Parallel decoding for sequence processing models

Abstract

Provided are machine learning processes and machine-learned devices and systems. More particularly, provided are systems and methods which leverage non-autoregressive (NAR) sequence processing models to perform parallel decoding with improved computational efficiency.

Description

PARALLEL DECODING FOR SEQUENCE PROCESSING MODELSRELATED APPLICATIONS

[0001] The present application is based on and claims priority to United States Provisional Application Number 63 / 703,766 having a filing date of October 4, 2024. The present application claims priority to and the benefit of each of such applications and incorporates all such applications herein by reference in their entirety.FIELD

[0002] The present disclosure relates generally to machine learning processes and machine-learned devices and systems. More particularly, the present disclosure relates to systems and methods which leverage non-autoregressive (NAR) sequence processing models to perform parallel decoding with improved computational efficiency.BACKGROUND

[0003] In the field of sequence processing models such as, for example, large language models (LLMs), the efficiency of sequence generation and processing poses significant technical challenges. Traditional autoregressive (AR) models, while capable of generating high-quality output sequences (e.g., sequences of output text tokens), require autoregressive processing where each token is predicted one-by-one based on the previously- generated tokens (e.g., including previously-predicted output tokens). This sequential dependency results in considerable computational overhead and latency, as the full context for each token must be processed as each output token is predicted individually, leading to increased time and resource consumption, particularly in real-time applications.

[0004] Moreover, the inherent design of AR models means that any errors in early token predictions propagate and magnify throughout the subsequent sequence generation, potentially degrading the overall quality and coherence of the generated text. This error propagation issue is exacerbated in longer sequences, where early inaccuracies can lead to significant deviations from preferred outputs.

[0005] Additionally, AR models’ reliance on their entire historical output for generating each subsequent token restricts their ability to effectively incorporate bidirectional context or make dynamic adjustments based on the full scope of the input sequence. Thislimitation is particularly problematic in tasks requiring high levels of contextual awareness and adaptability, such as interactive text editing, where the model must dynamically adjust to changes in the input text sequence.SUMMARY

[0006] A system of one or more computers can be configured to perform particular operations or actions by virtue of having software, firmware, hardware, or a combination of them installed on the system that in operation causes or cause the system to perform the actions. One or more computer programs can be configured to perform particular operations or actions by virtue of including instructions that, when executed by data processing apparatus, cause the apparatus to perform the actions.

[0007] One general aspect includes a computing system configured to perform sequence prediction with improved computational efficiency. The computing system also includes one or more processors. The system also includes one or more non-transitory computer readable media that collectively store computer-executable instructions for performing operations. The operations can include obtaining, by the computing system, a model input. The operations can include iteratively performing, by the computing system, a plurality of decoding iterations with a non-autoregressive sequence processing model to generate a model output based on the model input. The operations can include where, at each decoding iteration, the non-autoregressive sequence processing model processes a plurality of input tokens to generate a plurality of output tokens in parallel. The operations can include where, for one or more intermediate decoding iterations of the plurality of decoding iterations, the plurality of output tokens may include one or more edit tokens. The operations can include where performing, by the computing system, each of the one or more intermediate decoding iterations may include generating, by the computing system, the plurality of input tokens for a subsequent decoding iteration of the plurality of decoding iterations based on the one or more edit tokens. The operations can include providing, by the computing system, the model output as an output. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the methods.

[0008] Implementations may include one or more of the following features. The computing system of any preceding claim, where, for each of the one or more intermediate decoding iterations, generating the plurality of input tokens for the subsequent decoding iteration of the plurality of decoding iterations based on the one or more edit tokens mayinclude editing the plurality of output tokens for the current decoding iteration based on the one or more edit tokens to generate the plurality of input tokens for the subsequent decoding generation. At least one of the one or more edit tokens may include a delete token; and editing the plurality of output tokens based on the one or more edit tokens may include removing the delete token from the plurality of output tokens. At least one of the one or more edit tokens may include an insert token; and editing the plurality of output tokens based on the one or more edit tokens may include: inserting two or more mask tokens into the plurality of output tokens at positions adjacent to the insert token within the plurality of output tokens; and replacing the insert token with a corresponding input token located at a corresponding position within the plurality of input tokens. The insert token may include a value equal to an integer number within a range from two to a maximum insertion value; and inserting two or more mask tokens into the plurality of output tokens at positions adjacent to the insert token may include inserting the integer number of mask tokens into the plurality of output tokens at positions adjacent to the insert token. Iteratively performing, by the computing system, the plurality of decoding iterations may include iteratively performing, by the computing system, the plurality of decoding iterations until the plurality of output tokens for one of the decoding iterations contains only regular tokens. The one or more non-transitory computer-readable media further store the non-autoregressive sequence prediction model. The non- autoregressive sequence prediction model may include a decoder-only transformer model. The plurality of input tokens and the plurality of output tokens for each decoding iteration may include textual tokens. The plurality of input tokens and the plurality of output tokens for each decoding iteration may include image tokens or audio tokens. Iteratively performing, by the computing system, the plurality of decoding iterations may include iteratively performing, by the computing system, the plurality of decoding iterations until the plurality of output tokens contain an end of sequence token. For each of the one or more intermediate decoding iterations, generating the plurality of input tokens for the subsequent decoding iteration of the plurality of decoding iterations based on the one or more edit tokens may include editing the plurality of input tokens for the current decoding iteration based on the one or more edit tokens to generate the plurality of input tokens for the subsequent decoding generation. At least one of the one or more edit tokens may include a delete token; and editing the plurality of input tokens based on the one or more edit tokens may include removing from the plurality of input tokens a corresponding input token located at a corresponding position to the delete token within the plurality of input tokens. At least one of the one or more edit tokens may include an insert token; and editing the plurality of inputtokens based on the one or more edit tokens may include: inserting two or more mask tokens into the plurality of input tokens at positions adjacent to a corresponding position to the insert token within the plurality of input tokens. The insert token may include a value equal to an integer number within a range from two to a maximum insertion value; and inserting two or more mask tokens into the plurality of input tokens at positions adjacent to the corresponding position may include inserting the integer number of mask tokens into the plurality of input tokens at positions adjacent to the corresponding position. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer- accessible medium.

[0009] One general aspect includes a computer-implemented method for training a non-autoregressive sequence prediction model. The computer-implemented method includes obtaining, by a computing system may include one or more computing devices, a training pair may include a plurality of input tokens and plurality of target tokens, where the plurality of target tokens may include one or more edit tokens that indicate edits to be made to the plurality of input tokens to generate a corrected sequence of tokens. The method also includes processing, by the computing system, the plurality of input tokens with the non- autoregressive sequence prediction model to generate a plurality of predicted tokens. The method also includes evaluating, by the computing system, a loss function that compares the plurality of predicted tokens to the plurality of target tokens. The method also includes modifying, by the computing system, one or more values of one or more parameters of the non-autoregressive sequence prediction model based on the loss function. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the methods.

