Multimodal autoregressive model for time-aligned and contextual modalities

Abstract

The present disclosure provides multimodal processing models that decouple the multimodal modeling, dividing it into separate, focused autoregressive models, processing the inputs according to the characteristics of the modalities and their sampling rates. More specifically, example aspects of the present disclosure are directed a multimodal processing model that includes an autoregressive component for input data belonging to time-synchronized modalities, and an autoregressive component for input data belonging to modalities which are not necessarily aligned in time but are still sequential.

Description

MULTIMODAL AUTOREGRESSIVE MODEL FOR TIME- ALIGNED ANDCONTEXTUAL MODALITIESPRIORITY CLAIM

[0001] The present application is based on and claims priority to United States Provisional Application 63 / 586,200 having a filing date of September 28, 2023, which is incorporated by reference herein.FILED

[0002] The present disclosure relates generally to machine learning. More particularly, the present disclosure relates to a multimodal processing model configured to process input data comprising a mix of time-aligned data and time-unaligned data.BACKGROUND

[0003] One of the main challenges of multimodal models is that they need to combine heterogeneous modalities (e.g., video, audio, text), which have different characteristics. For example, an input can contain multiple modalities of data, where some of the modalities of data are time-aligned, and some of the modalities of data are time-unaligned.

[0004] As one example, for audio-visual data, video and audio are obtained at much higher rates than text and are roughly aligned in time. However, the video and audio are not necessarily time-aligned with text. For example, text may be present as a global context, e.g., a title or description, which is time-unaligned with respect to the audio-visual content. Furthermore, video and audio inputs are of much larger volumes, and grow as the video length increases, which naturally requires more compute dedicated to these modalities and makes modeling of long-range dependencies harder.SUMMARY

[0005] Aspects and advantages of embodiments of the present disclosure will be set forth in part in the following description, or can be learned from the description, or can be learned through practice of the embodiments.

[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 operationsor 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 processing of multimodal data that includes both time-aligned and time-unaligned data. The computing system includes one or more processors. The computing system includes one or more non-transitory computer-readable media that collectively store: a machine-learned multimodal processing model configured to receive and process a multimodal input data The machine-learned multimodal processing model may include: a time-aligned processing model configured to receive and jointly process a first portion of the multimodal input data and a second portion of the multimodal input data to autoregressively generate a sequence of combined feature representations, where the first portion of the multimodal input data is expressed in a first modality and the second portion of the multimodal input data expressed in a second, different modality, where the first portion of the multimodal input data is time- aligned with the second portion of the multimodal input data. The machine-learned multimodal processing model may include a time-unaligned processing model configured to receive and process a third portion of the multimodal input data to autoregressively generate a sequence of context feature representations, where the third portion of the multimodal input data is time-unaligned with the first portion and the second portion of the multimodal input data. In some implementations, the time-aligned processing model receives the sequence of context feature representations as an input. In some implementations, the time-unaligned processing model receives the sequence of combined feature representations as an input. 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 described functions.

[0008] Other aspects of the present disclosure are directed to various systems, apparatuses, non-transitory computer-readable media, user interfaces, and electronic devices.

[0009] These and other features, aspects, and advantages of various embodiments of the present disclosure will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate example embodiments of the present disclosure and, together with the description, serve to explain the related principles.

[0010] The attached Appendix, which is fully incorporated into and forms a portion of this disclosure, describes example implementations of the systems and methods describedherein. The present disclosure is not limited to the example implementations described in the attached Appendix.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Detailed discussion of embodiments directed to one of ordinary skill in the art is set forth in the specification, which makes reference to the appended figures, in which:

[0012] Figure 1 depicts a block diagram of an example multimodal processing model according to example embodiments of the present disclosure.

[0013] Figure 2 depicts a block diagram of an example time-aligned processing model according to example embodiments of the present disclosure.

