Multilevel tokenization for efficient execution of machine-learned models

Abstract

An example method includes generating, using an interface token encoder, a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence corresponding to a query. The example method includes generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens. The example method includes generating, using a machine-learned sequence processing model that processes the one or more contextual tokens, one or more predicted contextual tokens. The example method includes generating, using a contextual detokenizer that processes the one or more predicted contextual tokens, an output sequence including one or more predicted interface tokens that represent the one or more predicted contextual tokens. The example method includes generating, using an interface detokenizer that processes the output sequence, a response to the query.

Description

MULTILEVEL TOKENIZATION FOR EFFICIENTEXECUTION OF MACHINE-LEARNED MODELSPRIORITY CLAIM

[0001] The present application is based on and claims priority to United States Provisional Application Number 63 / 684,218 (filed August 16, 2024), which is hereby incorporated by reference herein in its entirety.FIELD

[0002] The present disclosure relates generally to machine learning processes and machine- learned devices and systems. More particularly, the present disclosure relates to multilevel tokenization.BACKGROUND

[0003] 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

[0004] 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.

[0005] In an aspect, the present disclosure provides a first example method. In some implementations, the first example method includes generating, using an interface token encoder, a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence (e.g., corresponding to a query). In some implementations, the first example method includes generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens. In some implementations, the first example method includes generating, using a machine-learned sequence processing model that processes the one or more contextual tokens, one or more predicted contextual tokens. In some implementations, the first example method includesgenerating, using a contextual detokenizer that processes the one or more predicted contextual tokens, an output sequence comprising one or more predicted interface tokens that represent the one or more predicted contextual tokens. In some implementations, the first example method includes generating, using an interface detokenizer that processes the output sequence, a response to the query.

[0006] In an aspect, the present disclosure provides a second example method. In some implementations, the second example method includes generating, using an interface token encoder, a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence. In some implementations, the second example method includes generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens, wherein the contextual tokenizer includes a quantization layer that maps each interface token encoding to a corresponding quantized representation. In some implementations, the second example method includes generating, using a contextual detokenizer that processes the one or more contextual tokens, one or more decoded interface token encodings that represent the one or more contextual tokens. In some implementations, the second example method includes training the contextual detokenizer using a loss function that computes a reconstruction loss between a decoded interface token encoding and a corresponding interface token encoding.

[0007] In an aspect, the present disclosure provides a third example method. In some implementations, the third example method includes freezing the contextual tokenizer and the contextual detokenizer (e.g., after training using, for instance, the second example method). In some implementations, the third example method includes training a machine-learned sequence processing model to generate predicted contextual tokens. In some implementations of the third example method, training the machine-learned sequence processing model includes generating, using the interface token encoder, a plurality of training interface token encodings respectively for a plurality of training interface tokens of a training data sequence. In some implementations of the third example method, training the machine-learned sequence processing model includes generating, using the contextual tokenizer, a sequence of one or more training contextual tokens representing the plurality of training interface tokens. In some implementations of the third example method, training the machine-learned sequence processing model includes generating, using the machine-learned sequence processing model that processes the one or more training contextual tokens, a contextual token prediction that indicates a likelihood associated with a contextual token for a particular location in the sequence of the one or more training contextual tokens. In some implementations of the third example method, training the machine-learned sequence processing model includes backpropagating a token prediction loss through themachine-learned sequence processing model to compute a model update, wherein the token prediction loss is based on a classification loss for the contextual token prediction evaluated using a reference contextual token present in the sequence of the one or more training contextual tokens at the particular location. In some implementations of the third example method, training the machine-learned sequence processing model includes updating, using the model update, the machine-learned sequence processing model.

[0008] In an aspect, the present disclosure provides a fourth example method. In some implementations, the fourth example method includes generating, using an interface token encoder (e.g., the interface token encoder of any of the preceding claims), a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence. In some implementations, the fourth example method includes generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens, wherein the contextual tokenizer includes a quantization layer that maps each interface token encoding to a corresponding quantized representation. In some implementations, the fourth example method includes storing, in a computer-readable storage medium, a compressed representation of the plurality of interface tokens, wherein the compressed representation includes the one or more contextual tokens.

[0009] In an aspect, the present disclosure provides one or more example non-transitory, computer readable media storing instructions that, when executed by one or more processors, cause a computing system to perform operations, wherein the operations include any implementation of the first example method, the second example method, or the third example method.

[0010] In an aspect, the present disclosure provides an example computing system. The example computing system includes one or more processors. The example computing system includes one or more non-transitory, computer readable media storing instructions that, when executed by the one or more processors, cause the example computing system to perform operations, wherein the operations include any implementation of the first example method, the second example method, or the third example method.

[0011] Other example aspects of the present disclosure are directed to other systems, methods, apparatuses, tangible non-transitory computer-readable media, computer program products, 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

[0012] Figure 1 is a block diagram of an example system according to example implementations of aspects of the present disclosure.

[0013] Figure 2 is a block diagram of aspects of an example system according to example implementations of aspects of the present disclosure.

[0014] Figure 3 is a block diagram of aspects of an example system according to example implementations of aspects of the present disclosure.

[0015] Figure 4 is a block diagram of aspects of an example system according to example implementations of aspects of the present disclosure.

[0016] Figure 5 is a block diagram of aspects of an example system according to example implementations of aspects of the present disclosure.

[0017] Figure 6 is a block diagram of aspects of an example system according to example implementations of aspects of the present disclosure.

[0018] Figure 7 is a block diagram of aspects of an example system according to example implementations of aspects of the present disclosure.

[0019] Figure 8 is a block diagram of aspects of an example system according to example implementations of aspects of the present disclosure.

[0020] Figure 9 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.

[0021] Figure 10 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.

[0022] Figure 11 is a block diagram of an example sequence processing model according to example implementations of aspects of the present disclosure.

[0023] Figure 12 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.

[0024] Figure 13 is a block diagram of an example model development platform according to example implementations of aspects of the present disclosure.

[0025] Figure 14 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.

[0026] Figure 15 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.

[0027] Figure 16 is a block diagram of an example networked computing system according to example implementations of aspects of the present disclosure.

[0028] Figure 17 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure.

[0029] Figure 18 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure.

[0030] Figure 19 is a flow chart diagram illustrating an example method for using a tokenizer according to example implementations of aspects of the present disclosure.

[0031] Figure 20 is a flow chart diagram illustrating an example method for training a tokenizer according to example implementations of aspects of the present disclosure.

[0032] Figure 21 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 22 is a flow chart diagram illustrating an example method for using a tokenizer according to example implementations of aspects of the present disclosure.DETAILED DESCRIPTION

[0034] Generally, the present disclosure is directed to techniques for multilevel tokenization. Query data can be provided for inference (e.g., a prompt for generating a response). A first tokenization scheme can be applied to query data to obtain a first level of tokens, which may be referred to as “interface” tokens. These tokens can provide a medium of interaction with input pipelines. A token encoder model can process the sequence of interface tokens to generate a corresponding sequence of interface token encodings that represent the tokens in context. These rich, contextualized representations can be input to a contextual tokenizer to generate a second level of tokens, contextual tokens.

[0035] Individual contextual tokens can be more expressive than an interface token. The contextual tokens can be selected based on the context and knowledge imbued by the token encodings. The contextual tokens can be constructed using product quantization schemes to compactly represent a large feature space. The contextual tokens can provide a dense medium for executing inference in a machine-learned sequence processing model (e.g., a model configured to predict next contextual token(s) based on input token(s)). By performing token processing in the dense token space, the model can be smaller or execute fewer cycles as compared to performing the same operations in a less information-dense medium. The predicted contextual tokens can be decoded using a detokenizing model to obtain a representation in an interface token space for outputting response data responsive to the query data.

[0036] In an example aspect, the present disclosure provides techniques for improved machine-learned model execution using a richly contextualized second level of contextual tokens that are generated based on a first level of interface tokens. An interface token encoder model can include a portion of a pre-trained sequence processing model that injects both implicit context (e.g., from other portions of the query data) and external context (e.g., world knowledge learned from a corpus of training data) into the interface token encodings. For instance, a pretrained language model can include one or more input encoding layers that operate to store the meaning of individual input tokens in context with the other input tokens in the sequence and store this contextualized information in a token encoding at a sequence position corresponding to a sequence position of the corresponding input token. For instance, an output of a transformer block in a transformer-based language processing model can include token encodings associated with token positions that reflect one or more rounds of self-attention operations over the input sequence.

[0037] In this manner, some of the task load for interpreting and contextualizing the input sequence can be offloaded from the machine-learned sequence processing model to the interface token encoder. The machine-learned sequence processing model can be provided with richly contextualized, dense tokenized representations. The machine-learned sequence processing model can subsequently predict dense tokenized representations, thereby concisely and compactly expressing complex information. Expressing such rich content using tokens (rather than continuous embeddings) can enable a lightweight tuple of token identifiers to represent the contents of large, high-dimensional vector representations. For instance, in some test implementations, contextual tokens according to the present disclosure can provide 10X, 100X, 1000X+ compression ratios (e.g., in data footprint) over native continuous embeddings.

[0038] Further, using more highly expressive tokens can provide for generating query responses with machine-learned sequence processing models with a decreased token count. For instance, a system can be configured to process a long-form query input, such as a document of 16,000 words, that is initially converted into a sequence of 24,000 interface tokens. A contextual tokenizer can then process this sequence to generate a compressed sequence of 1,500 contextual tokens, which corresponds to a 16-to-l compression ratio, thereby decreasing the number of tokens used to perform inference over the same document. This can decrease a token count per task, or a rate at which tokens are consumed on a per-task basis. As an example, this reduction in sequence length can lower the computational cost of a downstream model’s self-attention mechanism. Such a reduction can facilitate the processing of long-form inputs on a computing system (e.g., a server, a personal computer, a mobile device) that might otherwise encounter performance challenges due to memory or processing limitations.

[0039] In this manner, for instance, the machine-learned sequence processing model may be able to perform tasks using fewer decoding steps (e.g., using more compact token representations, such as representations that shorten a sequence length relative to the interface token sequence), smaller parameterizations (e.g., due to offloading task load to upstream encoder), etc. This in turn can provide for lower latency, fewer floating point operations, less electricity usage, etc.

[0040] For example, a machine-learned sequence processing model trained and configured to operate in a contextual token space can be significantly smaller than machine-learned sequence processing models trained and configured to operate in a traditional token space (e.g., word tokens, etc.). Such smaller models can execute on less energy-intensive hardware (e.g., with less memory, fewer processor cores, etc.) to facilitate more energy efficient inference.

[0041] Further, using more highly expressive tokens can provide for generating query responses with machine-learned sequence processing models with an increased temporal token output rate. For example, for a given set of computational resources that is able to perform forward passes through a model of a given size, outputting a single highly-expressive token that is mapped to a set of interface token(s) with some approximate multiplier N can result in a xN speedup in the rate at which interface tokens may be returned (as may be reduced by inference time of detokenizer stack), since each forward pass may now generate xN more interface tokens.