[0010] Implementations may include one or more of the following features. The computer-implemented method where obtaining, by the computing system, training pair may include constructing, by the computing system, the training pair by performing operations may include: obtaining, by the computing system, a ground truth sequence of tokens; corrupting, by the computing system, one or more of the tokens contained in the ground truth sequence of tokens to generate a corrupted sequence of tokens; and determining, by the computing system, an optimal edit to reconstruct the ground truth sequence of tokens from the corrupted sequence of tokens; where, for the training pair, the corrupted sequence of tokens serves as the plurality of input tokens, and the plurality of target tokens are based on the optimal edit. Corrupting, by the computing system, one or more of the tokens contained inthe ground truth sequence of tokens to generate the corrupted sequence of tokens may include one or more of: dropping one or more of the ground truth sequence of tokens; inserting one or more tokens into the ground truth sequence of tokens; flipping one or more of the ground truth sequence of tokens; and masking one or more of the ground truth sequence of tokens. Determining, by the computing system, the optimal edit to reconstruct the ground truth sequence of tokens from the corrupted sequence of tokens may include performing a dynamic programming algorithm to identify the optimal edit that contains a minimal number of operations required to minimize a Levenshtein distance to the ground truth sequence of tokens. Obtaining, by the computing system, training pair may include: performing, by the computing system, a training forward pass on a different training input with the non- autoregressive sequence prediction model to generate a different plurality of predicted tokens; and determining, by the computing system, an optimal edit to reconstruct a ground truth sequence of tokens from the different plurality of predicted tokens; where, for the training pair, the different plurality of predicted tokens serves as the plurality of input tokens, and the plurality of target tokens are based on the optimal edit. The one or more non- transitory computer-readable media where the non-autoregressive sequence prediction model may include a transformer model configured to process the plurality of input tokens to generate output embeddings, and where the categorical mixture prediction head may include a multi-layer perceptron that operates on the output embeddings of the transformer model. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-accessible medium.

[0011] Implementations may include one or more of the following features. The one or more non-transitory computer-readable media where the non-autoregressive sequence prediction model may include a transformer model configured to process the plurality of input tokens to generate output embeddings, and where the categorical mixture prediction head may include a multi-layer perceptron that operates on the output embeddings of the transformer model. Implementations of the described techniques may include hardware, a method or process, or computer software on a computer-accessible medium.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a graphical diagram illustrating an example approach for performing multiple decoding iterations with edit tokens to process a model input to generate a model output according to example implementations of aspects of the present disclosure;

[0013] Figure 2 is a graphical diagram illustrating an example approach for training a non-autoregressive sequence processing model according to example implementations of aspects of the present disclosure;

[0014] Figure 3 is a graphical diagram illustrating an example approach for training a non-autoregressive sequence processing model according to example implementations of aspects of the present disclosure;

[0015] Figure 4 is a graphical diagram illustrating an example architecture for a non- autoregressive sequence processing model according to example implementations of aspects of the present disclosure;

[0016] Figure 5 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;

[0017] Figure 6 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;

[0018] Figure 7 is a block diagram of an example sequence processing model according to example implementations of aspects of the present disclosure;

[0019] Figure 8 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;

[0020] Figure 9 is a block diagram of an example model development platform according to example implementations of aspects of the present disclosure;

[0021] Figure 10 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;

[0022] Figure 11 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;

[0023] Figure 12 is a block diagram of an example networked computing system according to example implementations of aspects of the present disclosure;

[0024] Figure 13 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure; and

[0025] Figure 14 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure.DETAILED DESCRIPTION

[0026] Example aspects of the present disclosure are directed to systems and methods which leverage NAR sequence processing models with improved computational efficiency. In particular, example implementations of the present disclosure address limitations found in previous NAR models by allowing dynamic text generation through iterative editing, which enables the predicted sequence to expand or contract as needed. In some implementations, this capability is supported by a unique mask-based training procedure that incorporates bidirectional attention across the entire input sequence and / or the use of novel edit tokens which provide a mechanism for the model to iteratively edit the predicted sequence over a number of decoding iterations. The proposed approaches facilitate the insertion, deletion, and / or prediction of missing tokens anywhere within the text. Additionally, some example training approaches provided herein employ student forcing for feedback, aiding the system in self-correcting its predictions. The proposed NAR sequence processing models can be applied to various tasks such as machine translation, token correction, infilling, language modeling, and / or other tasks across multiple different modalities. The proposed NAR sequence processing models have exhibited performance on par with or superior to traditional autoregressive models while requiring significantly fewer decoding steps, thereby enabling sequence prediction with reduced computational consumption.

[0027] More particularly, AR models are the dominant paradigm for all high- performing LLMs today, whether proprietary or open-source. They have shown excellent capabilities in many aspects of natural language generation such as text reasoning, STEM, and code generation tasks. Autoregression is a natural approach to language generation due to the uni-directional nature of text.

[0028] However, there are two shortcomings with the AR paradigm that are addressed by the present disclosure. First, autoregressive models require a network evaluation for each output token. Implementation tricks such as key value (KV) caching are helpful in practice, but have their own tradeoffs (such as memory usage). Secondly, AR models always condition on their full history, so mistakes are fed back into the model continuously, and errors compound. This has been hypothesized to be one source of the “hallucination problem” in LLMs. An ideal model would be able to identify and self-correct mistakes.

[0029] NAR models offer an alternative to the AR paradigm. One technical challenge is to model the joint probability distribution p(y1;y2, ■ ■ ■ > VN) °fasequence of N tokens. The AR formulation decomposes this into a chain of conditional distributions, p .yi)p .y2 \yi)- ■ ■ P .yN \yi>y2> ■ ■ ■ ' VN ~ !)• This decomposition can, at least in principle,model arbitrarily complex joint distributions; it also maps naturally onto neural network architectures like recurrent neural networks (RNNs) or causal Transformers. However, this representation power comes at the cost of requiring N network evaluations to sample a sequence.

[0030] Most existing NAR models have a number of technical and practical limitations. For example, most papers provide empirical results on a limited set of constrained tasks such as machine translation. Only a few papers have explored other tasks such as pre-training (language modeling) or editing. Many methods only work well with an AR model used as a teacher for knowledge distillation. This leads to pipeline complexity and upper bounds performance to that of the AR teacher. Speedups achieved by the NAR models often tend to be a trade-off with performance. Finally, the techniques often require significant departures in architecture from transformers. From a practical perspective, this is problematic due to the many good scaling properties of Transformers, and the extensive infrastructure that has built up around Transformers.

[0031] Example aspects of the present disclosure are directed to a NAR decoding approach that learns to iteratively edit its output. Some example implementations train on corrupted strings and learn to predict optimal edit sequences that move us closer to our data distribution. Some example implementations use student forcing to align the train / test distribution, reducing exposure bias. These techniques can be used independently and / or together. Example implementations also address the target dependency problem by introducing a thin latent variable model before the Transformer logit projection.

[0032] Example implementations of the proposed techniques consistently achieve similar or better performance than AR for a range of tasks such as machine translation, token correction and text inpainting, while requiring significantly fewer decoding steps (e.g., 4-14x fewer steps). The proposed models can also be to the language modeling task, thus opening the door to fully NAR modeling for both pre-training and fine-tuning tasks. Example implementations also demonstrate self-correction abilities.

[0033] More particularly, the present disclosure introduces computing systems and methods that can perform sequence prediction with enhanced computational efficiency. The proposed systems can include and / or leverage a NAR sequence processing model that allows for processing of multiple input tokens to generate corresponding output tokens in parallel. For instance, the NAR sequence processing model can iteratively operate over a plurality of decoding iterations. During each decoding iteration, the system can generate numerous outputtokens in parallel, which can lead to significant reductions in processing time compared to traditional autoregressive models that generate output tokens sequentially.

[0034] In particular, in some implementations, the NAR sequence processing model is or includes a decoder-only Transformer-based architecture. The architecture can apply bidirectional (non-causal) attention and / or contain a novel prediction head described elsewhere herein. In other implementations, the NAR sequence processing model can include or have an encoder-decoder architecture.

