withrespect to other systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of the example training flow can be performed additionally, or alternatively, by other systems.

[0311] Initially, development model 16 can persist in an initial state as an initialized model 21. Development model 16 can be initialized with weight values. Initial weight values can be random or based on an initialization schema. Initial weight values can be based on prior pre-training for the same or for a different model.

[0312] Initialized model 21 can undergo pre-training in a pre-training stage 22. Pretraining stage 22 can be implemented using one or more pre-training pipelines 17-2 over data from dataset(s) 17-1. Pre-training can be omitted, for example, if initialized model 21 is already pre-trained (e.g.. development model 16 contains, is, or is based on a pre-trained foundational model or an expert model).

[0313] Pre-trained model 23 can then be a new version of development model 16, which can persist as development model 16 or as a new development model. Pre-trained model 23 can be the initial state if development model 16 was already pre-trained. Pre-trained model 23 can undergo fine-tuning in a fine-tuning stage 24. Fine-tuning stage 24 can be implemented using one or more fine-tuning pipelines 17-3 over data from dataset(s) 17-1. Fine-tuning can be omitted, for example, if a pre-trained model has satisfactory performance, if the model was already fine-tuned, or if other tuning approaches are preferred.

[0314] Fine-tuned model 25 can then be a new version of development model 16, which can persist as development model 16 or as a new development model. Fine-tuned model 25 can be the initial state if development model 16 was already fine-tuned. Fine-tuned model 25 can undergo refinement with user feedback 26. For instance, refinement with user feedback 26 can include reinforcement learning, optionally based on human feedback from human users of fine-tuned model 25. As reinforcement learning can be a form of fine-tuning, it is to be understood that fine-tuning stage 24 can subsume the stage for refining with user feedback 26. Refinement with user feedback 26 can produce a refined model 27. Refined model 27 can be output to downstream system(s) 28 for deployment or further development.

[0315] In some implementations, computational optimization operations can be applied before, during, or after each stage. For instance, initialized model 21 can undergo computational optimization 29-1 (e.g., using computational optimization toolkit 19) before pre-training stage 22. Pre-trained model 23 can undergo computational optimization 29-2 (e.g., using computational optimization toolkit 19) before fine-tuning stage 24. Fine-tuned model 25 can undergo computational optimization 29-3 (e.g., using computationaloptimization toolkit 19) before refinement with user feedback 26. Refined model 27 can undergo computational optimization 29-4 (e.g., using computational optimization toolkit 19) before output to downstream system(s) 28. Computational optimization(s) 29-1,..., 29-4 can all be the same, all be different, or include at least some different optimization techniques.Example Machine-Learned Model Inference System

[0316] Figure 14 is a block diagram of an inference system for operating one or more machine-learned model(s) 1 to perform inference (e.g., for training, for deployment, etc.). A model host 31 can receive machine-learned model(s) 1. Model host 31 can host one or more model instance(s) 31-1, which can be one or multiple instances of one or multiple models. Model host 31 can host model instance(s) 31-1 using available compute resources 31-2 associated with model host 31.

[0317] Model host 31 can perform inference on behalf of one or more client(s) 32. Client(s) 32 can transmit an input request 33 to model host 31. Using input request 33, model host 31 can obtain input(s) 2 for input to machine-learned model(s) 1. Machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3. Using output(s) 3, model host 31 can return an output payload 34 for responding to input request 33 from client(s) 32. Output payload 34 can include or be based on output(s) 3.

[0318] Model host 31 can leverage various other resources and tools to augment the inference task. For instance, model host 31 can communicate with tool interfaces 35 to facilitate tool use by model instance(s) 31-1. Tool interfaces 35 can include local or remote APIs. Tool interfaces 35 can include integrated scripts or other software functionality. Model host 31 can engage online learning interface(s) 36 to facilitate ongoing improvements to machine-learned model(s) 1. For instance, online learning interface(s) 36 can be used within reinforcement learning loops to retrieve user feedback on inferences served by model host 31. Model host 31 can access runtime data source(s) 37 for augmenting input(s) 2 with additional contextual information. For instance, runtime data source(s) 37 can include a knowledge graph 37-1 that facilitates structured information retrieval for information associated with input request(s) 33 (e.g., a search engine service). Runtime data source(s) 37 can include public or private, external or local database(s) 37-2 that can store information associated with input request(s) 33 for augmenting input(s) 2. Runtime data source(s) 37 can include account data 37-3 which can be retrieved in association with a user account corresponding to a client 32 for customizing the behavior of model host 31 accordingly.