[0042] Further, shifting the primary contextualization workload to the interface token encoder can allow for relatively cheap autoregressive inference after a relatively more expensive initialization. After the query data is tokenized into the contextual token space, subsequent generations are already in the lightweight, information-dense contextual token space, such that autoregressive generations need not pass through the full encoding stack. Furthermore, decoding the contextual tokens to an interface token space can be a relatively lightweight operation, as the information-dense representations may readily admit more simplified formats.

[0043] Some traditional techniques, in contrast, predict next token values in the output token space (e.g., words, subwords) which must then be passed all the way through the input encoding layers for contextualization to then execute the next decoding step. That is, the relatively expensive input encoder layers are executed for each decoding step.

[0044] Advantageously, allocating the computational load according to example implementations of the present disclosure can facilitate more efficient use of memory and processor availability for inference systems. For example, model parameters for the interface token encoder can be loaded and unloaded to processor memory a fewer number of times, relaxing a potential memory bandwidth bottleneck. Further, because the sequence processing model itself may be smaller, it can make better use of a given amount of high speed or processormemory by requiring fewer load / unload cycles during a session of autoregressive inference. Similarly, the lower frequency, higher cost interface token encoding can be performed by hardware optimizing for processing-constrained operations (e.g., a higher number of cores) while hardware optimized for memory bandwidth-constrained operations (e.g., larger processor caches; faster or more numerous high-bandwidth-memory lanes; etc.) can be allocated for executing the machine-learned sequence processing model - an allocation which may be more efficient for smaller machine-learned sequence processing models according to the present disclosure.Allocating hardware can include distributing the operations across different devices (e.g., performing some steps locally and other remotely).

[0045] Similarly, in implementations in which input preprocessing is performed using a separate system (e.g., separate device, separate process or route, etc.), separating input contextualization from the token generation decoding steps can reduce a number of communication steps sent between the input processing and decoding systems (e.g., communications within a machine; across a network between machines; etc.).

[0046] This efficiency advantage can be further leveraged by caching contextual token representations of reference documents for later performing inference conditioned on the same materials. For example, after tokenizing a reference document that may be later recalled (for the same user or a different user), such as a product manual, event webpage, conversation history, etc., the contextual tokens can be cached in the information-dense state for immediate injection into the model as needed without needing to store and re-encode the reference material for later processing.

[0047] In an example aspect, the present disclosure provides techniques for improved machine-learned model execution by expanding an expressive capacity of the model output space with sub-linear, sub-exponential increase in cost. Advantageously, the dense contextual tokens of the present disclosure can leverage product quantization algorithms to concisely provide coverage of an expanded vocabulary domain.

[0048] Traditional approaches to increasing an expressive capacity of a “vocabulary” for a machine-learned model have generally involved increasing a number of tokens in an output layer of the model. For example, in a textual domain, a token vocabulary can roughly correspond to individual words. Naively expanding the token vocabulary to include expressive contexts for each word effectively applies a large multiplier to vocabulary size. For instance, the same word “play” can mean, among other things, a theatrical production (“I acted in the play last night.”), casual activity characterized by having fun (“The children went outside to play ”), an amount of movement between joined parts (“There was some play in the mechanism.”), a plan for action in a portion of a sports event (“They scored on first play.”), etc. Including a different token for eachsense of each word in a language can lead to an exploding size of the output layer, as each decoding step may involve generating predicted values (e.g., logit values) for each possible token, such that the output layer would include additional parameters (and associated computations) that would be activated for each decoding step.

[0049] Advantageously, using product quantization according to an example aspect of the present disclosure, N codebooks with M tokens each can effectively leverage a vocabulary of MAN unique codes with an inference cost scaled according to only M*N predictions. In an example, N prediction heads of size M can execute to cover the full prediction space. In this manner, N*M predictions can cover the full MAN output space (in contrast to requiring MAN predictions under traditional schemes). In this manner, for instance, the present disclosure provides new architectures that enable ingesting and predicting highly expressive tokens without extreme computational demands. Such increased vocabulary sizes can enable tokens to capture not only content representations but also contextual information associated with and further describing the content. Expressing such rich content using tokens (rather than continuous embeddings) can enable a lightweight tuple of token identifiers to represent the contents of large, high-dimensional vectors of floats, integers, or other numerical representations.

[0050] Another example technical problem solved by example implementations of aspects of the present disclosure may include the challenge of training a machine-learned sequence processing model that operates on a factorized token representation having a large effective vocabulary. The disclosed contextual tokens can be factorized, meaning each token may comprise a tuple of subtokens selected from multiple independent codebooks. The effective vocabulary size can be the Cartesian product of the individual codebook sizes, which can be in the millions, billions, or trillions. A conventional language model architecture, which may use a single output prediction head (e.g., a softmax layer) over the entire vocabulary, may not be suitable for such a large space. The prediction head naively extended to that vocabulary size might require a large number of parameters, thereby hindering the practical use of such expressive, factorized tokens in some circumstances.

[0051] Example implementations of aspects of the present disclosure may provide technical solutions to this problem by configuring the architecture of the machine-learned sequence processing model to handle the factorized token space efficiently. The model can be configured with a factorized prediction head that comprises a plurality of smaller, parallel subtoken decoding lanes, where each lane corresponds to one of the subtoken codebooks. Instead of predicting a single token from a large combined vocabulary, the model can (in series or parallel) predict one subtoken from each codebook via its respective decoding lane. To facilitate parallel computation, the model may be configured such that the prediction of subtokens for a givendecoding step occurs with conditional independence. This architectural configuration can replace a single, large computation with several smaller, independent, and parallelizable computations. A technical effect may be a reduction in the number of parameters in the model’s prediction head (either in toto or as loaded into memory at a given time for a given number of queries), which in turn may reduce the memory and computational resources required for training and inference.

[0052] Improved resource efficiency can benefit both powerful server systems as well as resource-constrained devices (e.g., mobile devices, edge devices, etc.). For instance, in environments where server resources are shared across multiple clients or services, such as in cloud computing platforms or large-scale data centers, the ability to minimize resource consumption and thermal output while maintaining high responsiveness is desirable. By performing inference using information-dense, contextualized contextual tokens, the system may reduce the number of tokens processed, generated, or otherwise stored in memory for a given task. This can lead to a decrease in the computational load and memory usage on the server. This reduction may, in turn, free up resources that can be allocated to handling additional client requests or improving service responsiveness. Furthermore, in scenarios involving thin-client devices that rely on server-side processing for content interaction tasks, such as text editing or media streaming, the efficiency improvements on the server side can directly translate to faster response times and smoother interactions on the client side. This is particularly relevant for devices with limited processing capabilities or in situations where network latency is a concern. By optimizing the tokenization process, the server can process requests more quickly and return results to the client in a shorter time frame.

[0053] A technical effect of example implementations of the present disclosure is high levels of data compression. Techniques according to the present disclosure can operate to achieve 10: 1, 100: 1, 1000: 1 or higher compression ratios (uncompressed size to compressed size) by compressing contextualized parts of an input data object rather than directly compressing the raw data object itself. For instance, in lieu of directly compressing words of a text file, in which a word inherits meaning from its context, an example implementation of the present disclosure can first obtain a richly contextualized representation of that word and then map that representation into a quantized token space, such that the quantized token operates as a compressed representation of the word and its context. In this manner, for instance, the compressed object can represent not only the word but also the word’s meaning in context. In this manner, for instance, a lightweight set of token values (e.g., token identifiers) can represent complex, contextualized portions of the underlying data.

[0054] Such highly compressed values can be retrieved to recover the original content (e.g., by detokenizing according to example aspects of the present disclosure). For example, thecompressed values can be used for efficient storage of information (e.g., using less energy to persist in memory, less disk space, etc.). Such highly compressed values can be retrieved to execute inference over the values using a machine-learned sequence processing model. For example, the compressed values can be retrieved from a cache and directly input to a machine- learned sequence processing model that performs generation (e.g., autoregressive generation) of additional values in the compressed value space (e.g., directly decoding token values for representations in the quantized vector space). By retrieving the cached compressed values directly, the inference cost of tokenization into the contextualized token space may be avoided on repeated sessions. For example, a retrieval-augmented generation system can retrieve, based on a query, reference material. A version of the reference material that is already available in the contextual token representation may be accessed and input into a context window of the machine-learned sequence processing model to condition generations based on the query.

[0055] A technical effect of example implementations of the present disclosure is increased energy efficiency in performing operations using machine-learned models, thereby improving the functioning of computers implementing such models. For instance, example implementations can provide for more energy-efficient runtime execution or inference. In some scenarios, increased energy efficiency can provide for less energy to be used to perform a given task (e.g., less energy expended to maintain the model in memory, less energy expended to perform calculations within the model, etc.). In some scenarios, increased energy efficiency can provide for more task(s) to be completed for a given energy budget (e.g., a larger quantity of tasks, more complex tasks, the same task but with more accuracy or precision, etc.).

[0056] In another example aspect, example implementations can provide for more energyefficient training operations or model updates. In some scenarios, increased energy efficiency can provide for less energy to be used to perform a given number of update iterations (e.g., less energy expended to maintain the model in memory, less energy expended to perform calculations within the model, such as computing gradients, backpropagating a loss, etc.). In some scenarios, increased energy efficiency can provide for more update iterations to be completed for a given energy budget (e.g., a larger quantity of iterations, etc.). In some scenarios, greater expressivity afforded by model architectures and training techniques of the present disclosure can provide for a given level of functionality to be obtained in fewer training iterations, thereby expending a smaller energy budget. In some scenarios, greater expressivity afforded by model architectures and training techniques of the present disclosure can provide for an extended level of functionality to be obtained in a given number of training iterations, thereby more efficiently using a given energy budget.

[0057] In this manner, for instance, the improved energy efficiency of example implementations of the present disclosure can reduce an amount of pollution or other waste associated with implementing machine-learned models and systems, thereby advancing the field of machine-learning and artificial intelligence as a whole. The amount of pollution can be reduced in toto (e.g., an absolute magnitude thereof) or on a normalized basis (e.g., energy per task, per model size, etc.). For example, an amount of CO2 released (e.g., by a power source) in association with training and execution of machine-learned models can be reduced by implementing more energy-efficient training or inference operations. An amount of heat pollution in an environment (e.g., by the processors / storage locations) can be reduced by implementing more energy-efficient training or inference operations.

[0058] Various example implementations are described herein with respect to the accompanying Figures.

[0059] Figure 1 is a block diagram of an example inference system 100 using multilevel tokenization to perform inference using a machine-learned sequence processing model according to an example aspect of the present disclosure. Query data 102 can be some data designated for input to a machine-learned model for performing inference. Interface tokenizer 104 can process query data 102 to generate interface tokens 106. Interface token encoder 108 can process interface tokens 106 to generate contextualized interface token encodings 110. Contextual tokenizer 112 can process interface token encodings 110 to generate contextual tokens 114.