[0035] In some implementations, the disclosed technology can utilize or output one or more “edit” tokens during intermediate decoding iterations. These edit tokens, such as delete or insert tokens, can facilitate dynamic modifications of the output sequence. For example, a delete token can instruct the system to remove specific tokens, thereby adjusting the sequence length during the decoding process. Similarly, insert tokens can lead to the addition of new tokens, enhancing the system’s flexibility in generating outputs that require modifications from the input sequence. Alternatively, the model can directly replace an input token with a traditional token (e.g., such as a regular word token). These tokens can be referred to as “regular” tokens to distinguish from “edit” tokens. The edit tokens enable the sequence processing model to control or modify the number of tokens contained in the modelgenerated output over a series of decoding iterations. This aspect distinguishes the proposed techniques from other NAR approaches, many of which are forced to use a second model strictly to predict the number of tokens before using their NAR model to fill in values for those tokens.

[0036] In particular, unlike traditional AR models that generate a sequence token by token, example techniques provided herein leverage ‘edit tokens’ that facilitate dynamic changes in sequence length during the decoding process. This method allows the model to adjust the sequence length iteratively, ensuring that the output length matches the input length unless modified by specific edit operations. This approach helps in maintaining efficient processing and reducing the iterative load by potentially minimizing the number of decoding steps required.

[0037] Specifically, example implementations of the present disclosure include an enhanced vocabulary that includes special edit tokens that include mask ([M]), delete ([D]), and multiple insert tokens ([ / - through [ / mflx]) These tokens allow the model to perform text generation by iteratively editing its input, thus enabling both the expansion and contraction of sequences dynamically. For instance, a delete token ([D]) removes its corresponding tokenfrom the sequence, and an insert token ([ ]) adds multiple mask tokens in the sequence, thereby adjusting the sequence length as needed.

[0038] The proposed systems can also include a set of editing logic for generating input tokens for subsequent decoding iterations based on the output tokens of current iterations. For example, the editing logic can edit the input tokens and / or the output tokens for a current iteration to generate the input tokens for a subsequent decoding iteration.

[0039] This iterative refinement can continue until the output tokens meet specific criteria, such as containing only regular tokens, containing an end of sequence token, and / or other stopping criteria. This iterative process allows the system to progressively edit the output to achieve a final output sequence with improved accuracy.

[0040] The computing system described in the present disclosure can be particularly effective in tasks that involve text, images, and / or audio tokens. For example, each decoding iteration can process these different types of tokens, making the system versatile across various applications, such as language translation, image captioning, or speech recognition.

[0041] According to another aspect of the present disclosure, the proposed models can be trained using pairs of input tokens and target tokens, where, in at least some training examples, the target tokens include edit tokens that guide the model in learning how to adjust the input tokens to match a desired output. This training method can also include evaluating a loss function that measures the difference between the predicted tokens and the target tokens, allowing the model to learn from its errors and improve over time.

[0042] In some implementations, the training of the NAR sequence processing models can include performing a corruption and editing process to generate the training pairs described above. An input ground truth sequence can first be corrupted. The corruption can include various actions such as dropping, inserting, flipping, or masking tokens in the ground truth sequence to create a corrupted sequence.

[0043] Thus, the training of the proposed models can include a corruption process where the ground truth sequence is altered using a defined set of actions, such as, for example, keep, drop, insert, flip, and / or mask, which may be selected based on a categorical distribution. This process not only introduces variability into the training data but also helps the model learn robustness by predicting necessary edits to revert the corrupted sequence back to its original form.

[0044] The training example generation process can also include the use of dynamic programming algorithms, such as those used to compute the Levenshtein distance, todetermine the optimal edits needed to transform the corrupted sequence back into the ground truth sequence.

[0045] Then, for training, the corrupted sequence can serve as input to the model while the target output is the optimal edits. By training on these pairs, the model learns to reconstruct the original sequence by predicting the necessary edits. This approach can help the model become robust to errors and variations in the input data. This method also ensures that the model learns the most efficient way to correct errors in the input tokens.

[0046] Thus, during training, each pair of input and target sequences can be adjusted to have the same length, where the target sequence is derived by calculating the optimal edit path (e.g., that minimizes the Levenshtein distance from the corrupted input). This optimal editing can be computed efficiently using dynamic programming, ensuring that the edits are cost-effective and minimal.

[0047] In some implementations, training of the models can also include performing student forcing during training, where the model’s predictions are used as new inputs to the training process. In particular, the model’s predictions can be recycled as new inputs. This method is akin to replay buffers in reinforcement learning This technique can allow the model to learn from its own predictions and gradually improve its error-correction capabilities.

[0048] Another aspect of the present disclosure is directed to a categorical mixture prediction head that utilizes a latent variable in generating output tokens. This head can process output embeddings from a transformer model and apply a multi-layer perceptron to enhance the prediction capabilities of the system. The latent variable can have multiple states, each corresponding to different modes of the data distribution, which can allow the system to capture complex patterns in the data.

[0049] In particular, the categorical mixture prediction head can introduce a discrete latent variable to model the output distribution as a mixture of mean field distributions. The latent variable can capture the underlying distribution modes and assists with modeling dependencies among output tokens. This method not only improves the data fitting but also allows for efficient parallel sampling of output tokens, thus optimizing both the performance and computational efficiency of the model.

[0050] The systems and methods of the present disclosure provide a number of technical effects and benefits. As one example, unlike AR models, which generate one token at a time in a sequential manner, the NAR model processes multiple tokens in parallel during each decoding iteration. This parallel processing capability inherently accelerates thegeneration process because it reduces the number of sequential operations required to generate a sequence. For example, while an AR model must wait for the previous token to be generated before proceeding to the next, the NAR model can generate multiple tokens simultaneously, which may enable the inference system to leverage modem multi-core processing architectures more effectively.

[0051] As another technical effect and benefit, the NAR model can significantly reduce the number of decoding steps required to generate a complete sequence. Since each step can produce multiple tokens at once, the total number of iterations needed to generate a sequence is lower. For instance, if an AR model requires N steps to generate N tokens, the NAR model might generate the same sequence in far fewer steps, depending on the number of tokens it can process in parallel at each iteration. This reduction in steps directly translates to lower computational time and resource usage.

[0052] As another technical effect and benefit, the technique of iteratively refining the output sequence using edit tokens (such as insert and delete operations) allows the NAR model to converge more quickly to the final sequence. This iterative process, which adjusts the sequence towards the target output in each step, can be more computationally efficient than the AR approach of predicting each token based solely on the preceding tokens. The ability to make larger adjustments (e.g., inserting or deleting multiple tokens at once) can lead to faster convergence in fewer steps.

[0053] Furthermore, the use of techniques like student forcing in training, where the model learns to correct its own predictions, can lead to a more robust model that requires fewer computational resources during inference. By learning to self-correct, the NAR model may achieve higher accuracy with fewer iterations, further reducing computational demands during actual deployment. Additionally, the model may also converge more quickly during training, enabling reduced performance of training operations.

[0054] As another technical benefit, the parallel nature of the proposed approach lends itself to improved parallelization across multiple hetero- and / or homogeneous processing devices and / or hardware accelerators and / or multiple cores of a single processing device and / or hardware accelerator. In one example, the parallelization can occur at a thread level of the processing devices.

[0055] In another example, because the number of output tokens at each particular decoding iterations is known at the beginning of the decoding iteration, the computing system can have knowledge of or otherwise be able to determine how much inference hardware to allocate for the decoding iteration to optimize the usage of the available inference hardwarefor the decoding iteration. This can result in reducing the amount of processors that are allocated to an inference task and therefore “locked up” or otherwise unavailable for use with other processing tasks, but which ultimately are not used for the inference task.