[0319] Model host 31 can be implemented by one or multiple computing devices or systems. Client(s) 2 can be implemented by one or multiple computing devices or systems, which can include computing devices or systems shared with model host 31.

[0320] For example, model host 31 can operate on a server system that provides a machine-learning service to client device(s) that operate client(s) 32 (e.g., over a local or wide-area network). Client device(s) can be end-user devices used by individuals. Client device(s) can be server systems that operate client(s) 32 to provide various functionality as a service to downstream end-user devices.

[0321] In some implementations, model host 31 can operate on a same device or system as client(s) 32. Model host 31 can be a machine-learning service that runs on-device to provide machine-learning functionality to one or multiple applications operating on a client device, which can include an application implementing client(s) 32. Model host 31 can be a part of a same application as client(s) 32. For instance, model host 31 can be a subroutine or method implemented by one part of an application, and client(s) 32 can be another subroutine or method that engages model host 31 to perform inference functions within the application. It is to be understood that model host 31 and client(s) 32 can have various different configurations.

[0322] Model instance(s) 31-1 can include one or more machine-learned models that are available for performing inference. Model instance(s) 31-1 can include weights or other model components that are stored in persistent storage, temporarily cached, or loaded into high-speed memory. Model instance(s) 31-1 can include multiple instance(s) of the same model (e.g., for parallel execution of more requests on the same model). Model instance(s) 31-1 can include instance(s) of different model(s). Model instance(s) 31-1 can include cached intermediate states of active or inactive model(s) used to accelerate inference of those models. For instance, an inference session with a particular model may generate significant amounts of computational results that can be re-used for future inference runs (e.g., using a KV cache for transformer-based models). These computational results can be saved in association with that inference session so that session can be executed more efficiently when resumed.

[0323] Compute resource(s) 31-2 can include one or more processors (central processing units, graphical processing units, tensor processing units, machine-learning accelerators, etc.) connected to one or more memory' devices. Compute resource(s) 31-2 can include a dynamic pool of available resources shared with other processes. Compute resource(s) 31-2 can include memory devices large enough to fit an entire model instance in asingle memory instance. Compute resource(s) 31-2 can also shard model instance(s) across multiple memory devices (e.g., using data parallelization or tensor parallelization, etc.). This can be done to increase parallelization or to execute a large model using multiple memory devices which individually might not be able to fit the entire model into memory.

[0324] Input request 33 can include data for input(s) 2. Model host 31 can process input request 33 to obtain input(s) 2. Input(s) 2 can be obtained directly from input request 33 or can be retrieved using input request 33. Input request 33 can be submitted to model host 31 via an API.

[0325] Model host 31 can perform inference over batches of input requests 33 in parallel. For instance, a model instance 31-1 can be configured with an input structure that has a batch dimension. Separate input(s) 2 can be distributed across the batch dimension (e.g., rows of an array). The separate input(s) 2 can include completely different contexts. The separate input(s) 2 can be multiple inference steps of the same task. The separate input(s) 2 can be staggered in an input structure, such that any given inference cycle can be operating on different portions of the respective input(s) 2. In this manner, for instance, model host 31 can perform inference on the batch in parallel, such that output(s) 3 can also contain the batch dimension and return the inference results for the batched input(s) 2 in parallel. In this manner, for instance, batches of input request(s) 33 can be processed in parallel for higher throughput of output payload(s) 34.

[0326] Output payload 34 can include or be based on output(s) 3 from machine-learned model(s) 1. Model host 31 can process output(s) 3 to obtain output payload 34. This can include chaining multiple rounds of inference (e.g., iteratively, recursively, across the same model(s) or different model(s)) to arrive at a final output for a task to be returned in output payload 34. Output payload 34 can be transmitted to client(s) 32 via an API.

[0327] Online learning interface(s) 36 can facilitate reinforcement learning of machine-learned model(s) 1. Online learning interface(s) 36 can facilitate reinforcement learning with human feedback (RLHF). Online learning interface(s) 36 can facilitate federated learning of machine-learned model(s) 1.