[0060] Machine-learned sequence processing model 116 can perform inference over contextual tokens 114 to generate predicted contextual tokens 118. Contextual detokenizer 120 can decode predicted contextual tokens 118 into predicted interface tokens 122 for interfacing with a downstream system. Response data 124 can include data obtained from predicted interface tokens 122.

[0061] Query data 102 can include various types of input data that can be processed by system 100. For example, query data 102 can include natural language text; multimedia content, such as images, audio, and video segments; structured data, such as database records, tables, graphs, or JSON objects; code sequences or other specialized text formats, such as markup languages (e.g., HTML, XML) or programming languages (e.g., Python, JavaScript). Query data 102 can include domain-specific data, such as medical records, sensor returns, etc.

[0062] Interface tokenizer 104 can process query data 102 to generate interface tokens 106 using a variety of tokenization approaches for different data modalities. For text data, for example, the tokenizer can use byte-level, character-level, or subword-level tokenization. Bytelevel tokenization can break down the text into individual bytes. Character-level tokenization can split the text into individual characters. Subword tokenization, such as Byte Pair Encoding orWordPiece, can be used to balance the granularity of tokens by splitting text into meaningful subword units, which can help in reducing the vocabulary size while still capturing semantic information. An example interface tokenizer 104 includes SentencePiece.

[0063] For image data, for example, interface tokenizer 104 can generate interface tokens 106 by dividing the image into smaller patches. Each patch can be treated as a token, allowing the model to process and analyze different parts of the image independently. Alternatively, the tokenizer can extract features from the image (e.g., using one or more CNN layers) and represent these features as tokens.

[0064] For audio data, for example, interface tokenizer 104 can generate interface tokens 106 by segmenting the audio signal into smaller time frames or windows. Each segment can be treated as a token, allowing the model to process the audio data sequentially. This approach can be useful for tasks such as speech recognition or audio event detection, where the temporal structure of the audio signal is important. Interface tokenizer 104 can use techniques like Mel- frequency cepstral coefficients (MFCCs) or spectrograms to extract features from the audio signal and represent the features in an image space, with the image data being subsequently tokenized as described above.

[0065] For video data, for example, interface tokenizer 104 can generate interface tokens 106 by dividing the video into smaller segments or frames. Each frame or segment can be treated as a token, allowing the model to process the video data sequentially. Alternatively, each frame or segment can be treated as one or more images and one or more audio segments and tokenized as described above for the respective modalities. Interface tokenizer 104 can use joint spatiotemporal processing (e.g., using 3D convolutional neural networks) to extract spatiotemporal features from the video and represent these features as tokens.

[0066] Interface tokenizer 104 can handle multimodal data by generating interface tokens 106 that represent different types of data within a single query. For example, in a scenario where both text and image data are present, the tokenizer can generate separate tokens for the text and image parts of the query. These tokens can then be processed together by the model, allowing it to analyze and integrate information from multiple modalities.

[0067] In an example, interface tokenizer 104 implements byte-level tokenization. Byte-level tokenization can include reading data as individual bytes. Byte data tokenization can be modality-agnostic. In an example, because each textual character is represented by its byte representation, a byte-level tokenizer may not need to maintain separate vocabularies for different languages. This can significantly reduce the complexity and memory requirements for handling text in multiple languages. For instance, a single byte-level tokenizer can process text in English, Chinese, and Arabic without needing separate tokenization rules or vocabularies foreach language. Similar benefits can accrue more generally because byte-level tokenization does not necessarily have a limited vocabulary, as arbitrary data can be represented in bytes.

[0068] Interface tokens 106, 122 can represent portions of input data. Interface tokens 106 can be represented with a token identifier. For instance, a token identifier can be stored in a lookup table in association with the contents associated with that token. Interface tokens 106, 122 can include portions of various data types, such as text, images, audio, and video. For instance, in a text-based application, interface tokens 106, 122 might include byte-level tokens, character-level tokens, or subword tokens. In contrast, for image processing, these tokens could be image patches, while in audio and video contexts, they might represent segments of the respective media.

[0069] In general, interface tokens 106, 122 can provide a medium for interacting with other systems. For instance, machine-learned sequence processing model 116 can perform its inference in a contextual token space, while external systems (e.g., user applications, interfaces, other software) can operate with content representations in other formats. Interface tokens 106, 122 can provide a structure for ingesting data from external sources and returning data to external sources.

[0070] Interface token encoder 108 can process interface tokens 106 to generate contextualized interface token encodings 110. Interface token encoder 108 can process interface tokens 106 to generate contextualized interface token encodings 110 for various data modalities, including text, images, audio, and video.

[0071] In general, interface token encoder 108 can include a machine-learned model that receives a token identifier or other representation of a token and outputs an encoded representation of the data communicated via that token. Interface token encoder 108 can include attention mechanisms so that the generated encoded representation is contextualized based on surrounding tokens.

[0072] In an example, interface token encoder 108 can use pretrained portions of transformer-based models. The input textual tokens can be embedded into a continuous vector space using a pre-trained embedding layer. This embedding layer may provide contextual embedding or may retrieve a stored initial, non-contextualized embedding for a given token. These initial embeddings may then be passed through multiple layers of transformer encoders, which can apply self-attention mechanisms to capture contextual relationships between the initial embeddings. The attention can be causal (e.g., leftward attention only, without dependence on future / sub sequent content in the sequence) or bidirectional (e.g., placing the embedding in context with respect to both preceding content and subsequent content). The encodings output from these transformer layers can represent encodings of the initial embeddings in context (e.g.,thereby incorporating information from other portions of the input sequence), thus providing a rich representation of each token in its context.

[0073] In an example, interface token encoder 108 can use convolutional neural networks (CNNs) to generate contextualized token encodings for input textual tokens with spatial feature dependence. For example, image patches, or initial embeddings thereof, can be processed using convolutional filters to capture local patterns and features. Pooling layers may be used to downsample the convolutional outputs, reducing the dimensionality while preserving important features. The final output from the convolutional layers can represent the contextualized token encodings or can be subsequently passed to one or more transformer blocks or other attention mechanisms.

[0074] In general, interface token encoder 108 can leverage the inherent knowledge of a pretrained model. For instance, for text data, the encoder can utilize a portion of a pre-trained language model such as one or more layers of a Gemini pre-trained model, or a Gemma pretrained model, available at https: / / github.com / google-deepmind / gemma. Such models can be pre-trained on large corpora of content and can generate high-quality token encodings that capture a wide range of semantic features. The pre-trained model can generate token encodings that reflect not only the implicit and explicit context from the query data, but also the rich contextual information learned during pre-training.

[0075] Interface token encoder 108 can be fine-tuned for specific tasks. This fine-tuning can enhance the relevance and performance of the encoded token encodings for various applications. In natural language processing (NLP) tasks, the token encoder can be fine-tuned on a large corpus of domain-specific text data. This fine-tuning process allows the model to learn the nuances and context-specific meanings of words and phrases within a specific domain.

[0076] In image processing, interface token encoder 108 can be fine-tuned using a dataset of images relevant to a specific application. Examples of such applications can include medical imaging or autonomous driving. For medical imaging tasks, the encoder can be fine-tuned on a dataset of annotated medical images, such as X-rays or MRIs. This can help the model learn to recognize and encode features more salient to diagnosing diseases. In autonomous driving, finetuning the encoder on datasets containing various driving scenarios and environmental conditions can improve the model’s ability to encode features relevant to the driving task. In general, the fine-tuning process can enhance the encoder’s ability to generate token encodings that are more relevant and informative for the specific application. This can lead to improved performance in tasks such as image classification, object detection, and segmentation.

[0077] For audio processing applications, interface token encoder 108 can be fine-tuned on a dataset of audio recordings relevant to the target task. For example, in speech recognition, fine-tuning the encoder on a dataset of spoken language data can improve transcription accuracy by training the encoder to encode features more salient for the transcription task. The dataset can include various accents, dialects, and speaking styles. This process can help the model capture the necessary acoustic features. In music genre classification, fine-tuning the encoder on a diverse set of music tracks from different genres can improve the model's ability to encode distinguishing features of each genre. In general, the fine-tuning process can enhance the encoder’s ability to generate token encodings that are more relevant and informative for the specific application. This can result in token encodings that are more effective for tasks such as speech-to-text conversion, speaker identification, etc.

[0078] In video processing, interface token encoder 108 can be fine-tuned using a dataset of videos related to the specific application. In video editing workflows, fine-tuning the encoder on a dataset of captured footage can help the model learn to recognize and encode features relevant to editing workflows (e.g., foreground subjects, background subjects, etc.). In sports analytics, fine-tuning the encoder on a dataset of sports videos can enhance the model’s ability to encode features related to player movements, game events, and team strategies. In general, the finetuning process can enhance the encoder’s ability to generate token encodings that are more relevant and informative for the specific application. This can improve the performance of tasks such as action recognition, event detection, and video summarization.

[0079] Fine-tuning the interface token encoder 108 on specific tasks can allow the model to adapt to the particular characteristics and requirements of different data modalities and applications. Utilizing domain-specific data during fine-tuning can enable the encoder to generate more relevant and informative token encodings. This can result in improved performance in various tasks across different fields.

[0080] Interface token encoder 108 may not be pre-trained. Interface token encoder 108 can be trained specifically for operation in system 100. For example, interface token encoder 108 can be trained jointly with one or more other models in system 100. System 100 as a whole can be trained end-to-end jointly.

[0081] Interface token encodings 110 can include token encodings representing the information communicated in each respective token, augmented by the context and world knowledge from interface token encoder 108. Generally, interface token encodings 110 can include vector representations in a continuous vector space. Interface token encodings 110 can have various dimensionalities.

[0082] Contextual tokenizer 112 can process interface token encodings 110 to generate contextual tokens 114. Contextual tokenizer 112 can generate contextual tokens 114 using various tokenization techniques. In general, contextual tokenizer 112 can include a machine-learned model that ingests and processes one or more interface token encodings and selects a token to represent the one or more interface token encodings. Contextual tokenizer 112 can include the machine-learned model as a preprocessor that flows into a quantization layer. The quantization layer can be used on the output to map the outputs from the preprocessor to a corresponding quantized representation.

[0083] Contextual tokenizer 112 can generate contextual tokens composed of multiple subtokens. Contextual tokenizer 112 can generate factorized contextual tokens. A factorized contextual token can include multiple subtokens that each represent independent or dependent factors. These subtokens can be selected from different subtoken codebooks. Contextual tokenizer 112 can include a machine-learned model with separate output heads designated for selecting each subtoken.

[0084] During setup of contextual tokenizer 112, the respective codebooks can be configured to represent independent factors of the tokenized content (e.g., based on a loss applied when training the tokenizer). For instance, a factor can correspond to a theme of the content. A factor can correspond to a tone or style of the content. A factor can correspond to a dialect or other attribute of the content. A factor can correspond to substantive meaning of the content.

[0085] Contextual tokenizer 112 can use product quantization to represent the contextual tokens. Each interface token encoding can be split into multiple sub-vectors, and each sub-vector can be quantized separately (e.g., using a same or different codebook).