[0056] The NAR sequence processing model described in the present disclosure can be applied to a diverse range of tasks across different data modalities, demonstrating its versatility and broad applicability in various fields of technology.

[0057] As one example, the model can be particularly beneficial for tasks such as machine translation and text generation. For translation, the model can take as input a sequence of text tokens in one language (e.g., a sentence in English) and output a sequence of text tokens in another language (e.g., the corresponding sentence in French). For text generation, the input might be a prompt or a partial sentence, and the output would be a continuation or completion of the input text, potentially used for creative writing aids, automated content creation, or chatbot responses.

[0058] In audio processing tasks, such as speech-to-text transcription or audio data compression, the model can handle sequences of audio tokens. For speech-to-text, the input would be an audio stream segmented into tokens representing phonetic components, and the output would be a textual representation of the spoken content. For audio compression, the input to the model may be raw audio data and the model can output a compressed version, preserving essential information while reducing the data size.

[0059] As another example, the model can also be extended to visual data processing tasks. For instance, in image captioning, the input could be an array of image tokens (derived from pixel data), and the output would be a sequence of text tokens that describe the image content. In video processing, the model could input sequences of image frames (video tokens) and output compressed or enhanced versions of these frames, useful in streaming services to reduce bandwidth usage while maintaining video quality.

[0060] Another potential application is in the biomedical field, where the model can process sequences of genetic data tokens or biomedical signal tokens. For example, in genetic sequence analysis, the input could be tokens representing segments of DNA or RNA sequences, and the output could be annotations or predictions about gene functions, mutations, or disease associations. As another example, in the case of biomedical signals, such as from electroencephalograms (EEGs) or electrocardiograms (ECGs), the model can input signal tokens and output diagnostic interpretations or anomaly detections.

[0061] Various example implementations are described herein with respect to the accompanying Figures.

[0062] Figure l is a graphical diagram illustrating an example approach for performing multiple decoding iterations with edit tokens to process a model input to generate a model output according to example implementations of aspects of the present disclosure.

[0063] In particular, Figure 1 illustrates a NAR sequence processing model 104 operating over multiple parallel decoding iterations. The model input 102 is obtained and processed by the NAR sequence processing model 104 to generate a model output 106. This process involves several decoding iterations (e.g., three decoding iterations are illustrated int he example) where the model 104 processes a plurality of input tokens to generate a plurality of output tokens in parallel.

[0064] During the first decoding iteration, the plurality of output tokens 108 is generated. These output tokens 108 can include regular tokens and may also comprise one or more edit tokens. The edit tokens facilitate modifications to the token sequence, such as deletions or insertions.

[0065] Edit logic 110 is employed to apply the necessary edits to the plurality of output tokens 108, based on the edit tokens contained within them. The edit logic 110 can, for example, remove tokens corresponding to delete tokens or insert mask tokens adjacent to insert tokens within the plurality of output tokens 108. This editing forms the plurality of input tokens 112 for the subsequent decoding iteration.

[0066] Thus, the plurality of input tokens 112 for the second decoding iteration is generated by the edit logic 110, which performs edits based on the output tokens from the previous iteration. This process can include inserting or deleting tokens as dictated by the edit tokens from the previous iteration's output.

[0067] The iterative process continues with each decoding iteration processing the updated sequence of input tokens to refine the output progressively. This iterative process can continue until the output tokens meet specific criteria, such as containing only regular tokens or an end of sequence token.

[0068] The described system and method can handle various types of tokens, including textual tokens, image tokens, or audio tokens, demonstrating the versatility of the NAR sequence processing model 104 across different data modalities. Each part and operation described in Figure 1 is an example of how the system can function and is not limited to these specific implementations.

[0069] Figure 2 is a graphical diagram illustrating an example approach for training a NAR sequence processing model according to example implementations of aspects of the present disclosure.

[0070] In particular, Figure 2 illustrates an example training approach for a NAR sequence processing model 212. The process begins with obtaining a ground truth sequence 202. This ground truth sequence 202 is then modified by corruption logic 204 to generate a corrupted sequence 206. The corruption logic 204 can perform various actions such as dropping, inserting, flipping, or masking tokens within the ground truth sequence 202.

[0071] Once the corrupted sequence 206 is generated, optimal edit logic 208 determines the necessary edits to transform the corrupted sequence 206 back into the ground truth sequence 202. This determination can involve a dynamic programming algorithm that calculates the minimal number of operations required to minimize the Levenshtein distance between the corrupted sequence 206 and the ground truth sequence 202.

[0072] The corrupted sequence 206 serves as the plurality of input tokens 200 for the NAR sequence processing model 212. The plurality of target tokens 201, which are derived based on the optimal edits determined by the optimal edit logic 208, guide the model 212 in correcting the corrupted sequence 206.

[0073] The NAR sequence processing model 212 processes the plurality of input tokens 200 to generate a plurality of predicted tokens 214. These predicted tokens 214 are then compared to the plurality of target tokens 201 using a loss function 216. The loss function 216 evaluates the discrepancies between the predicted tokens 214 and the target tokens 201.

[0074] Based on the evaluation by the loss function 216, the parameters of the NAR sequence processing model 212 can be modified to reduce the prediction errors in future iterations. This training method allows the model 212 to learn from the edits necessary to correct the input tokens 200 towards the target tokens 201, thereby improving its accuracy over time.

[0075] Figure 3 is a graphical diagram illustrating an example approach for training a NAR sequence processing model according to example implementations of aspects of the present disclosure.

[0076] In particular, Figure 3 illustrates an example training approach with student forcing for a NAR sequence processing model 310. The process begins with a ground truth sequence 302, which is modified by corruption logic 304 to generate a corrupted sequence 306. The corrupted sequence 306 serves as the first training input 308 for the NAR sequence processing model 310.

[0077] The NAR sequence processing model 310 processes the first training input 308 to generate a first plurality of predicted tokens 312. These predicted tokens 312 thenserve as a second training input for another iteration of processing by the NAR sequence processing model 310, demonstrating the student forcing method where the model's own predictions are used as subsequent inputs.

[0078] The second plurality of predicted tokens 318 is generated by the NAR sequence processing model 310 after processing the first plurality of predicted tokens 312. Optimal edit logic 314 determines the necessary edits to transform the first plurality of predicted tokens 312 back into a sequence that more closely resembles the ground truth sequence 302. These determined edits are used to generate the plurality of target tokens 316, which guide the model in correcting the predicted tokens towards the desired output.

[0079] The loss function 320 evaluates the discrepancies between the second plurality of predicted tokens 318 and the plurality of target tokens 316. Based on this evaluation, the parameters of the NAR sequence processing model 310 can be modified to reduce the prediction errors in future iterations. This training method allows the NAR sequence processing model 310 to learn from its own predictions and iteratively improve its accuracy over time.

[0080] Although shown separately for ease of explanation, the two training techniques shown in Figures 2 and 3 can optionally be combined or otherwise employed together. For example, with reference to Figure 2, the predicted tokens 214 could be used as an additional training input for the sequence processing model (and also evaluated by the optimal edit logic to generate an additional target) to perform the student forcing technique in addition to the operations shown in Figure 2.

[0081] Figure 4 is a graphical diagram illustrating an example architecture for a non- autoregressive sequence processing model according to example implementations of aspects of the present disclosure.

[0082] In particular, Figure 4 depicts an example architecture of an example NAR sequence processing model 400. The architecture includes several components configured to process a plurality of input tokens to generate a plurality of output tokens in parallel.

[0083] The model input 402 is processed by a bidirectional transformer 404. The bidirectional transformer 404 can handle the input tokens by applying bidirectional attention mechanisms, which allow it to consider both preceding and succeeding tokens in the sequence simultaneously for generating output embeddings.