[0328] Model host 31 can execute machine-learned model(s) 1 to perform inference for various tasks using various types of data. For example, various different input(s) 2 and output(s) 3 can be used for various different tasks. In some implementations, input(s) 2 can be or otherwise represent image data. Machine-learned model(s) 1 can process the image data to generate an output. As an example, machine-learned model(s) 1 can process the image data to generate an image recognition output (e.g., a recognition of the image data, a latentembedding of the image data, an encoded representation of the image data, a hash of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an image segmentation output. As another example, machine-learned model(s) 1 can process the image data to generate an image classification output. As another example, machine-learned model(s) 1 can process the image data to generate an image data modification output (e.g., an alteration of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an encoded image data output (e.g., an encoded and / or compressed representation of the image data, etc.). As another example, machine-learned model(s) 1 can process the image data to generate an upscaled image data output. As another example, machine-learned model(s) 1 can process the image data to generate a prediction output.

[0329] In some implementations, the task is a computer vision task. In some cases, input(s) 2 includes pixel data for one or more images and the task is an image processing task. For example, the image processing task can be image classification, where the output is a set of scores, each score corresponding to a different object class and representing the likelihood that the one or more images depict an object belonging to the object class. The image processing task may be object detection, where the image processing output identifies one or more regions in the one or more images and, for each region, a likelihood that region depicts an object of interest. As another example, the image processing task can be image segmentation, where the image processing output defines, for each pixel in the one or more images, a respective likelihood for each category in a predetermined set of categories. For example, the set of categories can be foreground and background. As another example, the set of categories can be object classes. As another example, the image processing task can be depth estimation, where the image processing output defines, for each pixel in the one or more images, a respective depth value. As another example, the image processing task can be motion estimation, where the network input includes multiple images, and the image processing output defines, for each pixel of one of the input images, a motion of the scene depicted at the pixel between the images in the network input.

[0330] In some implementations, input(s) 2 can be or otherwise represent natural language data. Machine-learned model(s) 1 can process the natural language data to generate an output. As an example, machine-learned model(s) 1 can process the natural language data to generate a language encoding output. As another example, machine-learned model(s) 1 can process the natural language data to generate a latent text embedding output. As another example, machine-learned model(s) 1 can process the natural language data to generate atranslation output. As another example, machine-learned model(s) 1 can process the natural language data to generate a classification output. As another example, machine-learned model(s) 1 can process the natural language data to generate a textual segmentation output. As another example, machine-learned model(s) 1 can process the natural language data to generate a semantic intent output. As another example, machine-learned model(s) 1 can process the natural language data to generate an upscaled text or natural language output (e.g., text or natural language data that is higher quality than the input text or natural language, etc.). As another example, machine-learned model(s) 1 can process the natural language data to generate a prediction output (e.g., one or more predicted next portions of natural language content).

[0331] In some implementations, input(s) 2 can be or otherwise represent speech data (e.g., data describing spoken natural language, such as audio data, textual data, etc.).Machine-learned model(s) 1 can process the speech data to generate an output. As an example, machine-learned model(s) 1 can process the speech data to generate a speech recognition output. As another example, machine-learned model(s) 1 can process the speech data to generate a speech translation output. As another example, machine-learned model(s) 1 can process the speech data to generate a latent embedding output. As another example, machine-learned model(s) 1 can process the speech data to generate an encoded speech output (e.g., an encoded and / or compressed representation of the speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate an upscaled speech output (e.g., speech data that is higher quality than the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a textual representation output (e.g., a textual representation of the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a prediction output.

[0332] In some implementations, input(s) 2 can be or otherwise represent latent encoding data (e.g., a latent space representation of an input, etc.). Machine-learned model(s) 1 can process the latent encoding data to generate an output. As an example, machine-learned model(s) 1 can process the latent encoding data to generate a recognition output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a reconstruction output. As another example, machine-learned model (s) 1 can process the latent encoding data to generate a search output. As another example, machine-learned model(s) 1 can process the latent encoding data to generate a reclustering output. Asanother example, machine-learned model(s) 1 can process the latent encoding data to generate a prediction output.

[0333] In some implementations, input(s) 2 can be or otherwise represent statistical data. Statistical data can be, represent, or otherwise include data computed and / or calculated from some other data source. Machine-learned model(s) 1 can process the statistical data to generate an output. As an example, machine-learned model(s) 1 can process the statistical data to generate a recognition output. As another example, machine-learned model(s) 1 can process the statistical data to generate a prediction output. As another example, machine-learned model(s) 1 can process the statistical data to generate a classification output. As another example, machine-learned model(s) 1 can process the statistical data to generate a segmentation output. As another example, machine-learned model(s) 1 can process the statistical data to generate a visualization output. As another example, machine-learned model(s) 1 can process the statistical data to generate a diagnostic output.