[0014] Figure 3 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;

[0002] Figure 4 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;

[0003] Figure 5 is a block diagram of an example sequence processing model according to example implementations of aspects of the present disclosure;

[0004] Figure 6 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;

[0005] Figure 7 is a block diagram of an example model development platform according to example implementations of aspects of the present disclosure;

[0006] Figure 8 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;

[0007] Figure 9 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;

[0008] Figure 10 is a block diagram of an example networked computing system according to example implementations of aspects of the present disclosure;

[0009] Figure 11 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure; and

[0015] Figure 12 is a block diagram of an example computing device according to example implementations of aspects of the present disclosure.DETAILED DESCRIPTION

[0016] Multimodal models aim to combine the signals from multiple varied sources, which makes them both universal and important tools for practical applications, e.g., summarizing the main points of a video or answering a question about it, using audio, video and text as inputs. However, these modalities have diverse characteristics and are challenging to combine uniformly by a single model. For example, video and text have disparate sampling rates: a video has many frames per second, but text or other similar types of global context, e.g., descriptions, meta information, question-answer pairs, can be provided once per video, or asynchronously to the video. Furthermore, video is a larger amount of data with many pixels per frame, thus taking a much larger portion of the input. At the same time, video and audio are naturally co-occurring and appear (almost) synchronously. They are roughly aligned and complementary. This co-occurrence in time can contribute to their joint learning and can serve as a rich self-supervisory learning signal, applied much more frequently than global text signals. So, ideally, these modalities need to be processed by differently-synchronized model components, which can more adequately process inputs of different frequencies and can allocate more parameters to the more important input modalities.

[0017] In view of these challenges, the present disclosure is directed to multimodal processing models that decouple the multimodal modeling, dividing it into separate, focused autoregressive models, processing the inputs according to the characteristics of the modalities and their sampling rates. More specifically, example aspects of the present disclosure are directed a multimodal processing model that includes an autoregressive component for input data belonging to time-synchronized modalities, and an autoregressive component for input data belonging to modalities which are not necessarily aligned in time but are still sequential. The descriptions herein focus on an example setting in which the time-aligned input data comprises audio data and video data and in which the time-unaligned input data comprises text data. However, this setting is an example only for purpose of explication, other forms of time-aligned and / or time-unaligned input data can be used as well.

[0018] More particularly, following the success of large language models, where text input sequences are processed autoregressively, many recent multimodal models reuse the autoregressive text models, feeding in other modalities, e.g., as embeddings, or by tokenizingthe visual inputs to be processed together with the text token sequence. However, the imbalance of the information volume is large and models which are well suited to encode / decode text sequences process only highly compressed image or video features. For example, one prior approach subsamples the video features significantly, dedicating less than 1% of the parameters to the image and video inputs, leaving the rest for text processing. Alternatively, methods that process the video by running each frame independently through an encoder or tokenizer, reach an upper bound on the number of frames (e.g., 8 frames) quickly. For longer videos and audio, these representations are insufficient to properly represent the modality or have to sample too many frames to span the input. This inherently limits the representation power of the video, or the ability to model fine-grained or long- range dependencies in it.

[0019] In contrast to these approaches, the present disclosure provides a multimodal model (e.g., audio-video-text) which autoregressively learns time-aligned representations, e.g., audio and video, to combine them in time and coordinates this information with time- unaligned modalities e.g., texts, meta information, and / or others. This decoupling allows for better parameter distribution within the model and sufficient capacity for the media modalities, and leads to smaller models overall.

[0020] Another aspect of the present disclosure relates to learning joint feature representations via a combiner model to balance the need for efficient media representations and ones which are sufficiently expressive to preserve the media content. The proposed approach extracts low-level spatio-temporal representation from the raw media inputs in order to capture the dynamic nature of videos in a high level representation and combines it with semantic and acoustic signals of video within concurrent timesteps.

[0021] In particular, according to one aspect of the present disclosure, to address the long-sequences of video-audio inputs, some example implementations can further partition the video and audio sequences into multiple time chunks and autoregressively process their compact learned representations. To that end, some example implementations can include a combiner model, which models the audio-video information within a time chunk of the input data. The combiner model first learns to extract audio and visual features from raw spatiotemporal signals, and then learns to fuse these features producing compact but expressive representations per time chunk, which are added to the subsequent autoregressive modeling in time.