[0084] Attached to the bidirectional transformer 404 is a categorical mixture prediction head 406. This component utilizes a latent variable that has a plurality of possible states to enhance the prediction capabilities of the model 400. The categorical mixtureprediction head 406 can include a multi-layer perceptron that operates on the output embeddings produced by the bidirectional transformer 404.

[0085] As one example, the multi-layer perceptron within the categorical mixture prediction head 406 can process the output embeddings and apply various transformations, potentially including linear transformations and non-linear activations such as the Gaussian Error Linear Unit (GeLU), to refine the predictions. This processing helps in capturing complex patterns and dependencies among the output tokens, facilitating the generation of a coherent and contextually appropriate sequence of output tokens. The arrangement of these components can vary, and the example shown is illustrative of one possible configuration.

[0086] Figure 5 depicts a flowchart of a method 500 for training one or more machine-learned models according to aspects of the present disclosure. For instance, an example machine-learned model can include a NAR sequence processing model.

[0087] One or more portion(s) of example method 500 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 500 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 500 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. Figure 5 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 5 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 500 can be performed additionally, or alternatively, by other systems.

[0088] At 502, example method 500 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 500 as a “training” instance, it is to be understood that runtime inferences can form training instances when a model is trained using an evaluation of the model’s performance on that runtime instance (e.g., online training / leaming). Example data types for the training instance and various tasks associated therewith are described throughout the present disclosure.

[0089] At 504, example method 500 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.

[0090] At 506, example method 500 can include receiving an evaluation signal associated with the output. The evaluation signal can be obtained using a loss function. Various determinations of loss can be used, such as mean squared error, likelihood loss, cross entropy loss, hinge loss, contrastive loss, or various other loss functions. The evaluation signal can be computed using known ground-truth labels (e.g., supervised learning), predicted or estimated labels (e.g., semi- or self-supervised learning), or without labels (e.g., unsupervised learning). The evaluation signal can be a reward (e.g., for reinforcement learning). The reward can be computed using a machine-learned reward model configured to generate rewards based on output(s) received. The reward can be computed using feedback data describing human feedback on the output(s).

[0091] At 508, example method 500 can include updating the machine-learned model using the evaluation signal. For example, values for parameters of the machine-learned model(s) can be learned, in some embodiments, using various training or learning techniques, such as, for example, backwards propagation. For example, the evaluation signal can be backpropagated from the output (or another source of the evaluation signal) through the machine-learned model(s) to update one or more parameters of the model(s) (e.g., based on a gradient of the evaluation signal with respect to the parameter value(s)). For example, system(s) containing one or more machine-learned models can be trained in an end-to-end manner. Gradient descent techniques can be used to iteratively update the parameters over a number of training iterations. In some implementations, performing backwards propagation of errors can include performing truncated backpropagation through time. Example method 500 can include implementing a number of generalization techniques (e.g., weight decays, dropouts, etc.) to improve the generalization capability of the models being trained.

[0092] In some implementations, example method 500 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.).

[0093] In some implementations, example method 500 can be implemented for particular stages of a training procedure. For instance, in some implementations, examplemethod 500 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.

[0094] In some implementations, example method 500 can be implemented for finetuning 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 500 uses adapter modules. Adapters can be small trainable layers that are inserted between pre-existing layers of a pre-trained model. During the finetuning process, the original parameters of the pre-trained model are typically frozen, and only the parameters of the adapters are updated.

[0095] In some implementations, example method 500 can be implemented to execute parameter-efficient fine-tuning methods, such as Layerwise Optimization of Residuals (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.

[0096] An example fine-tuning approach includes reinforcement learning. Reinforcement learning can be based on user feedback on model performance during use.

[0097] Figure 6 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.

[0098] Machine-learned model(s) 1 can be or include one or multiple machine- learned models or model components. Example machine-learned models can include neural networks (e.g., deep neural networks). Example machine-learned models can include nonlinear models or linear models. Example machine-learned models can use other architectures in lieu of or in addition to neural networks. Example machine-learned models can include decision tree based models, support vector machines, hidden Markov models, Bayesian networks, linear regression models, k-means clustering models, etc.

[0099] 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 thepreceding figures. For example, machine-learned model(s) 1 can be or include, or otherwise be representative of any one or more of NAR sequence processing models, etc. Although various features, variations, and implementations described below are described with respect to machine-learned model(s) 1, it is to be understood that such features, variations, and implementations are to be understood as described with respect to each of NAR sequence processing models, etc., any other machine-learned component described herein.

[0100] Example neural networks can include feed-forward neural networks, recurrent neural networks (RNNs), including long short-term memory (LSTM) based recurrent neural networks, convolutional neural networks (CNNs), diffusion models, generative-adversarial networks, or other forms of neural 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 multiheaded self-attention models.

[0101] 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.

[0102] 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).

[0103] 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 mechanismthat 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 compute-efficient forward passes.

[0104] 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.

[0105] Example data types for input(s) 2 or output(s) 3 include natural language text data, software code data (e.g., source code, object code, machine code, or any other form of computer-readable instructions or programming languages), machine code data (e.g., binary code, assembly code, or other forms of machine-readable instructions that can be executed directly by a computer’s central processing unit), assembly code data (e.g., low-level programming languages that use symbolic representations of machine code instructions to program a processing unit), 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.

[0106] 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.

[0107] 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.

[0108] Figure 7 is a block diagram of an example implementation of an example machine-learned model configured to process sequences of information. For instance, an example implementation of machine-learned model(s) 1 can include machine-learned sequence processing model(s) 4. An example system can pass input(s) 2 to sequence processing model(s) 4. Sequence processing model(s) 4 can include one or more machine- learned components. Sequence processing model(s) 4 can process the data from input(s) 2 to obtain an input sequence 5. Input sequence 5 can include one or more input elements 5-1, 5- 2, . . . , 5-A7, 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.

[0109] 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 al., MusicLM: Generating Music From Text, ARXIV:2301.1 1325V1 (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.

[0110] 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”).

[0111] 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 ofinput 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.

[0112] Elements 5-1, 5-2, . . . , 5-M can represent, in some cases, building blocks for capturing or expressing meaningful information in a particular data domain. For instance, the elements can describe “atomic units” across one or more domains. For example, for textual input source(s), the elements can correspond to groups of one or more words or sub-word components, such as sets of one or more characters.

[0113] For example, elements 5-1, 5-2, . . . , 5-M can represent tokens obtained using a tokenizer. For instance, a tokenizer can process a given portion of an input source and output a series of tokens (e.g., corresponding to input elements 5-1, 5-2, . . . , 5-M) that represent the portion of the input source. Various approaches to tokenization can be used. For instance, textual input source(s) can be tokenized using a byte-pair encoding (BPE) technique. See, e.g., Kudo et al., SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing, PROCEEDINGS OF THE 2018 CONFERENCE ON EMPIRICAL METHODS IN NATURAL LANGUAGE PROCESSING (System Demonstrations), pages 66-71 (October 31-November 4, 2018), https: / / aclanthology.org / D18-2012.pdf. Image-based input source(s) can be tokenized by extracting and serializing patches from an image.

[0114] 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 7 can be the tokens or can be the embedded representations thereof.

[0115] Prediction layer(s) 6 can predict one or more output elements 7-1, 7-2, . . . , 7- N based on the input elements. Prediction layer(s) 6 can include one or more machine-learned model architectures, such as one or more layers of learned parameters that manipulate and transform the input(s) to extract higher-order meaning from, and relationships between, input element(s) 5-1, 5-2, . . . , 5-M. In this manner, for instance, example prediction layer(s) 6 can predict new output element(s) in view of the context provided by input sequence 5.