[0086] In an example, contextual tokenizer 112 can use a clustering algorithm (e.g., k-means clustering). A clustering algorithm can partition the encoding space into clusters. Each cluster can be represented by a contextual token or subtoken. At runtime, contextual tokenizer 112 can use a distance measure to map interface token encodings (or portions thereof) to a contextual token (or a factor thereof). Each subtoken of the contextual token can correspond to a centroid of a region in an encoding space. Each subvector of the interface token encoding can be assigned to the closest centroid of a corresponding subspace. The subtokens corresponding to the centroids can be combined to compose an output contextual token.

[0087] In example, contextual tokenizer 112 can use a machine-learned classifier to classify input token encodings into groups associated with respective tokens. For example, at runtime, contextual tokenizer 112 can generate a distribution for each subtoken codebook using a machine-learned output layer. Subtokens can be selected based on these distributions (e.g., with greedy sampling, selecting the subtoken with the highest score in the distribution). Contextual tokenizer 112 can employ causal or bidirectional attention mechanisms. These mechanisms can operate over the plurality of interface token encodings 110. Causal attention can constrain theselection of each contextual token to depend only on previous tokens. Bidirectional attention can allow selection of contextual tokens considering both past and future tokens.

[0088] Contextual tokenizer 112 can be initialized using unsupervised learning techniques. Contextual tokenizer 112 can be pre-trained on a large dataset of input data and learn an efficient allocation of token representations based on a training objective (e.g., based on expected token probability of occurrence, etc.).

[0089] Contextual tokenizer 112 can be configured with a compression ratio. For instance, contextual tokenizer 112 can be configured with a hyperparameter that determines a batch size of token encodings to represent together with one contextual token.

[0090] Contextual tokens 114, 118 can be single tokens or include multiple subtokens. Contextual tokens 114, 118 can be constructed from a set of subtokens selected from corresponding subtoken codebooks. Contextual tokens 114, 118 can be stored as indices pointing to specific entries in the subtoken codebooks. For instance, each contextual token might be stored as a vector of indices, with each index corresponding to a particular subtoken in one of the codebooks.

[0091] Machine-learned sequence processing model 116 can perform inference over contextual tokens 114 to generate predicted contextual tokens 118. Machine-learned sequence processing model 116 can leverage various neural network architectures or other learned architectures that can receive an input sequence and generate an output sequence. Machine- learned sequence processing model 116 can use self-attention or cross-attention to compute predictions for selecting new contextual tokens to output in response to an input sequence of contextual tokens.

[0092] Machine-learned sequence processing model 116 can employ a transformer architecture. A transformer architecture can use self-attention mechanisms to capture dependencies across the entire sequence of contextual tokens or cross-attention mechanisms to attend into other context. Machine-learned sequence processing model 116 can use a Recurrent Neural Network or Long Short-Term Memory network.

[0093] Machine-learned sequence processing model 116 can be trained using unsupervised or supervised learning. In unsupervised learning, a training example from a training corpus can be tokenized into contextual tokens. Predicted contextual tokens for positions in the training example can be evaluated against the actual contextual tokens in those positions to compute a training loss (e.g., a classification loss in which the predicted token is the class, such as a negative log likelihood loss), and the loss can be used to train the model. In supervised learning, labeled input and output sequences of a training example can be tokenized into contextual tokens. Predicted contextual tokens responsive to the labeled input can be evaluated against thelabeled output tokens for the training example. The sequence processing model 116 can be finetuned for specific tasks or domains. Transfer learning can be applied. The model can be pretrained on a large corpus of general data. The model can then be fine-tuned on a smaller, taskspecific dataset.

[0094] Contextual detokenizer 120 can decode predicted contextual tokens 118 into predicted interface tokens 122 for interfacing with a downstream system. Contextual detokenizer 120 can include a machine-learned model configured to receive one or more token identifiers or other representations of contextual tokens and predict a corresponding set of one or more interface tokens 122 that represent the data communicated in the contextual tokens. Contextual detokenizer 120 can include various machine-learned model architectures that provide classification outputs over an available set of output classes (e.g., token identifiers) or other detokenizing outputs (e.g., for regressing continuous data from contextual token values, such as image data, audio data, etc.).

[0095] Interface detokenizer 123 can map generated interface tokens to content portions for constructing output data. In an example, a text detokenizer can map a token ID to a word or subword string and output a combined string. An image detokenizer can render a portion of an image indicated by a given token. An audio detokenizer can render a segment of audio indicated by a given token.

[0096] Response data 124 can include data obtained from predicted interface tokens 122. For instance, response data 124 can contain the contents indicated by predicted interface tokens 122. For example, predicted interface tokens 122 can contain a list of token identifiers, and response data 124 can contain the concatenated list of content stored in association with those token identifiers (e.g., subword strings).

[0097] Query data 102 and response data 124 can be received from and delivered to any variety of systems and applications via an application programming interface. For example, an assistant application can provide query data 102 to assist with performance of a task on a computing device or answering a user query. System 100 can return response data 124 as an intermediate or final result to answer the query or assist in performing the task.

[0098] Figure 2 is a block diagram of an example implementation of contextual detokenizer 120 according to example aspects of the present disclosure. Contextual detokenizer 120 can use a two-stage architecture. Contextual token decoder 202 can process predicted contextual tokens 118 to first generate predicted interface token encodings 204. Interface token encoding decoder 206 can then process predicted interface token encodings 204 to generate predicted interface tokens 122.

[0099] Contextual token decoder 202 can include a neural network configured to receive one or more contextual tokens and generate one or more token encodings in a continuous encoding space. The continuous encoding space can align with that output by interface token encoder 108.

[0100] Contextual token decoder 202 can include a causal or bidirectional decoder. A causal decoder can decode streams of latent tokens as they are available for input. A bidirectional decoder may wait to accumulate at least a buffer of future tokens (or the complete sequence of tokens) to apply bidirectional attention both over past and future tokens.

[0101] Interface token encoding decoder 206 can include a neural network configured to receive token encoding values and select or generate interface tokens to represent the data communicated in the token encoding value. Interface token encoding decoder 206 can generate a distribution for a vocabulary of available tokens using a machine-learned output layer. Interface tokens can be selected based on these distributions (e.g., with greedy sampling, selecting the token with the highest score in the distribution).

[0102] Figure 3 is a block diagram of an example implementation of machine-learned sequence processing model 116 according to example aspects of the present disclosure. Machine-learned sequence processing model 116 can include multiple output heads 300-1, 300- 2, . . . , 300-M configured to predict subtoken values 302-1, 302-2, . . . , 302-M for a predicted next contextual token 302. A head can be a model portion (e.g., one or more layers, or a portion of a layer, such as a designated channel) configured for outputting a particular output. A head can include one or more learned parameters dedicated to outputting the particular output (e.g., not shared with other output mechanisms or heads).

[0103] The multiple output heads can predict the subtoken values independently. For instance, subtoken values may not be conditioned on other predicted subtoken values. In an example, each output head receives the same context, such as all preceding contextual tokens and their constituent subtokens.

[0104] The multiple output heads can predict the subtoken values conditioned on one another. For instance, subtoken values may be conditioned on other predicted subtoken values. In an example, each output head in sequence from 1 to M receives the output from the preceding head. The multiple output heads can bidirectionally attend into one or more internal latent states within the other output heads.

[0105] The multiple output heads can predict the subtoken values conditioned on a high- level initial prediction value for content of predicted next contextual token 302. For example, a latent state in machine-learned sequence processing model 116 can include a high-dimensional numerical representation of content of predicted next contextual token 302. Each output head canprocess this high-level numerical representation of content of predicted next contextual token 302 and output the subtoken value predictions.

[0106] Nested decoding can be used to cover long range and local contexts. For instance, a high-level initial prediction value for content of predicted next contextual token 302 can be generated using full access to long-range context over the input. Machine-learned sequence generation model 116 can generate the subtoken predictions based on the high-level initial prediction value providing local context. This can reduce a computational cost of the subtoken predictions by relying on the context bottleneck in the high-level initial prediction value.

[0107] Figure 4 is a block diagram illustrating a system for performing an example nested decoding approach according to example aspects of the present disclosure. A decoding step encoding model 400 can operate at one time scale (e.g., for each step 1 . . . K) and, for each decoding step, a subtoken prediction model 402 can generate multiple subtoken values. In this manner, for instance, subtoken prediction model 402 can act as an inner loop decoder within the larger outer loop decoding of decoding step encoding model 400.

[0108] Decoding step encoding model 400 can receive a decoding step embedding input 404-1 for a decoding step at Ti. Decoding step encoding model 400 can generate for Ti a decoding step embedding output 410-1 that represents high level information for that decoding step. For Ti, subtoken prediction model 402 can generate (in parallel, autoregressively, etc.) subtoken values 412-1-1, 412-1-2, 412-1-3, etc.

[0109] Moving to a next decoding step at T2, a subtoken encoder model 404 can process the generated subtokens to generate a next decoding step encoding input 404-2. Based on this new input, decoding step encoding model 400 can generate a next decoding step encoding output 410-2. Subtoken prediction model 402 can then proceed to generate next subtoken values.

[0110] In this manner, for instance, subtoken prediction model 402 can generate multiple subtoken values for each decoding step, where one contextual token is generated at each step.

[0111] Figure 5 is a block diagram illustrating a system configuration for training one or more components of system 100 according to example aspects of the present disclosure. Figure 5 focuses on the training of the tokenization pipeline separately from the training of machine- learned sequence processing model 116. Figure 5 illustrates an example approach for unsupervised training of system components using query data 102.

[0112] Quantization loss 502 can measure a loss caused by the discretization of continuous encodings into the quantization space. Quantization loss 502 can measure a difference between a quantized representation of an encoding and the encoding itself. Quantization loss 502 can be weighted (e.g., entropy-based weighting, weighting based on expected losses, etc.).

[0113] Quantization loss 502 may not depend on input token encodings. Quantization loss 502 can be configured to balance two competing objectives. Quantization loss 502 can be configured to seek to produce confident assignments of input token encodings to clusters in the corresponding codebook(s). Quantization loss 502 can balance this first objective with a second objective that prevents a trivial solution, in which all elements of the token encodings are mapped to a single cluster. Thus, quantization loss 502 can encourage each input to be confidently assigned to one of the clusters, but also encourage mapping of the encodings across multiple clusters. Example quantization and clustering techniques are described in U.S. Patent No. 11,475,236, issued Oct. 18, 2022, from U.S. Patent Application Serial No. 16 / 880,456, filed May 21, 2020. The entire disclosure of U.S. Patent No. 11,475,236 is hereby incorporated by reference herein.

[0114] Quantization loss 502 can be used to update a codeword in a codebook used by contextual tokenizer 112 or to update a parameter that influences which input token encodings are mapped to a codeword.