[0022] The proposed processing approach strikes a balance between compact media representations which can be handled by subsequent autoregressive models and allow long-range dependencies while still sufficiently representing the events and activities in the videos and other concurrent modalities, e.g., audio. Different from prior work, in some implementations, the video / audio is processed sequentially in small pieces, but the information is still accumulated in time. Since video can contain both highly dynamic information and have large redundancies, some example implementations can apply sparse feature extraction, instead of regular-grid visual information previously used. This enables the creation of combined representations where more modalities are learned and combined together. This flexibility allows consuming multimodal inputs at different rates and allows for larger inputs for longer videos.

[0023] Example implementations of the proposed approach achieve the state-of-the-art on well-established multimodal benchmarks, outperforming much larger models. The proposed approach effectively addresses the high computational demand of media inputs by both learning compact representations, controlling the sequence length of the audio-video feature representations, and producing powerful visual and audio representations which allow for long-term dependencies and modeling of long-form video or video / audio inputs in time.

[0024] The systems and methods described herein provide a number of technical effects and benefits. In particular, one aspect of the present disclosure is directed to a multimodal processing model that includes components that are able to separately model time-aligned and time-unaligned data. By performing separate modeling in this fashion, the appropriate parameterization (e.g., number of parameters) can be dedicated to each modality of data, which may vary in length, volume, or density. For example, the time-aligned processing model can be scaled to handle longer input sequences, without needing to also scale parameters dedicated to processing the time-unaligned data. This can enable the creation of models that are “right-sized” for handling multimodal processing tasks. Stated differently, by enabling the separate optimization of model size for each modality, the proposed models can avoid over-parameterizing certain modalities. Over-parameterization results in wasted computational resources and, therefore, the present disclosure provides a technical benefit of conserving computational resources such as processor cycles, memory usage, network bandwidth, etc.

[0025] As another example technical effect, the multimodal processing model includes components that are able to separately model time-aligned and time-unaligned data. By performing separate modeling in this fashion, at least some of the processing of these different data types can, in some cases, be performed in parallel. Performing processing in parallel can result in reduced latency. Further, by performing separate modeling in thisfashion, at least some of the processing of these different data types can, in some cases, be performed on or by different computer hardware that is optimized for that particular type of data. For example, certain hardware accelerators or other hardware components may be optimized or otherwise more computationally efficient at processing different types of data. Therefore, by separately modeling time-aligned and time-unaligned data, the most appropriate hardware can be applied to each type of data, resulting in computational benefits.

[0026] The multimodal processing models described herein can be applied to a number of different applications or tasks. In some implementations, the multimodal processing model can perform multimodal sentiment analysis, where the model analyzes sentiment from text, audio, and visual data simultaneously, such as in social media monitoring and product reviews analysis. In some implementations, the multimodal processing model can perform human-computer interaction (HCI), using speech, gesture, and visual input to understand and respond to user commands and interactions. In some implementations, the multimodal processing model can perform tasks for multimodal search engines, employing text, image, and audio processing to provide accurate and relevant search results. In some implementations, the multimodal processing model can perform audio-visual speech recognition (AVSR), combining audio speech recognition with visual information like lip movement to improve accuracy, particularly in noisy environments. In some implementations, the multimodal processing model can perform tasks in health monitoring systems, integrating data from different modalities like text, images, and signals to monitor individual or population health conditions. In some implementations, the multimodal processing model can perform multimodal image captioning, generating descriptive text for images or videos by analyzing visual and contextual information. In some implementations, the multimodal processing model can perform video summarization, extracting key information from videos using both visual and audio data to create concise summaries. In some implementations, the multimodal processing model can perform robot navigation, utilizing visual, auditory, and sensory information to navigate robots through different environments. In some implementations, the multimodal processing model can perform cross-modal retrieval, retrieving related information from one modality using a query from another, like retrieving images using text queries. In some implementations, the multimodal processing model can perform natural language understanding, integrating text with other modalities like audio and images to understand the context better.