[0116] 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.”

[0117] A transformer is an example architecture that can be used in prediction layer(s) 4. See, e.g., Vaswani et al., Atention 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, . . . , 7-N. A transformer block can include one or more attention layer(s) and one or more post-attention layer(s) (e.g., feedforward layer(s), such as a multi-layer perceptron).

[0118] Prediction layer(s) 6 can include other machine-learned model architectures in addition to or in lieu of transformer-based architectures. For example, recurrent neural networks (RNNs) and long short-term memory (LSTM) models can also be used, as well as convolutional neural networks (CNNs). In general, prediction layer(s) 6 can leverage various kinds of artificial neural networks that can understand or generate sequences of information.

[0119] Output sequence 7 can include or otherwise represent the same or different data types as input sequence 5. For instance, input sequence 5 can represent textual data, and output sequence 7 can represent textual data. Input sequence 5 can represent image, audio, or audiovisual data, and output sequence 7 can represent textual data (e.g., describing the image, audio, or audiovisual data). It is to be understood that prediction layer(s) 6, and any other interstitial model components of sequence processing model(s) 4, can be configured to receive a variety of data types in input sequence(s) 5 and output a variety of data types in output sequence(s) 7.

[0120] 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.

[0121] In some implementations, output sequence 7 can be generated non- autoregressively. For instance, for some applications, outputs (e.g., embeddings) of one or more prediction layer(s) 6 can be passed through one or more output layers (e.g., a softmax layer) to obtain a respective probability distributions over an output vocabulary (e.g., a textual or symbolic vocabulary) for each of multiple output tokens. The outputs can beconditioned on a set of input elements in a context window. In this manner, for instance, prediction layer 6 can operate to generate multiple output elements (e.g., output elements 7-1, 7-2, 7-3) in parallel. For instance, multiple output elements of output sequence 7 can be predicted together.

[0122] 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.

[0123] Figure 8 is a block diagram of an example technique for populating an example input sequence 8. Input sequence 8 can include various functional elements that form part of the model infrastructure, such as an element 8-0 obtained from a task indicator 9 that signals to any model(s) that process input sequence 8 that a particular task is being performed (e.g., to help adapt a performance of the model(s) to that particular task). Input sequence 8 can include various data elements from different data modalities. For instance, an input modality 10-1 can include one modality of data. A data-to-sequence model 11-1 can process data from input modality 10-1 to project the data into a format compatible with input sequence 8 (e.g., one or more vectors dimensioned according to the dimensions of input sequence 8) to obtain elements 8-1, 8-2, 8-3. Another input modality 10-2 can include a different modality of data. A data-to-sequence model 11-2 can project data from input modality 10-2 into a format compatible with input sequence 8 to obtain elements 8-4, 8-5, 8- 6. Another input modality 10-3 can include yet another different modality of data. A data-to- sequence model 11-3 can project data from input modality 10-3 into a format compatible with input sequence 8 to obtain elements 8-7, 8-8, 8-9.

[0124] Input sequence 8 can be the same as or different from input sequence 5. Input sequence 8 can be a multimodal input sequence that contains elements that represent data from different modalities using a common dimensional representation. For instance, an embedding space can have P dimensions. Input sequence 8 can be configured to contain a plurality of elements that have P dimensions. In this manner, for instance, example implementations can facilitate information extraction and reasoning across diverse datamodalities by projecting data into elements in the same embedding space for comparison, combination, or other computations therebetween.

[0125] 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.

[0126] In some implementations, the expressive power of the embedding space may not be limited to meanings associated with any particular set of tokens or other building blocks. For example, a continuous embedding space can encode a spectrum of high-order information. An individual piece of information (e.g., a token) can map to a particular point in that space: for instance, a token for the word “dog” can be projected to an embedded value that points to a particular location in the embedding space associated with canine-related information. Similarly, an image patch of an image of a dog on grass can also be projected into the embedding space. In some implementations, the projection of the image of the dog can be similar to the projection of the word “dog” while also having similarity to a projection of the word “grass,” while potentially being different from both. In some implementations, the projection of the image patch may not exactly align with any single projection of a single word. In some implementations, the projection of the image patch can align with a combination of the projections of the words “dog” and “grass.” In this manner, for instance, a high-order embedding space can encode information that can be independent of data modalities in which the information is expressed.

[0127] 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). Forinstance, the input value represented by element 8-0 can be a learned within a continuous embedding space.

[0128] 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).

[0129] 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.).

[0130] 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.

[0131] Figure 9 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.

[0132] Model development platform 12 can provide one or more model libraries 13 containing building blocks for new models. Model libraries 13 can include one or more pretrained foundational models 13-1, which can provide a backbone of processing power across various tasks. Model libraries 13 can include one or more pre-trained expert models 13-2, which can be focused on performance in particular domains of expertise. Model libraries 13 can include various model primitives 13-3, which can provide low-level architectures or components (optionally pre-trained), which can be assembled in various arrangements as desired. 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.

[0133] 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.

[0134] 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.

[0135] 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).

[0136] 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.

[0137] Pre-training pipelines 17-2 can include a machine-learned model training workflow configured to update development model 16 over large-scale, potentially noisy datasets. For example, pre-training can leverage unsupervised learning techniques (e.g., denoising, etc.) to process large numbers of training instances to update model parameters from an initialized state and achieve a desired baseline performance. Pre-training pipelines 17-2 can leverage unlabeled datasets in dataset(s) 17-1 to perform pre-training. Workbench 15 can implement a pre-training pipeline 17-2 to pre-train development model 16.

[0138] 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 signalsfrom user feedback signals. Workbench 15 can implement a fine-tuning pipeline 17-3 to finetune development model 16.

[0139] 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.

[0140] 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.

[0141] 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).

[0142] 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 one or more training iterations. Workbench 15 can implement prompt engineering tools in development model 16.

[0143] Prompt libraries 17-4 can include pipelines for prompt generation. For example, inputs can be generated using development model 16 itself or other machine- learned models. In this manner, for instance, a first model can process information about a task and output a input for a second model to process in order to perform a step of the task. The second model can be the same as or different from the first model. Workbench 15 can implement prompt generation pipelines in development model 16.

[0144] Prompt libraries 17-4 can include pipelines for context injection. For instance, a performance of development model 16 on a particular task can improve if provided with additional context for performing the task. Prompt libraries 17-4 can include software components configured to identify desired context, retrieve the context from an external source (e.g., a database, a sensor, etc.), and add the context to the input prompt. Workbench 15 can implement context injection pipelines in development model 16.

[0145] 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 techniquesadapted for training a wide variety of machine-learned models. Example training techniques can correspond to the example training method 500 described above.

[0146] 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.

[0147] 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”).

[0148] 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.

[0149] 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.

[0150] 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 canreceive, 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.

[0151] 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.

[0152] 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.

[0153] Figure 10 is a block diagram of an example training flow for training a machine-learned development model 16. One or more portion(s) of the example training flow can be implemented by a computing system that includes one or more computing devices such as, for example, computing systems described with reference to the other figures. Each respective portion of the example training flow can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of the example training flow can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. FIG. 10 depicts elements performed in a particular order for purposes of illustration and discussion. Those ofordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. FIG. 10 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.

[0154] Initially, development model 16 can persist in an initial state as an initialized model 21. Development model 16 can be initialized with weight values. Initial weight values can be random or based on an initialization schema. Initial weight values can be based on prior pre-training for the same or for a different model.