[0334] In some implementations, input(s) 2 can be or otherwise represent sensor data. Machine-learned model(s) 1 can process the sensor data to generate an output. As an example, machine-learned model(s) 1 can process the sensor data to generate a recognition output. As another example, machine-learned model(s) 1 can process the sensor data to generate a prediction output. As another example, machine-learned model(s) 1 can process the sensor data to generate a classification output. As another example, machine-learned model(s) 1 can process the sensor data to generate a segmentation output. As another example, machine-learned model(s) 1 can process the sensor data to generate a visualization output. As another example, machine-learned model(s) 1 can process the sensor data to generate a diagnostic output. As another example, machine-learned model(s) 1 can process the sensor data to generate a detection output.

[0335] In some implementations, machine-learned model(s) 1 can be configured to perform a task that includes encoding input data for reliable and / or efficient transmission or storage (and / or corresponding decoding). For example, the task may be an audio compression task. The input may include audio data and the output may comprise compressed audio data. In another example, the input includes visual data (e.g. one or more images or videos), the output comprises compressed visual data, and the task is a visual data compression task. In another example, the task may comprise generating an embedding for input data (e.g. input audio or visual data). In some cases, the input includes audio data representing a spoken utterance and the task is a speech recognition task. The output may comprise a text output which is mapped to the spoken utterance. In some cases, the task comprises encrypting ordecrypting input data. In some cases, the task comprises a microprocessor performance task, such as branch prediction or memory address translation.

[0336] In some implementations, the task is a generative task, and machine-learned model(s) 1 can be configured to output content generated in view of input(s) 2. For instance, input(s) 2 can be or otherwise represent data of one or more modalities that encodes context for generating additional content.

[0337] In some implementations, the task can be a text completion task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent textual data and to generate output(s) 3 that represent additional textual data that completes a textual sequence that includes input(s) 2. For instance, machine-learned model(s) 1 can be configured to generate output(s) 3 to complete a sentence, paragraph, or portion of text that follows from a portion of text represented by input(s) 2.

[0338] In some implementations, the task can be an instruction following task.Machine-learned model(s) 1 can be configured to process input(s) 2 that represent instructions to perform a function and to generate output(s) 3 that advance a goal of satisfying the instruction function (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent data of the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g.. natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially process and accomplish steps toward accomplishing the requested functionality. For instance, an initial output can be executed by an external system or be processed by machine-learned model(s) 1 to complete an initial step of performing a function. Multiple steps can be performed, with a final output being obtained that is responsive to the initial instructions.

[0339] In some implementations, the task can be a question answering task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent a question to answer and to generate output(s) 3 that advance a goal of returning an answer to the question (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent dataof the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the question (e.g., natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the question (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially process and accomplish steps toward answering the question. For instance, an initial output can be executed by an external system or be processed by machine-learned model(s) 1 to complete an initial step of obtaining an answer to the question (e.g., querying a database, performing a computation, executing a script, etc.). Multiple steps can be performed, with a final output being obtained that is responsive to the question.

[0340] In some implementations, the task can be an image generation task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of image content. The context can include text data, image data, audio data, etc. Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent image data that depicts imagery related to the context. For instance, machine-learned model(s) 1 can be configured to generate pixel data of an image. Values for channel(s) associated with the pixels in the pixel data can be selected based on the context (e.g., based on a probability determined based on the context).

[0341] In some implementations, the task can be an audio generation task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of audio content. The context can include text data, image data, audio data, etc. Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent audio data related to the context. For instance, machine-learned model(s) 1 can be configured to generate waveform data in the form of an image (e.g., a spectrogram). Values for channel(s) associated with pixels of the image can be selected based on the context. Machine-learned model(s) 1 can be configured to generate waveform data in the form of a sequence of discrete samples of a continuous waveform. Values of the sequence can be selected based on the context (e.g.. based on a probability determined based on the context).