[0027] With reference now to the Figures, example embodiments of the present disclosure will be discussed in further detail.Example Techniques

[0028] Autoregressive models are powerful generative models that are well suited for data which appears in a sequence, modeling the probability of the current value, conditioned of previous ones. Video and audio information is sequential but also roughly time- synchronized. At the same time, other modalities e.g., text, might be provided globally per video as context and applied to the full video rather than to specific parts. To address the challenges of modeling diverse multimodal inputs, some example implementations of the present disclosure can subdivide the autoregressive modeling by learning separate autoregressive models: one for the time-aligned modalities (e.g., audio-video), and another one for modalities which are not necessarily aligned in time but are still sequential (e.g., text). Learning across these can be coordinated by cross-attention mechanisms. In some implementations, the time-aligned inputs can be allocated relatively more parameters and are properly modeled in time. A learning module or model portion, which can be called the Combiner, combines the lower-level signals from time-aligned snippets. Here information can be processed spatio-temporally, extracting features particularly relevant to dynamic changes in the inputs.

[0029] Overview of example architecture: At a high level, one example architecture can include two main learning components: The first one is an autoregressive component which is designed to process (almost) synchronized multimedia inputs e.g., video+audio and combine their inputs in time. In order to process the video and audio signals, and to accommodate longer video / audio inputs, they can be partitioned into smaller chunks (e.g., roughly synchronized in time) for which a joint audio-visual representation is learned via the Combiner as described herein. The second component processes the context, or the signals not aligned in time, such as global text information, which are often still sequential. It can be autoregressive as well, and can use the combined latent space as cross-attention inputs.

[0030] Example model inputs: Some example implementations may receive an input f f f video sequence of N frames v = {v( , v2, ... vN}, and audio wave signal of M timesteps a = f f f{a' , a2, ... aM}, where the audio signal is captured during the duration of the video and corresponds to the given video input. Additionally the example input can include an input f f f text sequence t = {t[, t2, ... t],}, which is related to the video / audio and might vary according to the task e.g. it can be a description, a question-answer pair, meta information, or other textual data.

[0031] Example partitioning of the media inputs: In order to process the video sequence efficiently and to learn the correlation of features in time, some example implementaitnos can partition the input video into T non-overlapping segments or chunks, with vtand atdenoting the video and audio input per segment (let K = N / T). Here each chunk captures all input data between two timestamps (e.g., video and audio snippets), as follows:

[0032] Thus the video can be represented by its chunks instead, v = {v1(v2, ... vT], and more specifically latent features will be extracted from each chunk to represent the video. A similar partitioning is done for the audio signals, where they are partitioned in T chunks to be synchronized in time to the video chunks, a = {a1(a2, ... aT}. Audio features can be extracted from the raw audio signal.

[0033] Example audio / video features

[0034] Example video features: Prior models captured video information at individual sparsely sampled frames, which lacks the temporal information essential to video understanding and which might miss dynamic events. Alternatively, 3D convolutions, sparse 3D tubes, and others learn spatio-temporally, which can capture key dynamic changes in the videos. Example implementations can extract sparse 3D tubes from the videos which span all 3 dimensions of the video snippet starting at various locations. The sparse 3D tubes, together with standard 2D patches can be processed via a ViT encoder. Rather than applying offsets, the tubes can be applied at the beginning of the snippet. Per video chunk t, denote vtas the time-aligned features for this chunk, and thus v = {v1(v2, ... , vT) are the time-aligned video representations for the whole video.

[0035] Example audio features: Audio inputs arrive at a predefined frequency and can be processed in various ways. As one example, some implementations can represent the audio as a spectrogram. The spectrogram can be created so that the time bands match the frames per second (e.g., 25 fps) used in the videos, and thus can easily be split into snippets aligned with the video. The spectrogram for each snippet can be processed by a ViT model, after an audio input projection layer. The ViT backbone can, in some cases, be the same as the one used for video features. Reusing the visual component is previously shown to be advantageous. Similarly to above, denote atto be the audio feature per chunk t and a = {a1(a2, ... , aT) for the full video.

[0036] Example Modality Combiners

[0037] The task of the Combiner module is two-fold: 1) to combine the video (and audio) features at a specific snippet of time, learning their joint representation and 2) effectively compress the representation from each video / audio snippet, which allows our model to scale to longer videos.