[0155] Initialized model 21 can undergo pre-training in a pre-training stage 22. Pretraining stage 22 can be implemented using one or more pre-training pipelines 17-2 over data from dataset(s) 17-1. Pre-training can be omitted, for example, if initialized model 21 is already pre-trained (e.g., development model 16 contains, is, or is based on a pre-trained foundational model or an expert model).

[0156] 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 as satisfactory performance, if the model was already fine-tuned, or if other tuning approaches are preferred.

[0157] Fine-tuned model 29 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 29 can be the initial state if development model 16 was already fine-tuned. Fine-tuned model 29 can undergo refinement with user feedback 26. For instance, refinement with user feedback 26 can include reinforcement learning, optionally based on human feedback from human users of fine-tuned model 25. As reinforcement learning can be a form of fine-tuning, it is to be understood that fine-tuning stage 24 can subsume the stage for refining with user feedback 26. Refinement with user feedback 26 can produce a refined model 27. Refined model 27 can be output to downstream system(s) 28 for deployment or further development.

[0158] In some implementations, computational optimization operations can be applied before, during, or after each stage. For instance, initialized model 21 can undergocomputational 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.

[0159] Figure 11 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.

[0160] 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.

[0161] Model host 31 can leverage various other resources and tools to augment the inference task. For instance, model host 31 can communicate with tool interfaces 35 to facilitate tool use by model instance(s) 31-1. Tool interfaces 35 can include local or remote APIs. Tool interfaces 35 can include integrated scripts or other software functionality. Model host 31 can engage online learning interface(s) 36 to facilitate ongoing improvements to machine-learned model(s) 1. For instance, online learning interface(s) 36 can be used within reinforcement learning loops to retrieve user feedback on inferences served by model host 31. Model host 31 can access runtime data source(s) 37 for augmenting input(s) 2 with additional contextual information. For instance, runtime data source(s) 37 can include a knowledge graph 37-1 that facilitates structured information retrieval for information associated with input request(s) 33 (e.g., a search engine service). Runtime data source(s) 37 can include public or private, external or local database(s) 37-2 that can store information associated with input request(s) 33 for augmenting input(s) 2. Runtime data source(s) 37 can include accountdata 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.

[0162] 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.

[0163] For example, model host 31 can operate on a server system that provides a machine-learning service to client device(s) that operate client(s) 32 (e.g., over a local or wide-area network). Client device(s) can be end-user devices used by individuals. Client device(s) can be server systems that operate client(s) 32 to provide various functionality as a service to downstream end-user devices.

[0164] 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.

[0165] 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 high-speed memory. Model instance(s) 31-1 can include multiple instance(s) of the same model (e.g., for parallel execution of more requests on the same model). Model instance(s) 31-1 can include instance(s) of different model(s). Model instance(s) 31-1 can include cached intermediate states of active or inactive model(s) used to accelerate inference of those models. For instance, an inference session with a particular model may generate significant amounts of computational results that can be re-used for future inference runs (e.g., using a KV cache for transformer-based models). These computational results can be saved in association with that inference session so that session can be executed more efficiently when resumed.

[0166] 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 caninclude a dynamic pool of available resources shared with other processes. Compute resource(s) 31-2 can include memory devices large enough to fit an entire model instance in 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.

[0167] 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.

[0168] 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.

[0169] 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.

[0170] 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.

[0171] 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 modelhost 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.

[0172] 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.

[0173] 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 ormore 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.

[0174] In some implementations, input(s) 2 can be or otherwise represent natural language data. Machine-learned model(s) 1 can process the natural language data to generate an output. As an example, machine-learned model(s) 1 can process the natural language data to generate a language encoding output. As another example, machine-learned model(s) 1 can process the natural language data to generate a latent text embedding output. As another example, machine-learned model(s) 1 can process the natural language data to generate 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).

[0175] 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 speechdata, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a prediction output.

[0176] 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.

[0177] 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.

[0178] 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.

[0179] 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.

[0180] 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.

[0181] 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.

[0182] 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., naturallanguage 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.

[0183] 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.

[0184] 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).

[0185] 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).

[0186] 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).

[0187] Figure 12 is a block diagram of an example networked computing system that can perform aspects of example implementations of the present disclosure. The system can include a number of computing devices and systems that are communicatively coupled over a network 49. An example computing device 50 is described to provide an example of a computing device that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). An example server computing system 60 is described as an example of a server computing system that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Computing device 50 and server computing system(s) 60 can cooperatively interact (e.g., over network 49) to perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Model development platform system 70 is an example system that can host or serve model development platform(s) 12 for development of machine-learned models. Third-party system(s) 80 are example system(s) with which any of computing device 50, server computing system(s) 60, or model development platform system(s) 70 can interact in theperformance of various aspects of the present disclosure (e.g., engaging third-party tools, accessing third-party databases or other resources, etc.).

[0188] 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 12 can be co-located with, contained by, or otherwise integrated into one or more other devices or systems.

[0189] Computing device 50 can be any type of computing device, such as, for example, a personal computing device (e.g., laptop or desktop), a mobile computing device (e.g., smartphone or tablet), a gaming console or controller, a wearable computing device, an embedded computing device, a server computing device, a virtual machine operating on a host device, or any other type of computing device. Computing device 50 can be a client computing device. Computing device 50 can be an end-user computing device. Computing device 50 can be a computing device of a service provided that provides a service to an end user (who may use another computing device to interact with computing device 50).

[0190] 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.

[0191] 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.

[0192] 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.

[0193] 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.

[0194] 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.

[0195] Server computing system 60 can store or otherwise include one or more machine-learned models 65. Machine-learned model(s) 65 can be the same as or different from machine-learned model(s) 55. Machine-learned models 65 can include one or more machine-learned model(s) 1, such as a sequence processing model 4. Machine-learned models 65 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 65 can be received from computing device 50, model development platform system 70, third party system(s) 80, or developed locally on server computing system(s) 60. Machine-learned model(s) 65 can be loaded into memory 62 and used or otherwise implemented byprocessor(s) 61. Server computing system(s) 60 can implement multiple parallel instances of machine-learned model(s) 65.

[0196] 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.

[0197] 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.

[0198] Third-party system(s) 80 can include one or more processors 81 and a memory 82. Processor(s) 81 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. Memory 82 can include one or more non-transitory computer-readable storage media, such as HBM, RAM, ROM,EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. Memory 82 can store data 83 and instructions 84 which can be executed by processor(s) 81 to cause third-party system(s) 80 to perform operations. The operations can implement any one or multiple features described herein. The operations can implement example methods and techniques described herein. Example operations include the functionality described herein 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).

[0199] Figure 12illustrates 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).

[0200] Figure 13 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 13, 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.

[0201] Figure 14 is a block diagram of an example computing device 99 that performs according to example embodiments of the present disclosure. Computing device 99 can be the same as or different from computing device 98. Computing device 99 can be a user computing device or a server computing device (e.g., computing device 50, server computing system(s) 60, etc.). Computing device 98 can implement model host 31. For instance, computing device 99 can include a number of applications (e.g., applications 1 through N). Each application can be in communication with a central intelligence layer. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. In 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).

[0202] The central intelligence layer can include a number of machine-learned models. For example, as illustrated in Figure 14, 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.

[0203] 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 14, 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).

[0204] 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.

[0205] While the present subject matter has been described in detail with respect to various specific example embodiments thereof, each example is provided by way of explanation, not limitation of the disclosure. Those skilled in the art, upon attaining an understanding of the foregoing, can readily produce alterations to, variations of, and equivalents to such embodiments. Accordingly, the subject disclosure does not preclude inclusion of such modifications, variations or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. For instance, features illustrated 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.