[0115] Reconstruction loss 504 can provide a reconstruction loss measure that evaluates how well token encodings predicted within contextual detokenizer 120 align with original interface token encodings 110 generated by interface token encoder 108. For example, this loss can penalize information loss via the transformation into contextual tokens 114. Reconstruction loss 504 can include a distance measure or other measure of a magnitude of a comparison between a reference token encoding and a predicted token encoding associated therewith (e.g., mean squared error).

[0116] Reconstruction loss 504 can be used to train contextual token decoder 202 to improve its ability to regress continuous token encodings that fully represent contextual tokens 114. Reconstruction loss 504 can be used to train contextual tokenizer 112 to select a tokenization scheme (e.g., adjusting codewords or a learned mapping thereto) that improves retention of important information in the contextual token space.

[0117] Token loss 506 can provide a measure of how well predicted interface tokens 122 match the original interface tokens 106 (e.g., using a cross-entropy loss).

[0118] Token loss 506 can be used to train interface token encoding decoder 206 to accurately map continuous token encodings to interface tokens. To disentangle effects of the separate parts of the pipeline, interface token encoding decoder 206 can be trained based on inputting interface token encodings 110 directly to predict predicted interface tokens 122.

[0119] Figure 6 is a block diagram illustrating a system configuration for training one or more components of system 100 according to example aspects of the present disclosure. In Figure 6, an implementation is illustrated in which the tokenization pipeline is trained end-to-endto recover input query data 102. Specifically, interface token encoding decoder 206 can be trained using the outputs of the preceding layers, such that interface token encoding decoder 206 can learn to adapt and counteract any upstream errors.

[0120] Figure 7 is a block diagram illustrating a system configuration for training one or more components of system 100 according to example aspects of the present disclosure. In Figure 7, an implementation is illustrated in which the tokenization pipeline is trained end-to-end to recover input query data 102. Specifically, a single token loss 506 can be used to train the pipeline.

[0121] Interface token encoder 108 can be trained based on any of the losses described herein in the training configurations shown in Figures 5 to 7.

[0122] Figure 8 is a block diagram illustrating a system configuration for training machine- learned sequence processing model 116 according to example aspects of the present disclosure. Training data 800 can be tokenized into contextual tokens 114. Machine-learned sequence processing model 116 can process contextual tokens 114 to generate predictions 802. A loss 804 can evaluate the predictions to generate updates to model 116.

[0123] Predictions 802 can include predicted distributions over candidate subtokens (e.g., subtoken IDs). For example, a contextual token space can be configured to use M subtokens. The Mth subtoken can have a codebook of N codewords or subtoken values. Predictions 802 can include predicted scores for each subtoken for each position in a sequence of contextual tokens 114.

[0124] In normal decoding at inference time, a subtoken can be sampled based on the predicted distribution (e.g., greedy sampling, etc.) and returned as a selected subtoken for that contextual token for that decoding step. In training, the probabilities can be evaluated directly, as a “ground truth” highest probability can be imputed to be 1 for the actual subtoken in the corresponding contextual token in contextual tokens 114.

[0125] Loss 804 can evaluate a quality of a prediction by the model. Loss 804 can include a classification loss that indicates a measure of how well a prediction over candidate classes (e.g., tokens) corresponds to a ground truth class (e.g., the actual token). An example loss is a crossentropy loss or negative log likelihood loss.

[0126] Loss 804 can be backpropagated through model 116 for training model 116. In this manner, for instance, model 116 can be trained (e.g., pre-trained) in an unsupervised manner over a training corpus.

[0127] Figure 9 depicts a flowchart of a method 900 for training one or more machine- learned models according to aspects of the present disclosure. For instance, an example machine- learned model can include one or more models of interface tokenizer 104, interface tokenencoder 108, contextual tokenizer 112, machine-learned sequence processing model 116, contextual detokenizer 120, etc.

[0128] One or more portion(s) of example method 900 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 900 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 900 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 9 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 9 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 900 can be performed additionally, or alternatively, by other systems.

[0129] At 902, example method 900 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 900 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 / learning). Example data types for the training instance and various tasks associated therewith are described throughout the present disclosure.

[0130] At 904, example method 900 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.

[0131] At 906, example method 900 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 becomputed 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).

[0132] At 908, example method 900 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 b ackpropagation through time. Example method 900 can include implementing a number of generalization techniques (e.g., weight decays, dropouts, etc.) to improve the generalization capability of the models being trained.

[0133] In some implementations, example method 900 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 profile, such as based on accuracy, precision, recall, etc.).

[0134] In some implementations, example method 900 can be implemented for particular stages of a training procedure. For instance, in some implementations, example method 900 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.

[0135] In some implementations, example method 900 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)). In some implementations, example method 900 uses adapter modules. Adapters can be small trainable layers that are inserted between preexisting layers of a pre-trained model. During the fine-tuning process, the original parameters of the pre-trained model are typically frozen, and only the parameters of the adapters are updated.

[0136] In some implementations, example method 900 can be implemented to execute parameter-efficient fine-tuning methods, such as Low-Rank Adaptation (LoRA). LoRA can refine pre-trained models with minimal adjustments to the original parameters. This can be achieved by introducing trainable low-rank matrices that modify the behavior of the pre-trained weights without directly altering them. In some implementations, during fine-tuning, only these auxiliary matrices are updated, which significantly reduces the number of parameters that are trained.

[0137] An example fine-tuning approach includes reinforcement learning. Reinforcement learning can be based on user feedback on model performance during use.

[0138] Figure 10 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.

[0139] 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 non-linear 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.

[0140] Machine-learned model(s) 1 can be or include, or otherwise be representative of any one or more of the machine-learned models described above with respect to the preceding figures. For example, machine-learned model(s) 1 can be or include, or otherwise be representative of one or more models of interface tokenizer 104, interface token encoder 108, contextual tokenizer 112, machine-learned sequence processing model 116, contextual detokenizer 120, contextual token decoder 202, interface encoding decoder 206, etc. Attributes described herein with respect to machine-learned model(s) 1 are to be understood as describing any of, each of, or all of one or more models of interface tokenizer 104, interface token encoder 108, contextual tokenizer 112, machine-learned sequence processing model 116, contextual detokenizer 120, contextual token decoder 202, interface encoding decoder 206, etc.

[0141] Example neural networks can include feed-forward neural networks, recurrent neural networks, including long short-term memory based recurrent neural networks, convolutional neural networks, diffusion models, generative-adversarial networks, or other forms of neural networks. Example neural networks 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 multi-headed self-attention models.

[0142] 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 multiple different models or multiple different model portions configured to operate on data from input(s) 2.

[0143] 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, a model ensemble can include multiple models that have different attributes (e.g., different architectures, trained with different recipes, etc.). The ensemble can output an overall output based on the individual outputs of the constituent models. In this manner, for instance, the diverse constituent models can work together to provide system-level robustness by effectively aggregating over individual strengths and weaknesses of any given model. The respective individual outputs can be combined in a weighted combination, using a voting or routing mechanism, or a learned output layer (e.g., one or more feedforward or fully-connected layers).

[0144] 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). For example, different portions of a model can learn (explicitly or implicitly) different expertise areas, with pathways through the model being selected by a learned routing mechanism that engages the appropriate expert for a given input (e.g., a given portion of an input, such as on a per-token basis). For example, a feedforward network can be sparsely activated for a given portion of an input based on an output of a routing mechanism that processes the portion of the input. In this manner, for instance, the group of activated weights can form an “expert” that is selected by the router. On each forward pass, only a subset of the total model weights may be engaged, thereby decreasing a quantity of operations performed for processing a given input compared to a densely activated model. In this manner, for instance, the expressive and interpretive power of a high-parameter-count model can be achieved with more computeefficient forward passes.

[0145] 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 the same type(s) or of different types of data as compared to input(s) 2. Output(s) 3 can include one type or many different types of data.

[0146] 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 programminglanguages that use symbolic representations of machine code instructions to program a processing unit), genetic data or other chemical or biochemical data, image data, audio data, audiovisual data, haptic data, biometric data, medical data, financial 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.

[0147] 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 biometric data, sensor data and medical data, etc. It is to be understood that any combination of data types in an input 2 or an output 3 can be present.

[0148] 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.

[0149] Figure 11 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 sequence5. Input sequence 5 can include one or more input elements 5-1, 5-2, . . . , 5-A , 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, . . . , 7-7V, etc. generated based on input sequence 5. The system can generate output(s) 3 based on output sequence 7.

[0150] 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, GOOGLE, 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:2010.1 1929V2 (Jun. 3, 2021), audio domains, see, e.g., Agostinelli et a\.. Music!. : Generating Music From Text, ARXIV: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.

[0151] 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”).

[0152] 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.

[0153] Elements 5-1, 5-2, . . . , 5M 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.

[0154] 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, . . . , 5M) 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 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.

[0155] 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 11 can be the tokens or can be the embedded representations thereof.

[0156] Prediction layer(s) 6 can predict one or more output elements 7-1, 7-2, . . . , 1-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.

[0157] 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.”

[0158] 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, ARXIV: 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 or more output element(s) 7-1, 7-2, . . . , 1-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).

[0159] 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 and long short-term memory models can also be used, as well as convolutional neural networks. In general, prediction layer(s) 6 can leverage various kinds of artificial neural networks that can understand or generate sequences of information.

[0160] 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 otherinterstitial 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.

[0161] 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.

[0162] 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 regenerating the probability distribution based on the updated context window, and sampling a likely next output element, and so forth.

[0163] 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).

[0164] 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.

[0165] Figure 12 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 includeone 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.

[0166] 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 / Jdimensions. 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.

[0167] 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 in the 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 data types can be broken down into continuously defined portions (e.g., image patches) that can be described using continuously distributed locations within the embedding space.

[0168] 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, theprojection 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.

[0169] 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 a learned within a continuous embedding space.

[0170] 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).

[0171] 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.).

[0172] 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.

[0173] Figure 13 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.

[0174] 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 pre-trained 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. Model primitives 13-3 can include a library of pre-trained adapters or LoRA modules that can adapt a baseline foundational model to align its outputs with a desired performance profile, augment model capabilities (e.g., to adapt to a different input modality, etc.), and the like.

[0175] Model libraries 13 can include a tokenizer system (e.g., interface tokenizers, contextual tokenizers or detokenizers, etc.) according to the present disclosure. Model libraries 13 can include a machine-learned sequence processing model 116.

[0176] 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.

[0177] Workbench 15 can facilitate further refinement and adaptation of development model 16 by leveraging a number of different toolkits integrated with model development platform 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.

[0178] 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).

[0179] 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.

[0180] 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., de-noising, etc.) toprocess 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.

[0181] 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 fine-tune development model 16.

[0182] 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 desired runtime 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.

[0183] 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.

[0184] 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).

[0185] 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.

[0186] 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 an 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.

[0187] 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 withadditional 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.

[0188] 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 900 described above.

[0189] 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.

[0190] 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”).

[0191] 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.

[0192] 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 instruction that initiate API calls to send or obtain data via external systems.

[0193] 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.

[0194] Model development platform 12 can include a computational optimization toolkit 19 for optimizing a computational performance of development model 16. For instance, 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.