[0038] When partitioning the inputs, the features for each modality, video and audio in this case, are (roughly) time-aligned latent features v = {v1(v2, vT) and a ={a1(a2, ... , aT), where, in some implementations, the maximum timestamp for any data incorporated into vtor atis less than the minimum timestamp of any data incorporated into vt+1or at+1. Explicitly vtis composed of f features of size d giving it a shape of ( , d) and atis composed of s features also of size d with shape (s, d). The role of the Combiner is to map such time-aligned modal latent features into a smaller set of shared latent features. Specifically let u = {u1(u2, ... , uT] where ut= (vt, at) having size (n, d) and n = f + s be the set of all time-aligned features from all modalities. The Combiner then maps u to a lower dimensional latent feature space x = {x1, x2, ... , xT] where xthas shape (m, d) where, in some implementations, n » m.

[0039] Since the Combiner is reducing the dimensionality of video+audio features, it can effectively use all features in the sequence, not only the ones per chunk. However, since the features produced by the Combiner are, in some implementations, going to be used in the sequential autoregressive modeling of video / audio, it is important for the Combiner to not break causality thus, in some implementations: xt= Combiner(u1, u2, ... , ut) (2)

[0040] The Combiner may have various different architectures. Two example architectures for the Combiner include a standard Transformer one and a memory based one, based on the Token Turing Machines, to reduce memory.

[0041] Example Causal Transformer Combiner: One example Combiner includes a standard Transformer model, here of R layers (e.g., R = 8). For each step t it maps the original set of features utto xtwhere xtis of much lower dimensionality, i.e., effectively reducing the number of tokens (e.g., m = 32). The inputs to the Combiner are the latent features of the video and audio, which can be concatenated before being fed to the Combiner. A causal version of the transformer can mask out inputs from future timestamps (i.e., > t). The attention mechanism of the transformer can be modified to mask features at the timechunk level, thus all features from utand the previous time steps are used to compute eachoutput feature in xtas in Eq. (2). This effectively applies attention mechanisms to all the modality inputs jointly while respecting causality.

[0042] Example Token Turing Machine Combiner: The Token Turing Machine (TTM) (See Ryoo et al. (2023)) is a recurrent sequential model with Transformers and token-based operations. It maintains an external ‘memory’ Mtas a set of features, and updates it at every time step by reading and writing. Given inputs utat each time step, it first ‘reads’ features to be processed, from input features as well as memory features. Such features, zt, are passed to the ‘processor’, which can be implemented as a standard Transformer, generating a set of intermediate output features ot. These intermediate outputs are then used to update Mt(i.e., memory ‘write’) as well as to produce the final output xt. zt= Read(ut, Mt) (3) ot= Process(zt) (4) Mt+i= Write(Mt, ot, ut) (5) xt= Output(ot) (6)

[0043] A central idea is to make the TTM processor generate the outputs by utilizing memory Mtinstead of the entire history of features {u1(u^ }. Once trained, the differentiable read and write operations are optimized to maintain Mtso that it stores important features from the previous time steps {u1(and updates it, at every step.

[0044] Some example implementations of the present disclosure can implement TTM as the Combiner module to sequentially combine u. The function ‘Process’ can be implemented with a standard Transformer with layers of multi-head self-attention and MLPs. The functions ‘Read’, ‘Write’, and ‘Output’ can be implemented with TokenLearner (See Ryoo et al. (2021)). Note that some implementations are able to separately control the number of features in the memory as well as the number of ‘Output’ function features, allowing efficient Combiner computation and feature generation.

[0045] One advantage of the TTM Combiner is its utilization of memory features to sequentially process ut. The number of such memory features are much smaller than the total number of history features ({u1;u^ }) in general (e.g., 256 vs. ~ 10k). This not only makes TTM a natural fit for the model, but also reduces the total time complexity of the TTM Combiner to be constant with respect to t, instead of O(t) or O(t2) in Transformers. The TTM Combiner saves memory in both training and inference, using about 30% less memory and reduces the runtime by about 18%.