[0342] In some implementations, the task can be a data generation task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of data (e.g., data from various data domains, such as sensor data, image data, multimodal data, statistical data, etc.). The desired data can be, for instance, synthetic data for training other machine-learned models. The context can include arbitrary data type(s).Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent data that aligns with the desired data. For instance, machine-learned model(s) 1 can be configured to generate data values for populating a dataset. Values for the data object(s) can be selected based on the context (e.g., based on a probability determined based on the context).Example Computing Systems and Devices

[0343] Figure 15 is a block diagram of an example networked computing system that can perform aspects of example implementations of the present disclosure. The system can include a number of computing devices and systems that are communicatively coupled over a network 49. An example computing device 50 is described to provide an example of a computing device that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). An example server computing system 60 is described as an example of a server computing system that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Computing device 50 and server computing system(s) 60 can cooperatively interact (e.g., over network 49) to perform any aspect of the present disclosure (e g., implementing model host 31, client(s) 32, or both). Model development platform system 70 is an example system that can host or serve model development platform(s) 12 for development of machine-learned models. Third-party system(s) 80 are example system(s) with which any of computing device 50, server computing system(s) 60, or model development platform system(s) 70 can interact in the performance of various aspects of the present disclosure (e.g., engaging third-party tools, accessing third-party databases or other resources, etc.).

[0344] Network 49 can be any type of communications network, such as a local area network (e.g., intranet), wide area network (e.g., Internet), or some combination thereof and can include any number of wired or wireless links. In general, communication over network 49 can be carried via any type of wired or wireless connection, using a wide variety of communication protocols (e.g., TCP / IP, HTTP, SMTP, FTP), encodings or formats (e.g., HTML, XML), or protection schemes (e.g., VPN. secure HTTP, SSL). Network 49 can also be implemented via a system bus. For instance, one or more devices or systems of Figure 15can be co-located with, contained by, or otherwise integrated into one or more other devices or systems.

[0345] Computing device 50 can be any type of computing device, such as, for example, a personal computing device (e.g., laptop or desktop), a mobile computing device (e.g., smartphone or tablet), a gaming console or controller, a wearable computing device, an embedded computing device, a server computing device, a virtual machine operating on a host device, or any other type of computing device. Computing device 50 can be a client computing device. Computing device 50 can be an end-user computing device. Computing device 50 can be a computing device of a service provided that provides a service to an end user (who may use another computing device to interact with computing device 50).

[0346] Computing device 50 can include one or more processors 51 and a memory 52. Processor(s) 51 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 52 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 52 can store data 53 and instructions 54 which can be executed by processor(s) 51 to cause computing device 50 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein.

[0347] Computing device 50 can also include one or more input components that receive user input. For example, a user input component can be a touch-sensitive component (e.g., a touch-sensitive display screen or a touch pad) that is sensitive to the touch of a user input object (e.g., a finger or a stylus). The touch-sensitive component can serve to implement a virtual keyboard. Other example user input components include a microphone, camera, LIDAR, a physical keyboard or other buttons, or other means by which a user can provide user input.

[0348] Computing device 50 can store or include one or more machine-learned models 55. Machine-learned models 55 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 55 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 55 can be received from server computing system(s) 60, model development platform system 70. third party system(s) 80 (e.g., an application distribution platform), or developed locally on computing device 50. Machine-learned model(s) 55 can be loaded into memory 52 and used or otherwiseimplemented by processor(s) 51. Computing device 50 can implement multiple parallel instances of machine-learned model(s) 55.

[0349] Server computing system(s) 60 can include one or more processors 61 and a memory 62. Processor(s) 61 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory’ 62 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory’ 62 can store data 63 and instructions 64 which can be executed by processor(s) 61 to cause server computing system(s) 60 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein.

[0350] In some implementations, server computing system 60 includes or is otherwise implemented by one or multiple server computing devices. In instances in which server computing system 60 includes multiple server computing devices, such server computing devices can operate according to sequential computing architectures, parallel computing architectures, or some combination thereof.

[0351] Server computing system 60 can store or otherwise include one or more machine-learned models 65. Machine-learned model(s) 65 can be the same as or different from machine-learned model(s) 55. Machine-learned models 65 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 65 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 65 can be received from computing device 50, model development platform system 70, third party system(s) 80. or developed locally on server computing system(s) 60. Machine-learned model(s) 65 can be loaded into memory 62 and used or otherwise implemented by processor(s) 61. Server computing system(s) 60 can implement multiple parallel instances of machine-learned model(s) 65.