[0195] 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.

[0196] Figure 14 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 respectiveportion 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. Figure 14 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 14 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 the example training flow can be performed additionally, or alternatively, by other systems.

[0197] 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 pretraining for the same or for a different model.

[0198] Initialized model 21 can undergo pre-training in a pre-training stage 22. Pre-training 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 pretrained (e.g., development model 16 contains, is, or is based on a pre-trained foundational model or an expert model).

[0199] 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.

[0200] 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 finetuned 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. Refinementwith 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.

[0201] 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 computational optimization 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.

[0202] Figure 15 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.

[0203] 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.

[0204] 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, externalor 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.

[0205] 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.

[0206] For example, model host 31 can operate on a server system that provides a machinelearning 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.

[0207] 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.

[0208] 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 on in persistent storage, temporarily cached, or loaded into highspeed 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.

[0209] 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 includememory devices large enough to fit an entire model instance in a single 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.

[0210] 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.

[0211] 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.

[0212] 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.

[0213] 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.

[0214] Model host 31 can access a library of pre-trained adapters or LoRA modules that can adapt a baseline model to align its outputs with a desired performance profile, augment model capabilities (e.g., to adapt to a different input modality, etc.), and the like. For instance, model host 31 can receive an input request to load a customized model, and model host 31 can retrieve one or more components to adapt a baseline model to the custom profile. Model host 31 can determine that a particular functionality is needed for a particular task (e.g., based on an output of a model that preprocesses an input) and retrieve a pre-trained component accordingly.

[0215] 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 latent embedding 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.

[0216] 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.

[0217] 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 alanguage 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 a translation 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).

[0218] 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.

[0219] 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. As another example, machine- learned model(s) 1 can process the latent encoding data to generate a prediction output.

[0220] 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.

[0221] 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.

[0222] 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 or decrypting input data. In some cases, the task comprises a microprocessor performance task, such as branch prediction or memory address translation.

[0223] 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.

[0224] 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.

[0225] 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.

[0226] In some implementations, the task can be an instruction following task that is to be performed by a robot and / or autonomous vehicle (e.g., navigation task, assembly task, movement or object manipulation task, and / or any other robotic / vehicular task and the like). Machine- learned model(s) 1 can be configured to process input(s) 2 that represent instructions for the robot or autonomous vehicle 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 for controlling the robot / autonomous vehicle 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 speechor textual data (e.g., natural language instructions or spoken request for a task to be performed by the robot / autonomous vehicle) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that control the robot or autonomous vehicle to be responsive to the instruction function. The output(s) 3 may, without limitation, for example represent textual data and / or computer / machine 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 / spoken instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data and / or machine 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 controlling the robot and / or autonomous vehicle in accomplishing the requested functionality (e.g., navigation function / task, assembly function / task, movement or object manipulation function / task, and / or any other robotic / vehicular function / task and the like). For instance, an initial output can be executed by an external system or control system of the robot / vehicle or be processed by machine-learned model(s) 1 to complete an initial step of the robot / vehicle performing a function. Multiple steps can be performed for controlling the robot / vehicle, with a final output being obtained that is responsive to the initial instructions.

[0227] 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 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 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.

[0228] 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).

[0229] 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).

[0230] 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).

[0231] Figure 16 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 servercomputing 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.).

[0232] 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 16 can be co-located with, contained by, or otherwise integrated into one or more other devices or systems.

[0233] 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 provider that provides a service to an end user (who may use another computing device to interact with computing device 50).

[0234] 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.

[0235] 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.

[0236] 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 otherwise implemented by processor(s) 51.Computing device 50 can implement multiple parallel instances of machine-learned model(s) 55.

[0237] 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.

[0238] 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.

[0239] 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, ordeveloped 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.

[0240] 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 network on 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.

[0241] 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.

[0242] 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, flashmemory 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 with 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).

[0243] Figure 16 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).

[0244] Figure 17 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 17, 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.

[0245] Figure 18 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 computingdevice 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 some implementations, 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).

[0246] The central intelligence layer can include a number of machine-learned models. For example, as illustrated in Figure 18, 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.

[0247] 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 18, 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).

[0248] Figure 19 depicts a flowchart of a method 1900 for using a tokenizer according to aspects of the present disclosure.

[0249] One or more portion(s) of example method 1900 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 1900 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 1900 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 19 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 19 is described with reference to elements / terms described with respect toother systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of example method 1900 can be performed additionally, or alternatively, by other systems.

[0250] At 1902, example method 1900 can include generating, using an interface token encoder, a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence. For example, an interface token can include interface token encoder 108. For example, the plurality of interface token encodings can include interface token encodings 110. For example, the plurality of interface tokens can include interface tokens 106 generated by interface tokenizer 104.

[0251] At 1904, example method 1900 can include generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens. For example, the contextual tokenizer can include contextual tokenizer 112. An example of the one or more contextual tokens is contextual tokens 114.

[0252] At 1906, example method 1900 can include generating, using a machine-learned sequence processing model that processes the one or more contextual tokens, one or more predicted contextual tokens. For example, machine-learned sequence processing model 116 can process contextual tokens 114 to generate predicted contextual tokens 118.

[0253] At 1908, example method 1900 can include generating, using a contextual detokenizer that processes the one or more predicted contextual tokens, an output sequence comprising one or more predicted interface tokens that represent the one or more predicted contextual tokens. In an example, the contextual detokenizer can include contextual detokenizer 120. Contextual detokenizer 120 can detokenize predicted contextual tokens 118 to obtain predicted interface tokens 122.

[0254] At 1910, example method 1900 can include generating, using an interface detokenizer that processes the output sequence, a response to the query. For example, interface detokenizer 123 can detokenize a sequence of predicted interface tokens into response data 124.

[0255] In some implementations of example method 1900, the machine-learned sequence processing model generates the one or more predicted contextual tokens using self-attention over the one or more contextual tokens. For example, machine-learned sequence processing model 116 can include a self-attention block that computes relative attention values for each pairwise set of contextual tokens. A self-attention block can include a transformer-based attention architecture.

[0256] In some implementations of example method 1900, the contextual tokenizer includes a quantization layer that maps each interface token encoding to a corresponding quantizedrepresentation. For example, a quantization layer can receive an input token encoding and map the token encoding (or portions thereof) to one or more codewords in a codebook. The codebook can contain a discretized domain of representative token encoding vectors and a corresponding identifier. A contextual token can contain one or more identifiers of corresponding representative token encoding vectors.

[0257] In some implementations of example method 1900, the quantization layer performs product quantization. For example, under a product quantization approach, the input token encoding can be split into portions, and each portion can be separately quantized (e.g., using vector quantization) by mapping the portion to a codeword selected from a discretized domain of a given codebook. The codebook can be generated by clustering a slice of training token encodings across the corresponding portion (e.g., executing a k-means algorithm over the portions from the training token encodings).

[0258] In some implementations of example method 1900, each respective contextual token includes a respective plurality of subtokens respectively selected from a plurality of subtoken codebooks. A subtoken can contain an identifier that points to a representative vector for a corresponding portion of an interface token encoding.

[0259] In some implementations of example method 1900, generating, using the contextual tokenizer that processes the plurality of interface token encodings, the one or more contextual tokens includes, for each contextual token: for each respective subtoken codebook of a plurality of subtoken codebooks: generating, using a machine-learned output layer of the contextual tokenizer, a respective distribution over the respective subtoken codebook; and selecting a respective subtoken value based on the generated distribution. For example, a classification network can classify an input token encoding (or a portion thereof) as associated with a particular codeword in a codebook. The classification network can have an output head that predicts a distribution over the token identifiers (e.g., a softmax). An identifier can be selected based on having a highest value.

[0260] In some implementations of example method 1900, the respective distributions for each respective subtoken codebook are generated independently. For example, the machine- learned output layer can include a plurality of output heads that independently predict the respective distributions.

[0261] In some implementations of example method 1900, the contextual tokenizer uses causal attention over the plurality of interface token encodings to select each respective contextual token. For example, the contextual tokenizer can include one or more machine- learned layers that apply transformations to the plurality of interface token encodings. The one or more machine-learned layers can execute attention operations that, for a given internal stateassociated with a given interface token encoding, attends over input token encodings that precede the given interface token encoding in a sequence of interface token encodings.

[0262] In some implementations of example method 1900, the contextual tokenizer uses bidirectional attention over the plurality of interface token encodings to select each respective contextual token. The one or more machine-learned layers can execute attention operations that, for a given internal state associated with a given interface token encoding, attends over input token encodings that precede and follow the given interface token encoding in a sequence of interface token encodings.

[0263] In some implementations of example method 1900, the contextual tokenizer maps a batch of one or more interface token encodings to a respective contextual token. For example, multiple interface token encodings can be batched for representation by a single contextual token. The batch ratio can define a quantity of interface token encodings per contextual token. The batch ratio can influence a compression ration of the tokenization step. One or more interface token encodings can be mapped to a contextual token using a distance measure. For instance, a distance between a token encoding (or a portion thereof) and a centroid of a cluster (e.g., a representative codeword in a codebook) can be used to map the token encoding to the cluster.

[0264] In some implementations of example method 1900, each subtoken of the respective contextual token corresponds to a centroid of a region in an encoding space. For instance, the subtoken can contain an identifier that points to the centroid vector as a representative vector for the region.

[0265] In some implementations of example method 1900, the machine-learned sequence processing model includes a factorized prediction head that includes a plurality of subtoken decoding lanes respectively for a plurality of subtoken codebooks. For example, the factorized prediction head can include a plurality of output pathways (e.g., output prediction layers that each cover a respective codebook) that independently predict the respective distributions.

[0266] In some implementations of example method 1900, the factorized prediction head, for a given decoding step, outputs predicted subtokens with conditional independence among the predicted subtokens for the decoding step.

[0267] In some implementations of example method 1900, a predicted subtoken from a first subtoken decoding lane is, for a given decoding step, conditioned on a subtoken from a second subtoken decoding lane predicted at a prior decoding step. For example, in some instances, subtokens within each lane are predicted independently with respect to a current time step but conditioned on multiple different subtoken lane outputs from prior time steps.

[0268] In some implementations of example method 1900, the interface token encoder includes one or more encoder layers extracted from a pretrained language processing model. For example, a pretrained language processing model can include one or more encoder layers that allow information to propagate across input locations. For instance, the encoder layers can include a feedforward network, a transformer block, etc.

[0269] In some implementations of example method 1900, the interface token encoder is frozen during training of the contextual detokenizer and the machine-learned sequence processing model.

[0270] In some implementations of example method 1900, the contextual detokenizer and the machine-learned sequence processing model are trained end-to-end based on an evaluation of the output sequence. In some implementations of example method 1900, the contextual detokenizer and the machine-learned sequence processing model are trained end-to-end based on an evaluation of the one or more predicted interface tokens. For example, the output sequence can include token predictions that are compared to reference token values to compute a reconstruction loss for training the contextual detokenizer and the machine-learned sequence processing model.