[0046] Example Time-Aligned Video / Audio Autoregressive Modeling

[0047] This subsection describes the autoregressive modeling of time-aligned video and audio. Some example implementations can apply an autoregressive modeling strategy where the model conditions video / audio representations corresponding to a time interval on feature representations from previous time intervals. These representations can be learned jointly by the Combiner. As mentioned, the video can be first partitioned in T smaller video snippets. Each of the snippets itself can be, e.g., of size 4-64 frames (overlap is possible but not required). Some example implementations extract spatio-temporal information into latent video features vtand audio features atfrom the same video partition, apply Combiner to produce xt. The feature representations per video chunk xtare then fed sequentially to the autoregressive model, where at each step the features are reconstructed from the previous step, conditioned on the prior inputs and the latent vector h which corresponds to the latent representations learned within the autoregressive model:where {v1(v2, ... vT], and {a1(a2, ... aT] are the feature representations from the video and audio, p(xt|vt, at) is estimated by the Combiner, and p(ht|xt) is estimated by the latent causal model, p(vt+1, at+1|ht) by the modality reconstruction model. This allows for learning from previous representations in the sequence (in time) and aims to predict the nextstep feature representation.

[0048] While autoregressive modeling has been used for videos and images, it is often done on a pixel-by-pixel basis, which is highly inefficient and captures only short-term dependencies. With the proposed approach, autoregressive modeling and the Combiner address both shortcomings. Note that the Combiner also accumulates information from prior chunks; however, the autoregressive model works at a higher level of abstraction with already learned features from the Combiner. Both mechanisms work together beneficially.

[0049] Example Latent Causal Modeling: The autoregressive latent model estimates: n =i P(ht|xt). This can be done by applying an autoregressive transformer to x ={x1(x2, ... , xT] to produce h = {h1(h2, ... , hT] where the target of htis xt+1so the difference between x2 Tand hx.... r-i is used as a loss to control the latent representation of theCombiner output x. Since example implementations are modeling data autoregressively in time, care must be taken with the attention mechanism during training, the transformer can use a modified attention mechanism.

[0050] Example Modality Reconstruction: The autoregressive modality reconstruction models estimate. This can be done by applying a separate transformer to h to produce reconstructions of the audio and video signals v and a, which can be added as an optional loss. To save on computation, the video input v can be down sampled to vsmal1for the reconstruction target, thus the actual reconstruction can be vsmal1.

[0051] Example Attention Mechanisms for Autoregressive Modeling

[0052] Since the autoreggressive models are trained in time, masking can be done to satisfy causality. The attention mechanisms within and across chunks can in some cases be modified when masking. This can apply to both the Combiner and the Autoregressive learning. When masking features for autoregressive modeling, the standard pattern of masking each feature individually would mask features from within the same time-chunk from each other. While this would still satisfy causality, it unnecessarily restricts the model, preventing features from within the same time-chunk from interacting based on position within the time-chunk. To allow features in the same chunk to interact, the autoregressive mask between all features i, which fall in a time-chunk t, and another feature j can be computed as follows (A is the number of features and T the number of time-chunks): mask‘ = (° i <= ^(f> T / N) , N / T JI otherwise.

[0053] Example Approach for Combining Aligned and Non-aligned Autoregressive Modeling

[0054] Text, or other context information, might not necessarily be aligned in time with the video and audio modalities. It is still sequential. So here it can be modeled by a separate autoregressive model, devoted to text representations and to combining the visual-audio f f f information together. Assuming tokenization for the input text t = {t] , tJ2, ... tp] is provided, obtaining a tokenized text sequence t = {w1(w2, ... wL] of length L, some example implementations model the text sequentially as conditioned on audio and video. In order to combine the outputs of the video / audio autoregressive model some example implementations can use cross-attention strategy. Here, unlike prior work, all feature representations h = {h!, h2hT} from the latent causal model can, in some implementations, be used in the main text model.

[0055] The autoregressive text model can estimate Eq. (9) by applying a transformer to the input text sequence w = {w1(w2, ... , wL] and using the latent model output h as crossattention to produce w. The loss can be the standard cross-entropy loss between target w and output text sequences w. This can provide further feedback to the Combiner latent representation h through the cross-attention layer. Of note is that since all parts of the model are autoregressive, it is naturally applicable to streaming videos.

[0056] Example Model Losses

[0057] Various model losses can be used. Some example implementations use two main losses, each driving the corresponding autoregressive model:

[0058] A latent space reconstruction loss for time-aligned inputs is the difference between .2,...;r and hx.... r-i in the autoregressive setting such that ht= xt+1. Some example implementations apply a L2normalization and then take dot product between the feature vectors as the loss (i.e., cosine similarity).