[0352] In an example configuration, machine-learned models 65 can be included in or otherwise stored and implemented by server computing system 60 to establish a client-server relationship with computing device 50 for serving model inferences. For instance, server computing system(s) 60 can implement model host 31 on behalf of client(s) 32 on computing device 50. For instance, machine-learned models 65 can be implemented by server computing system 60 as a portion of a web service (e.g., remote machine-learned model hosting service, such as an online interface for performing machine-learned model operations over a networkon server computing system(s) 60). For instance, server computing system(s) 60 can communicate with computing device 50 over a local intranet or internet connection. For instance, computing device 50 can be a workstation or endpoint in communication with server computing system(s) 60, with implementation of machine-learned models 65 being managed by server computing system(s) 60 to remotely perform inference (e.g., for runtime or training operations), with output(s) returned (e.g., cast, streamed, etc.) to computing device 50. Machine-learned models 65 can work cooperatively or interoperatively with machine-learned models 55 on computing device 50 to perform various tasks.

[0353] Model development platform system(s) 70 can include one or more processors 71 and a memory 72. Processor(s) 71 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA. a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 72 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 72 can store data 73 and instructions 74 which can be executed by processor(s) 71 to cause model development platform system(s) 70 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein. Example operations include the functionality described herein with respect to model development platform 12. This and other functionality can be implemented by developer tool(s) 75.

[0354] Third-party system(s) 80 can include one or more processors 81 and a memory 82. Processor(s) 81 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA. a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory’ 82 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory’ 82 can store data 83 and instructions 84 which can be executed by processor(s) 81 to cause third-party system(s) 80 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein. Example operations include the functionality described herein wi th respect to tools and other external resources called when training or performing inference with machine-learned model(s) 1, 4, 16, 20, 55, 65, etc. (e.g., third-party resource(s) 85).

[0355] Figure 15 illustrates one example arrangement of computing systems that can be used to implement the present disclosure. Other computing system configurations can be used as well. For example, in some implementations, one or both of computing system 50 or server computing system(s) 60 can implement all or a portion of the operations of model development platform system 70. For example, computing system 50 or server computing system(s) 60 can implement developer tool(s) 75 (or extensions thereof) to develop, update / train, or refine machine-learned models 1, 4, 16, 20, 55, 65, etc. using one or more techniques described herein with respect to model alignment toolkit 17. In this manner, for instance, computing system 50 or server computing system(s) 60 can develop, update / train, or refine machine-learned models based on local datasets (e.g., for model personalization / customization. as permitted by user data preference selections).

[0356] Figure 16 is a block diagram of an example computing device 98 that performs according to example embodiments of the present disclosure. Computing device 98 can be a user computing device or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 98 can include a number of applications (e.g., applications 1 through N). Each application can contain its own machine learning library and machine-learned model(s). For example, each application can include a machine-learned model.Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. As illustrated in Figure 16, each application can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, each application can communicate with each device component using an API (e.g., a public API). In some implementations, the API used by each application is specific to that application.

[0357] Figure 17 is a block diagram of an example computing device 99 that performs according to example embodiments of the present disclosure. Computing device 99 can be the same as or different from computing device 98. Computing device 99 can be a user computing device or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 99 can include a number of applications (e.g., applications 1 through N). Each application can be in communication with a central intelligence layer. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. In someimplementations, each application can communicate with the central intelligence layer (and model(s) stored therein) using an API (e.g., a common API across all applications).

[0358] The central intelligence layer can include a number of machine-learned models. For example, as illustrated in Figure 17, a respective machine-learned model can be provided for each application and managed by the central intelligence layer. In other implementations, two or more applications can share a single machine-learned model. For example, in some implementations, the central intelligence layer can provide a single model for all of the applications. In some implementations, the central intelligence layer is included within or otherwise implemented by an operating system of computing device 99.

[0359] The central intelligence layer can communicate with a central device data layer. The central device data layer can be a centralized repository of data for computing device 99. As illustrated in Figure 17, the central device data layer can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, the central device data layer can communicate with each device component using an API (e.g., a private API).Additional Disclosure

[0360] The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sent to and from such systems. The inherent flexibility of computer-based systems allows for a great variety of possible configurations, combinations, and divisions of tasks and functionality between and among components. For instance, processes discussed herein can be implemented using a single device or component or multiple devices or components working in combination. Databases and applications can be implemented on a single system or distributed across multiple systems. Distributed components can operate sequentially or in parallel.