[0271] In some implementations of example method 1900, the contextual detokenizer includes a first stage that generates one or more predicted interface token encodings for each predicted contextual token and a second stage that generates the one or more predicted interface tokens based on the one or more predicted interface token encodings. For example, the first stage can include a contextual token decoder 202 that regresses continuous-valued token encodings in an token encoding space compatible with an interface token decoder 206 that maps the regressed continuous-valued token encodings to discrete tokens (e.g., by comparison to precomputed token embeddings, such as non-contextualized embeddings).

[0272] In some implementations of example method 1900, the first stage and the second stage are trained jointly using a classification loss applied to the second stage. A classification loss can include a token loss computed to evaluate an error in predicting a correct token based on a known reference token for a given position in the output sequence.

[0273] In some implementations of example method 1900, the first stage is trained using a reconstruction loss (e.g., to encourage the first stage to reconstruct a reference embedded value output by the interface token encoder) and the second stage is trained using a classification loss (e.g., to encourage the second stage to correctly classify a given token encoding to the correct token).

[0274] In some implementations of example method 1900, the first stage is trained using a reconstruction loss (e.g., to reconstruct, from the contextual tokens, the original interface tokenencodings input to the contextual tokenizer). In some implementations of example method 1900, the second stage is trained over a set of ground truth input token encodings (e.g., in lieu of using the reconstructed token encodings) using a classification loss. In this manner, for instance, the training can be at least partially disentangled.

[0275] In some implementations of example method 1900, the first stage is trained using a reconstruction loss and a quantization loss.

[0276] In some implementations, example method 1900 includes streaming the response while concurrently generating, using the machine-learned sequence processing model that processes the one or more contextual tokens, one or more additional predicted contextual tokens. For example, streaming the response can include transmitting one or more decoded interface tokens to a recipient interface (e.g., an API) while further decoding of additional contextual and interface tokens. For example, in an interface user interface, streamed text can provide a low- latency initiation of interaction (e.g., in lieu of holding back the full response until complete). In a filtering or controlled decoding context, streamed content (e.g., text, or otherwise) can be filtered or otherwise post-processed in real time to decrease a total latency of serial processes. For example, decoded tokens can be streamed to a queue, and a second serial process can retrieve the tokens from the queue for further processing in a different thread, using different processor cores, etc.

[0277] In some implementations, example method 1900 includes caching one or more of the contextual tokens. For example, the cached contextual tokens can be stored for later use to perform inference over the data encoded in the contextual tokens.

[0278] In some implementations, example method 1900 includes retrieving the cached one or more contextual tokens. In some implementations, example method 1900 includes inputting the retrieved cached one or more contextual tokens to the machine-learned sequence processing model to generate one or more additional predicted contextual tokens.

[0279] In some implementations of example method 1900, the cached one or more contextual tokens were initially generated for processing a first query and are retrieved for processing a second query.

[0280] Figure 20 depicts a flowchart of a method 2000 for training a tokenizer according to aspects of the present disclosure.

[0281] One or more portion(s) of example method 2000 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 2000 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 2000 can be implemented on the hardware components ofthe device(s) described herein, for example, to train one or more systems or models. Figure 20 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 20 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 2000 can be performed additionally, or alternatively, by other systems.

[0282] At 2002, example method 2000 can include generating, using an interface token encoder, a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence. For example, an interface token can include interface token encoder 108. For example, the plurality of interface token encodings can include interface token encodings 110. For example, the plurality of interface tokens can include interface tokens 106 generated by interface tokenizer 104.

[0283] At 2004, example method 2000 can include generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens, wherein the contextual tokenizer includes a quantization layer that maps each interface token encoding to a corresponding quantized representation. For example, the contextual tokenizer can include contextual tokenizer 112. An example of the one or more contextual tokens is contextual tokens 114.

[0284] At 2006, example method 2000 can include generating, using a contextual detokenizer that processes the one or more contextual tokens, one or more decoded interface token encodings that represent the one or more contextual tokens. In some implementations of example method 2000, the contextual detokenizer includes a first stage that generates one or more decoded interface token encodings for each contextual token and a second stage that generates one or more decoded interface tokens based on the one or more decoded interface token encodings. For example, the first stage can include a contextual token decoder 202 that regresses continuous-valued token encodings in an encoding space compatible with an interface token decoder 206 that maps the regressed continuous-valued token encodings to discrete tokens (e.g., by comparison to precomputed token encodings).

[0285] At 2008, example method 2000 can include training the contextual detokenizer using a loss function that computes a reconstruction loss between a decoded interface token encoding and a corresponding interface token encoding. For example, encoding loss 504 can include areconstruction loss that measures a difference between the decoded interface token encodings and the original interface token encodings input to the contextual tokenizer.

[0286] In some implementations of example method 2000, the loss function computes a quantization loss for the quantization layer. A quantization loss (e.g., loss 502) can measure a contribution of the quantization layer to an overall loss. For instance, a quantization loss can measure a difference between a quantized representation and the underlying encoding represented by that quantized representation.

[0287] In some implementations, example method 2000 includes generating, using a second stage of the contextual detokenizer, one or more decoded interface tokens (e.g., interface tokens 122) that represent the one or more decoded interface token encodings.

[0288] In some implementations, example method 2000 includes training the second stage of the contextual detokenizer using a second loss function that computes a classification loss based on a comparison of the decoded interface tokens and the input interface tokens.

[0289] In some implementations, example method 2000 includes generating, using a second stage of the contextual detokenizer, one or more decoded interface tokens based on one or more ground truth interface token encodings that are respectively associated with one or more ground truth interface token labels. For example, a ground truth interface token encoding can be input to the second stage to predict a token label or identifier. In some implementations of example method 2000, the method includes training the second stage of the contextual detokenizer using a second loss function that computes a classification loss based on a comparison of the decoded interface tokens and the one or more ground truth interface token labels.

[0290] In some implementations, example method 2000 includes initializing the contextual tokenizer using a clustering algorithm applied over the input encoding space. For example, clustering can be performed for each desired codebook. For a product quantization approach, an encoding space can be sliced into separate portions, and each portion can be clustered separately. In an example, a plurality of training token encodings can be sliced into portions. The slicing can be hand-tuned (e.g., subdividing 1024-length vector into a fixed 16 portions) or learned (e.g., jointly learning the vector subdivision size or positions while clustering to reduce a quantization or reconstruction error). Clustering can be performed over each group of corresponding slices to obtain a codebook for that group.

[0291] In some implementations, example method 2000 includes training the contextual tokenizer jointly with the contextual detokenizer using the loss function.

[0292] In some implementations of example method 2000, the interface token encoder is frozen.

[0293] In some implementations of example method 2000, the interface token encoder is multimodal.

[0294] In some implementations of example method 1900 or 2000, the interface tokens comprise byte-level tokens. In some implementations of example method 1900 or 2000, the interface tokens comprise character-level tokens. In some implementations of example method 1900 or 2000, the interface tokens comprise subword tokens. In some implementations of example method 1900 or 2000, the interface tokens comprise image patches. In some implementations of example method 1900 or 2000, the interface tokens comprise audio segments. In some implementations of example method 1900 or 2000, the interface tokens comprise video segments. In some implementations of example method 1900 or 2000, the interface tokens comprise tokens selected from a token vocabulary constructed for an input data modality.

[0295] In some implementations, example method 2000 includes freezing the contextual tokenizer and the contextual detokenizer. In some implementations, example method 2000 includes training a machine-learned sequence processing model to generate predicted contextual tokens, wherein training the machine-learned sequence processing model includes unsupervised training over contextual tokens generated by the contextual tokenizer for a training dataset. An example method for training the machine-learned sequence processing model is provided below with respect to Figure 21.

[0296] Figure 21 depicts a flowchart of a method 2100 for training a machine-learned sequence model according to aspects of the present disclosure.

[0297] One or more portion(s) of example method 2100 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 2100 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 2100 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 21 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 21 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 2100 can be performed additionally, or alternatively, by other systems.

[0298] At 2102, example method 2100 can include generating, using an interface token encoder (e.g., interface token encoder 108), a plurality of training interface token encodings respectively for a plurality of training interface tokens of a training data sequence.

[0299] At 2104, example method 2100 can include generating, using a contextual tokenizer (e.g., contextual tokenizer 112), a sequence of one or more training contextual tokens representing the plurality of training interface tokens.

[0300] At 2106, example method 2100 can include generating, using a machine-learned sequence processing model (e.g., model 116) that processes the one or more training contextual tokens, a contextual token prediction that indicates a likelihood associated with a contextual token for a particular location in the sequence of the one or more training contextual tokens.

[0301] At 2108, example method 2100 can include backpropagating a token prediction loss through the machine-learned sequence processing model to compute a model update. In some implementations, the token prediction loss can be based on a classification loss for the contextual token prediction evaluated using a reference contextual token present in the sequence of the one or more training contextual tokens at the particular location. Example predictions 802 illustrate an example implementation of loss 804 in Figure 8.

[0302] At 2108, example method 2100 can include updating, using the model update, the machine-learned sequence processing model.

[0303] Figure 22 depicts a flowchart of a method 2200 for compressing data according to aspects of the present disclosure.

[0304] One or more portion(s) of example method 2200 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 2200 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 2200 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 22 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 22 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 2200 can be performed additionally, or alternatively, by other systems.

[0305] At 2202, example method 2200 can include generating, using an interface token encoder (e.g., the interface token encoder of any of the preceding claims), a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence.

[0306] At 2204, example method 2200 can include generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens. The contextual tokenizer can include a quantization layer that maps each interface token encoding to a corresponding quantized representation.

[0307] At 2206, example method 2200 can include storing, in a computer-readable storage medium, a compressed representation of the plurality of interface tokens, wherein the compressed representation includes the one or more contextual tokens.

[0308] In some implementations of example method 2200, the contextual tokenizer includes a contextual tokenizer (e.g., contextual tokenizer 112). The contextual tokenizer can be trained according to example method 2000.

[0309] In some implementations, example method 2200 includes retrieving the compressed representation from the computer-readable storage medium. In some implementations, example method 2200 includes inputting the one or more contextual tokens of the compressed representation to a machine-learned sequence processing model to generate one or more predicted contextual tokens. In some implementations, example method 2200 includes generating, using a contextual detokenizer that processes the one or more predicted contextual tokens, an output sequence comprising one or more predicted interface tokens that represent the one or more predicted contextual tokens. In some implementations, example method 2200 includes generating, using an interface detokenizer that processes the output sequence, a response to a query associated with the input data sequence.

[0310] 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. In further examples, there is provided a computer program product including computer-executable instructions which, when executed by at least one computing apparatus (or one or more processors), cause the at least one computing apparatus (or one or more processors) to perform one or more of the method(s) or process(es) described with reference to the Figures and as described herein.

[0311] While the present subject matter has been described in detail with respect to various specific example embodiments or implementations 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 or described 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.

[0312] 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.”

[0313] 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.