[0059] An unaligned text cross entropy loss is the standard cross-entropy loss between w and w for the unaligned text output.

[0060] Additionally, some example implementations include a loss to encourage modality reconstruction. More specifically, some example implementations add a video reconstruction loss which is commonly used for video (audio reconstruction loss can also be added). Similar to latent space reconstruction above, the video reconstruction loss approximates the difference between vsmal1and vsmal1also in an autoregressive setting such that vtsmal1Some example implementations use the same distance metric on the video reconstruction as on the latent space reconstruction problem. While this loss can be useful, especially for generation tasks, it is not required.

[0061] Various combinations of the losses described above (and potentially others alternatively or additionally) can be weighted to compute the final loss.

[0062] Example Implementation Details

[0063] Example Models: One example model has 3B parameters; without audio it is 2.9B. A little over half of the parameters are for the audio+video autoregressive model. Example models work on 128 frames customarily (16 chunks, 8 frames), but can handle more for longer videos (e.g., 512=16 chunks x 32 frames). Some example implementations can use Combiner dimension m = 32. Some example implementations can apply random masking to the Combiner output features at a ratio of 0.75% as a form of dropout regularization as we found this stabilizes the causal model latent reconstruction. Due to the design of examplemodels (partitioning and Combiner), adding more frames, or increasing the chunk size, number of chunks, etc. lead to only marginal increase in parameters. Increasing the number of chunks, while not leading to parameter increases, increases memory, which underscores the importance of the Combiner and particularly the TTM.

[0064] Example Model Training: Example models can be pretrained on the Video-Text Pairs (VTP) dataset which is collected from noisy video-text pairs from the web. Some example implementations can use only about 12% of the data, 3M samples. All losses can be given equal weight during pretraining. During finetuning the unaligned text loss can be increased 10-fold to better align the training loss with the final evaluation.Example Model Visualizations

[0065] Figure 1 depicts a block diagram of an example multimodal processing model 100 according to example embodiments of the present disclosure. The multimodal processing model 100 can be configured to receive and process a multimodal input data to generate a model prediction 118. The multimodal processing model 100 can include a time-aligned processing model 108 and a time-unaligned processing model 110.

[0066] The time-aligned processing model 108 can be configured to receive and jointly process a first portion 102 of the multimodal input data and a second portion 104 of the multimodal input data to autoregressively generate a sequence of combined feature representations 112. For example, the first portion 102 of the multimodal input data can be expressed in a first modality and the second portion 104 of the multimodal input data expressed in a second, different modality, and where the first portion 102 of the multimodal input data is time-aligned with the second portion 104 of the multimodal input data.

[0067] The time-unaligned processing model 110 can be configured to receive and process a third portion 106 of the multimodal input data to autoregressively generate a sequence of context feature representations 114. The third portion 106 of the multimodal input data is time-unaligned with the first portion 102 and the second portion 104 of the multimodal input data. The third portion 106 of the multimodal input data may be expressed in a third modality. The third modality may be the same as or different from the first modality and / or the second modality.

[0068] As one example, the first input modality can be imagery, the second input modality can be audio, and the third input modality can be text. Other combinations of modalities are possible as well. For example, the first and / or second modality can be any form of time series or tabular data. The third modality can be any form of context data.

[0069] In some implementations, the time-aligned processing model 112 receives the sequence of context feature representations 114 as an input. For example, cross-attention can be performed between the context feature representations 114 and the internal representations generated by the time-aligned processing model 108.

[0070] In some implementations, the time-unaligned processing model 110 receives the sequence of combined feature representations as an input. For example, cross-attention can be performed between the combined feature representations 112 and the internal representations generated by the time-unaligned processing model 114.

[0071] In some implementations, the combined feature representations 112 and / or the context feature representations 114 can be provided as an output of the multimodal processing model 100. For example, the representations 112 and / or 114 can represent and / or encode a meaningful and useful representation of the multimodal input data in a learned latent space. For example, the representations 112 and / or 114 can be referred to as “embeddings”. These embeddings can be used for various tasks or predictions or may be useful in their own right (e.g., for performing clustering, similarity measurement, etc.).