[0206] 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.”

[0207] 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.

[0208] 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 asindicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

Claims

WHAT IS CLAIMED IS:

1. A computing system configured to perform sequence prediction with improved computational efficiency, the computing system comprising: one or more processors; and one or more non-transitory computer readable media that collectively store computerexecutable instructions for performing operations, the operations comprising: obtaining, by the computing system, a model input; iteratively performing, by the computing system, a plurality of decoding iterations with a non-autoregressive sequence processing model to generate a model output based on the model input; wherein, at each decoding iteration, the non-autoregressive sequence processing model processes a plurality of input tokens to generate a plurality of output tokens in parallel; wherein, for one or more intermediate decoding iterations of the plurality of decoding iterations, the plurality of output tokens comprise one or more edit tokens; and wherein performing, by the computing system, each of the one or more intermediate decoding iterations comprises generating, by the computing system, the plurality of input tokens for a subsequent decoding iteration of the plurality of decoding iterations based on the one or more edit tokens; and providing, by the computing system, the model output as an output.

2. The computing system of any preceding claim, wherein, for each of the one or more intermediate decoding iterations, generating the plurality of input tokens for the subsequent decoding iteration of the plurality of decoding iterations based on the one or more edit tokens comprises editing the plurality of output tokens for the current decoding iteration based on the one or more edit tokens to generate the plurality of input tokens for the subsequent decoding generation.

3. The computing system of claim 2, wherein: at least one of the one or more edit tokens comprises a delete token; andediting the plurality of output tokens based on the one or more edit tokens comprises removing the delete token from the plurality of output tokens.

4. The computing system of claim 2 or claim 3, wherein: at least one of the one or more edit tokens comprises an insert token; and editing the plurality of output tokens based on the one or more edit tokens comprises: inserting two or more mask tokens into the plurality of output tokens at positions adjacent to the insert token within the plurality of output tokens; and replacing the insert token with a corresponding input token located at a corresponding position within the plurality of input tokens.

5. The computing system of claim 4, wherein: the insert token comprises a value equal to an integer number within a range from two to a maximum insertion value; and inserting two or more mask tokens into the plurality of output tokens at positions adjacent to the insert token comprises inserting the integer number of mask tokens into the plurality of output tokens at positions adjacent to the insert token.

6. The computing system of claim 1, wherein, for each of the one or more intermediate decoding iterations, generating the plurality of input tokens for the subsequent decoding iteration of the plurality of decoding iterations based on the one or more edit tokens comprises editing the plurality of input tokens for the current decoding iteration based on the one or more edit tokens to generate the plurality of input tokens for the subsequent decoding generation.

7. The computing system of claim 6, wherein: at least one of the one or more edit tokens comprises a delete token; and editing the plurality of input tokens based on the one or more edit tokens comprises removing from the plurality of input tokens a corresponding input token located at a corresponding position to the delete token within the plurality of input tokens.

8. The computing system of claim 6 or claim 7, wherein: at least one of the one or more edit tokens comprises an insert token; and editing the plurality of input tokens based on the one or more edit tokens comprises: inserting two or more mask tokens into the plurality of input tokens at positions adjacent to a corresponding position to the insert token within the plurality of input tokens.

9. The computing system of claim 8, wherein: the insert token comprises a value equal to an integer number within a range from two to a maximum insertion value; and inserting two or more mask tokens into the plurality of input tokens at positions adjacent to the corresponding position comprises inserting the integer number of mask tokens into the plurality of input tokens at positions adjacent to the corresponding position.

10. The computing system of any preceding claim, wherein iteratively performing, by the computing system, the plurality of decoding iterations comprises iteratively performing, by the computing system, the plurality of decoding iterations until the plurality of output tokens for one of the decoding iterations contains only regular tokens.

11. The computing system of any of claims 1-9, wherein iteratively performing, by the computing system, the plurality of decoding iterations comprises iteratively performing, by the computing system, the plurality of decoding iterations until the plurality of output tokens contain an end of sequence token.

12. The computing system of any preceding claim, wherein the one or more non- transitory computer-readable media further store the non-autoregressive sequence prediction model.

13. The computing system of any preceding claim, wherein the non-autoregressive sequence prediction model comprises a decoder-only transformer model.

14. The computing system of any preceding claim, wherein the plurality of input tokens and the plurality of output tokens for each decoding iteration comprise textual tokens.

15. The computing system of any preceding claim, wherein the plurality of input tokens and the plurality of output tokens for each decoding iteration comprise image tokens or audio tokens.

16. A computer-implemented method for training a non-autoregressive sequence prediction model, the method comprising, for each of one or more training iterations: obtaining, by a computing system comprising one or more computing devices, a training pair comprising a plurality of input tokens and plurality of target tokens, wherein the plurality of target tokens comprise one or more edit tokens that indicate edits to be made to the plurality of input tokens to generate a corrected sequence of tokens; processing, by the computing system, the plurality of input tokens with the non- autoregressive sequence prediction model to generate a plurality of predicted tokens; evaluating, by the computing system, a loss function that compares the plurality of predicted tokens to the plurality of target tokens; and modifying, by the computing system, one or more values of one or more parameters of the non-autoregressive sequence prediction model based on the loss function.

17. The computer-implemented method of claim 16, wherein obtaining, by the computing system, training pair comprises constructing, by the computing system, the training pair by performing operations comprising: obtaining, by the computing system, a ground truth sequence of tokens; corrupting, by the computing system, one or more of the tokens contained in the ground truth sequence of tokens to generate a corrupted sequence of tokens; and determining, by the computing system, an optimal edit to reconstruct the ground truth sequence of tokens from the corrupted sequence of tokens; wherein, for the training pair, the corrupted sequence of tokens serves as the plurality of input tokens, and the plurality of target tokens are based on the optimal edit.

18. The computer-implemented method of claim 17, wherein corrupting, by the computing system, one or more of the tokens contained in the ground truth sequence of tokens to generate the corrupted sequence of tokens comprises one or more of: dropping one or more of the ground truth sequence of tokens; inserting one or more tokens into the ground truth sequence of tokens; flipping one or more of the ground truth sequence of tokens; and masking one or more of the ground truth sequence of tokens.

19. The computer-implemented method of claim 17 or 18, wherein determining, by the computing system, the optimal edit to reconstruct the ground truth sequence of tokens from the corrupted sequence of tokens comprises performing a dynamic programming algorithm to identify the optimal edit that contains a minimal number of operations required to minimize a Levenshtein distance to the ground truth sequence of tokens.

20. The computer-implemented method of claim 16, wherein obtaining, by the computing system, training pair comprises: performing, by the computing system, a training forward pass on a different training input with the non-autoregressive sequence prediction model to generate a different plurality of predicted tokens; and determining, by the computing system, an optimal edit to reconstruct a ground truth sequence of tokens from the different plurality of predicted tokens; wherein, for the training pair, the different plurality of predicted tokens serves as the plurality of input tokens, and the plurality of target tokens are based on the optimal edit.

21. One or more non-transitory computer-readable media that collectively store: a non-autoregressive sequence prediction model configured to process a plurality of input tokens to generate a plurality of output tokens in parallel; wherein the non-autoregressive sequence prediction model comprises a categorical mixture prediction head that receives a latent variable as an input, wherein the latent variable has a plurality of possible states.

22. The one or more non-transitory computer-readable media of claim 21, wherein the non-autoregressive sequence prediction model comprises a transformer model configured to process the plurality of input tokens to generate output embeddings, and wherein the categorical mixture prediction head comprises a multi-layer perceptron that operates on the output embeddings of the transformer model.

Images