[0314] 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 understoodthat, 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 computer-implemented method for generating query responses with machine-learned sequence processing models with a decreased token count, the method comprising: generating, using an interface token encoder, a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence corresponding to a query; generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens; generating, using a machine-learned sequence processing model that processes the one or more contextual tokens, one or more predicted contextual tokens; generating, using a contextual detokenizer that processes the one or more predicted contextual tokens, an output sequence comprising one or more predicted interface tokens that represent the one or more predicted contextual tokens; and generating, using an interface detokenizer that processes the output sequence, a response to the query.

2. The computer-implemented method of claim 1, wherein the machine-learned sequence processing model generates the one or more predicted contextual tokens using self-attention over the one or more contextual tokens.

3. The computer-implemented method of any preceding claim, wherein the contextual tokenizer comprises a quantization layer that maps each interface token encoding to a corresponding quantized representation.

4. The computer-implemented method of any preceding claim (e.g., claim 3), wherein the quantization layer performs product quantization.

5. The computer-implemented method of any preceding claim (e.g., claim 4), wherein each respective contextual token comprises a respective plurality of subtokens respectively selected from a plurality of subtoken codebooks.

6. The computer-implemented method of any preceding claim (e.g., claim 5), wherein generating, using the contextual tokenizer that processes the plurality of interface token encodings, the one or more contextual tokens comprises:for each contextual token: for each respective subtoken codebook of a plurality of subtoken codebooks: generating, using a machine-learned output layer of the contextual tokenizer, a respective distribution over the respective subtoken codebook; and selecting a respective subtoken value based on the generated distribution.

7. The computer-implemented method of any preceding claim (e.g., claim 6), wherein the respective distributions for each respective subtoken codebook are generated independently.

8. The computer-implemented method of any preceding claim (e.g., claim 6), wherein the contextual tokenizer uses causal attention over the plurality of interface token encodings to select each respective contextual token.

9. The computer-implemented method of any preceding claim (e.g., claim 6), wherein the contextual tokenizer uses bidirectional attention over the plurality of interface token encodings to select each respective contextual token.

10. The computer-implemented method of any preceding claim (e.g., claim 5), wherein the contextual tokenizer maps a batch of one or more interface token encodings to a respective contextual token using a distance measure.

11. The computer-implemented method of any preceding claim (e.g., claim 10), wherein each subtoken of the respective contextual token corresponds to a centroid of a region in an encoding space.

12. The computer-implemented method of any preceding claim, wherein the machine-learned sequence processing model comprises a factorized prediction head that comprises a plurality of subtoken decoding lanes respectively for a plurality of subtoken codebooks.

13. The computer-implemented method of any preceding claim (e.g., claim 12), wherein the factorized prediction head, for a given decoding step, outputs predicted subtokens with conditional independence among the predicted subtokens for the decoding step.

14. The computer-implemented method of any preceding claim (e.g., claim 13), wherein a predicted subtoken from a first subtoken decoding lane is, for a given decoding step, conditioned on a subtoken from a second subtoken decoding lane predicted at a prior decoding step.

15. The computer-implemented method of any preceding claim, wherein the interface token encoder comprises one or more encoder layers extracted from a pretrained language processing model.

16. The computer-implemented method of any preceding claim, wherein the interface token encoder is frozen during training of the contextual detokenizer and the machine-learned sequence processing model.

17. The computer-implemented method of any preceding claim, wherein the contextual detokenizer and the machine-learned sequence processing model are trained end-to-end based on an evaluation of the output sequence or the one or more predicted interface tokens.

18. The computer-implemented method of any preceding claim, wherein the contextual detokenizer comprises a first stage that generates one or more predicted interface token encodings for each predicted contextual token and a second stage that generates the one or more predicted interface tokens based on the one or more predicted interface token encodings.

19. The computer-implemented method of any preceding claim (e.g., claim 18), wherein the first stage and the second stage are trained jointly using a classification loss applied to the second stage.

20. The computer-implemented method of any preceding claim (e.g., claim 18), wherein the first stage is trained using a reconstruction loss and the second stage is trained using a classification loss.

21. The computer-implemented method of any preceding claim (e.g., claim 18), wherein the first stage is trained using a reconstruction loss and the second stage is trained over a set of ground truth input token encodings using a classification loss.

22. The computer-implemented method of any preceding claim (e.g., claim 18), wherein the first stage is trained using a reconstruction loss and a quantization loss.

23. The computer-implemented method of any preceding claim, comprising: streaming the response while concurrently: generating, using the machine-learned sequence processing model that processes the one or more contextual tokens, one or more additional predicted contextual tokens.

24. The computer-implemented method of any preceding claim, comprising: caching one or more of the contextual tokens.

25. The computer-implemented method of any preceding claim (e.g., claim 24), comprising: retrieving the cached one or more contextual tokens; and inputting the retrieved cached one or more contextual tokens to the machine-learned sequence processing model to generate one or more additional predicted contextual tokens.

26. The computer-implemented method of any preceding claim (e.g., claim 25), wherein the cached one or more contextual tokens were initially generated for processing a first query and are retrieved for processing a second query.

27. A computer-implemented method, the method comprising: generating, using an interface token encoder, a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence; generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens, wherein the contextual tokenizer comprises a quantization layer that maps each interface token encoding to a corresponding quantized representation; generating, using a contextual detokenizer that processes the one or more contextual tokens, one or more decoded interface token encodings that represent the one or more contextual tokens; and training the contextual detokenizer using a loss function that computes a reconstruction loss between a decoded interface token encoding and a corresponding interface token encoding.

28. The computer-implemented method of any preceding claim (e.g., claim 27), wherein the loss function computes a quantization loss for the quantization layer.

29. The computer-implemented method of any preceding claim (e.g., claim 27 or 28), wherein a first stage of the contextual tokenizer generates the one or more decoded interface token encodings, and wherein the first stage is trained using the loss function, and wherein the method comprises: generating, using a second stage of the contextual detokenizer, one or more decoded interface tokens based on the one or more decoded interface token encodings; and training the second stage of the contextual detokenizer using a second loss function that computes a classification loss based on a comparison of the decoded interface tokens and the input interface tokens.

30. The computer-implemented method of any preceding claim (e.g., claim 27 or 28), wherein a first stage of the contextual tokenizer generates the one or more decoded interface token encodings, and wherein the first stage is trained using the loss function, and wherein the method comprises: generating, using a second stage of the contextual detokenizer, one or more decoded interface tokens that represent one or more ground truth interface token encodings that are respectively associated with one or more ground truth interface token labels; and training the second stage of the contextual detokenizer using a second loss function that computes a classification loss based on a comparison of the decoded interface tokens and the one or more ground truth interface token labels.

31. The computer-implemented method of any preceding claim, comprising: initializing the contextual tokenizer using a clustering algorithm applied over the encoding space.

32. The computer-implemented method of any preceding claim, comprising: training the contextual tokenizer jointly with the contextual detokenizer using the loss function.

33. The computer-implemented method of any preceding claim, wherein the interface token encoder is frozen.

34. The computer-implemented method of any preceding claim, comprising: freezing the contextual tokenizer and the contextual detokenizer; andtraining a machine-learned sequence processing model to generate predicted contextual tokens, wherein training the machine-learned sequence processing model comprises unsupervised training over contextual tokens generated by the contextual tokenizer for a training dataset.

35. The computer-implemented method of any preceding claim, comprising: freezing the contextual tokenizer and the contextual detokenizer; and training a machine-learned sequence processing model to generate predicted contextual tokens, wherein training the machine-learned sequence processing model comprises: generating, using the interface token encoder, a plurality of training interface token encodings respectively for a plurality of training interface tokens of a training data sequence; generating, using the contextual tokenizer, a sequence of one or more training contextual tokens representing the plurality of training interface tokens; generating, using the machine-learned sequence processing model that processes the one or more training contextual tokens, a contextual token prediction that indicates a likelihood associated with a contextual token for a particular location in the sequence of the one or more training contextual tokens; backpropagating a token prediction loss through the machine-learned sequence processing model to compute a model update, wherein the token prediction loss is based on a classification loss for the contextual token prediction evaluated using a reference contextual token present in the sequence of the one or more training contextual tokens at the particular location; and updating, using the model update, the machine-learned sequence processing model.

36. A computer-implemented method, comprising: training a machine-learned sequence processing model to generate predicted contextual tokens, wherein training the machine-learned sequence processing model comprises: generating, using an interface token encoder, a plurality of training interface token encodings respectively for a plurality of training interface tokens of a training data sequence; generating, using a contextual tokenizer, a sequence of one or more training contextual tokens representing the plurality of training interface tokens; generating, using the machine-learned sequence processing model that processes the one or more training contextual tokens, a contextual token prediction that indicates alikelihood associated with a contextual token for a particular location in the sequence of the one or more training contextual tokens; backpropagating a token prediction loss through the machine-learned sequence processing model to compute a model update, wherein the token prediction loss is based on a classification loss for the contextual token prediction evaluated using a reference contextual token present in the sequence of the one or more training contextual tokens at the particular location; and updating, using the model update, the machine-learned sequence processing model.

37. A computer-implemented method, the method comprising: generating, using an interface token encoder (e.g., the interface token encoder of any of the preceding claims), a plurality of interface token encodings respectively for a plurality of interface tokens of an input data sequence; generating, using a contextual tokenizer that processes the plurality of interface token encodings, one or more contextual tokens representing the plurality of interface tokens, wherein the contextual tokenizer comprises a quantization layer that maps each interface token encoding to a corresponding quantized representation; and storing, in a computer-readable storage medium, a compressed representation of the plurality of interface tokens, wherein the compressed representation comprises the one or more contextual tokens.

38. The computer-implemented method of claim 35 or 36, wherein the contextual tokenizer comprises a contextual tokenizer trained according to any of claims 27 to 35.

39. The computer-implemented method of claim 37 or 38, comprising: retrieving the compressed representation; inputting the one or more contextual tokens of the compressed representation to a machine-learned sequence processing model to generate one or more predicted contextual tokens; generating, using a contextual detokenizer (e.g., the contextual tokenizer of any of the preceding claims) that processes the one or more predicted contextual tokens, an output sequence comprising one or more predicted interface tokens that represent the one or more predicted contextual tokens; andgenerating, using an interface detokenizer that processes the output sequence, a response to a query associated with the input data sequence.

40. The computer-implemented method of any preceding claim, wherein the interface token encoder is multimodal.

41. The computer-implemented method of any preceding claim, wherein the interface tokens comprise: byte-level tokens; character-level tokens; subword tokens; image patches; audio segments; video segments; or tokens selected from a token vocabulary constructed for an input data modality.

42. 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 the method of any of the preceding claims.

43. A computing system, comprising: one or more processors; and one or more non-transitory computer-readable media storing instructions that are executable by the one or more processors to cause the computing system to perform operations, the operations comprising the method of any of the preceding claims.

44. A computer program product comprising instructions that are executable by one or more processors to cause a computing system to perform operations, the operations comprising the method of any of the preceding claims.

Images