[0361] While the present subject matter has been described in detail with respect to various specific example embodiments thereof, each example is provided by way of explanation, not limitation of the disclosure. Those skilled in the art, upon attaining an understanding of the foregoing, can readily produce alterations to, variations of, and equivalents to such embodiments. Accordingly, the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. For instance, features illustrated ordescribed as part of one embodiment can be used with another embodiment to yield a still further embodiment. Thus, it is intended that the present disclosure cover such alterations, variations, and equivalents.

[0362] Aspects of the disclosure have been described in terms of illustrative embodiments thereof. Any and all features in the following claims can be combined or rearranged in any way possible, including combinations of claims not explicitly enumerated in combination together, as the example claim dependencies listed herein should not be read as limiting the scope of possible combinations of features disclosed herein. Accordingly, the scope of the present disclosure is by way of example rather than by way of limitation, and the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. Moreover, terms are described herein using lists of example elements joined by conjunctions such as “and,” “or,” “but,” etc. It should be understood that such conjunctions are provided for explanatory purposes only. Clauses and other sequences of items joined by a particular conjunction such as “or,” for example, can refer to “and / or,” “at least one of’, “any combination of’ example elements listed therein, etc. Terms such as “based on” should be understood as “based at least in part on.”

[0363] The term “can” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability that is necessarily present in every implementation. For example, the phrase “X can perform Y” should be understood as indicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

[0364] The term “may” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability' that is necessarily present in every implementation. For example, the phrase “X may perform Y” should be understood as indicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

Claims

WHAT IS CLAIMED IS:

1. A method comprising:obtaining, by a computing system comprising one or more computing devices, a dataset comprising a plurality of data entries, each data entry of the plurality of data entries correlating a first data item of the data entry to one or more ground truth classification labels associated with the first data item;obtaining, by the computing system, a plurality of machine-learned confidence values comprising, for each data entry of the plurality of data entries, one or more respective machine-learned confidence values indicative of a confidence that the first data item of the data entry belongs to one or more respective classes associated with the one or more ground truth classification labels;assigning, by the computing system based at least in part on the plurality of machine-learned confidence values, each data entry of the plurality of data entries to a corresponding confidence bin of a plurality of confidence bins;determining, by the computing system based at least in part on the plurality of confidence bins and the one or more ground truth classification labels of the plurality of data entries, a count matrix indicative of a plurality of ground truth classification label counts, each count of the plurality of ground truth classification label counts being associated respectively with a corresponding confidence bin of the plurality of confidence bins and a corresponding class of a plurality of possible classes; anddetermining, by the computing system based at least in part on the count matrix, a probability matrix correlating each respective confidence bin of the plurality of confidence bins with one or more corresponding classification probabilities associated with the respective confidence bin.

2. The method of claim 1, wherein determining the probability matrix comprises determining one or more parameters of a parametrized probability matrix based at least in part on a relationship between the one or more parameters and a probability of randomly sampling the count matrix given the probability matrix.

3. The method of claim 1, wherein the probability matrix is a matrix product of a fixed basis matrix and a matrix of coefficient values, and determining the probability matrix comprises determining the matrix of coefficient values using fixed point iteration.

4. The method of claim 3, wherein the fixed point iteration comprises:initializing, by the computing system, an initial value for the matrix of coefficient values; andfor each of a plurality of iterations:determining, by the computing system, an updated value for the matrix of coefficient values, wherein the updated value is determined based at least in part on a current value of the matrix of coefficient values.

5. The method of claim 1, wherein determining the probability matrix comprises Poisson factorization.

6. The method of claim 1, wherein the count matrix comprises a weighted sum of count values.

7. The method of claim 6, wherein the plurality of machine-learned confidence values were generated using a first machine-learned model, and wherein the weighted sum of count values comprises one or more count values indicative of one or more classification labels generated by a second machine-learned model different from the first machine-learned model.

8. The method of claim 1, further comprising:determining, by the computing system based at least in part on the probability matrix, an estimated performance metric associated with one or more first machine-learned models;selecting, by the computing system based at least in part on the estimated performance metric, an action; andcausing, by the computing system, the action to be performed.

9. The method of claim 8, wherein the estimated performance metric is associated with one or more outputs of the one or more first machine-learned models, and selecting the action comprises selecting, based at least in part on a determination that the estimated performance metric satisfies a performance metric threshold, an action to output the one or more outputs.

10. The method of claim 8, wherein the estimated performance metric is associated with one or more first outputs of the one or more first machine-learned models, and wherein selecting the action comprises selecting, based at least in part on a determination that the estimated performance metric does not satisfy a performance metric threshold, a corrective action.

11. The method of claim 10, wherein the corrective action comprises generating one or more second outputs.

12. The method of claim 11, wherein generating the one or more second outputs comprises generating the one or more second outputs using one or more second machine-learned models different from the one or more first machine-learned models.

13. The method of claim 8, wherein the action is a machine-learned inference operation, and wherein selecting the action comprises selecting, based at least in part on the estimated performance metric, whether to use the one or more first machine-learned models to perform the machine-learned inference operation.

14. The method of claim 8, wherein the estimated performance metric comprises an estimated likelihood that one or more outputs of the one or more first machine-learned models belongs to a set of one or more target classes.

15. The method of claim 1, wherein the one or more ground truth classification labels comprise one or more of:one or more entailment classification labels;one or more relevance classification labels;one or more risk classification labels;one or more conciseness classification labels;one or more creativity classification labels;one or more genre classification labels; andone or more classification labels indicative of a degree to which a first portion of the firstdata item complies with instructions contained in a second portion of the first data item.

16. The method of claim 1, further comprising determining, by the computing system based at least in part on a number of data entries of the plurality of data entries, a number of bins of the plurality of confidence bins.

17. A computing system comprising one or more processors and one or more non-transitory computer-readable media storing instructions that are executable by one or more processors to cause the computing system to perform operations, the operations comprising:obtaining a dataset comprising a plurality of data entries, each data entry of the plurality of data entries correlating a first data item of the data entry to one or more ground truth classification labels associated with the first data item;obtaining a plurality of machine-learned confidence values comprising, for each data entry of the plurality of data entries, one or more respective machine-learned confidence values indicative of a confidence that the first data item of the data entry belongs to one or more respective classes associated with the one or more ground truth classification labels; assigning, based at least in part on the plurality of machine-learned confidence values, each data entry of the plurality of data entries to a corresponding confidence bin of a plurality of confidence bins;determining, based at least in part on the plurality of confidence bins and the one or more ground truth classification labels of the plurality of data entries, a count matrix indicative of a plurality of ground truth classification label counts, each count of the plurality of ground truth classification label counts being associated respectively with a corresponding confidence bin of the plurality of confidence bins and a corresponding class of a plurality of possible classes; anddetermining, based at least in part on the count matrix, a probability matrix correlating each respective confidence bin of the plurality of confidence bins with one or more corresponding classification probabilities associated with the respective confidence bin.

18. The computing system of claim 17, wherein determining the probability matrix comprises Poisson factorization.

19. The computing system of claim 17, wherein the operations further comprise:determining, based at least in part on the probability matrix, an estimated performance metric associated with one or more first machine-learned models;selecting, based at least in part on the estimated performance metric, an action; andcausing the action to be performed.

20. One or more non-transitory computer-readable media storing instructions that are executable by one or more processors to cause a computing system to perform operations, the operations comprising:obtaining a dataset comprising a plurality of data entries, each data entry of the plurality of data entries correlating a first data item of the data entry to one or more ground truth classification labels associated with the first data item;obtaining a plurality of machine-learned confidence values comprising, for each data entry of the plurality of data entries, one or more respective machine-learned confidence values indicative of a confidence that the first data item of the data entry belongs to one or more respective classes associated with the one or more ground truth classification labels; assigning, based at least in part on the plurality of machine-learned confidence values, each data entry of the plurality of data entries to a corresponding confidence bin of a plurality of confidence bins:determining, based at least in part on the plurality of confidence bins and the one or more ground truth classification labels of the plurality of data entries, a count matrix indicative of a plurality of ground truth classification label counts, each count of the plurality of ground truth classification label counts being associated respectively with a corresponding confidence bin of the plurality of confidence bins and a corresponding class of a plurality of possible classes; anddetermining, based at least in part on the count matrix, a probability' matrix correlating each respective confidence bin of the plurality of confidence bins with one or more corresponding classification probabilities associated with the respective confidence bin.

Images