[0072] Additionally or alternatively to providing the combined feature representations 112 and / or the context feature representations 114 as an output, in some implementations, the multimodal processing model 100 can also a prediction model 116. The prediction model 116 can be configured to generate a model prediction 118 based on one or both of: the sequence of combined feature representations 112 and the sequence of context feature representations 114. The model prediction 118 can be a specific prediction or other output for any number of different multimodal processing tasks, examples of which are described elsewhere herein.

[0073] Figure 2 depicts a block diagram of an example time-aligned processing model according to example embodiments of the present disclosure. The example model shown in Figure 2 provides one possible example of the model 108 shown in Figure 1.

[0074] As illustrated in Figure 2, the time-aligned processing model includes a first modality encoder, a second modality encoder, a combiner model, and an autoregressive prediction model. For example, one or both of the first modality encoder and the second modality encoder can be or include a vision transformer.

[0075] As illustrated in Figure 2, the time-aligned processing model can operate iteratively over a number of time chunks for the input data. Specifically, for each of the plurality of time chunks of the first portion and the second portion of the multimodal input data: the first modality encoder can be configured to encode the first portion of the multimodal input data included in the time chunk to one or more first features. The secondmodality encoder can be configured to encode the second portion of the multimodal input data included in the time chunk to one or more second features. The combiner model can be configured to combine the one or more first features and the one or more second features into one or more intermediate combined features.

[0076] Then, the autoregressive prediction model can be configured to predict a next combined feature representation in the sequence of combined feature representations based on the one or more intermediate combined features generated for the time chunk and one or more prior combined feature representations in the sequence of combined feature representations.

[0077] As one example, the combiner model can be a token turing machine combiner. As another example, the combiner model can be a causal transformer combiner.

[0078] Referring collectively to Figures 1 and 2, in some implementations, machine- learned multimodal processing model has been trained using a latent space reconstruction loss. For example, this loss can evaluate and provide a loss signal based on the ability of the time-aligned processing model to correctly predict the intermediate combined features for each time chunk.

[0079] In some implementations, the machine-learned multimodal processing model has been trained using a video reconstruction loss. For example, this loss can evaluate and provide a loss signal based on the ability of the encoder that is handling the video data to correctly predict the feature(s) extracted from the video data.

[0080] In some implementations, the machine-learned multimodal processing model has been trained using an unaligned text cross entropy loss. For example, this loss can evaluate and provide a loss signal based on the ability of the model(s) that are handling the text data to correctly predict the feature(s) extracted from the text data. For example, in some instances this may correspond to the ability of the time-unaligned model 110 to correctly predict the sequence of context representations 114.Example Methods

[0081] FIG. 3 depicts a flowchart of a method 300 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 multimodal processing model.

[0082] One or more portion(s) of example method 300 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 ofexample method 300 can be performed by any (or any combination) of one or more computing devices. Moreover, one or more portion(s) of example method 300 can be implemented on the hardware components of the device(s) described herein, for example, to train one or more systems or models. FIG. 3 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. FIG. 3 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 300 can be performed additionally, or alternatively, by other systems.

[0083] At 302, example method 300 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 300 as a “training” instance, it is to be understood that runtime inferences can form training instances when a model is trained using an evaluation of the model’s performance on that runtime instance (e.g., online training / learning). Example data types for the training instance and various tasks associated therewith are described throughout the present disclosure.

[0084] At 304, example method 300 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.

[0085] At 306, example method 300 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).

[0086] At 308, example method 300 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 300 can include implementing a number of generalization techniques (e.g., weight decays, dropouts, etc.) to improve the generalization capability of the models being trained.

[0087] In some implementations, example method 300 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.).

[0088] In some implementations, example method 300 can be implemented for particular stages of a training procedure. For instance, in some implementations, example method 300 can be implemented for pre-training a machine-learned model. Pre-training can include, for instance, large-scale training over potentially noisy data to achieve a broad base of performance levels across a variety of tasks / data types. In some implementations, example method 300 can be implemented for fine-tuning a machine-learned model. Fine-tuning can include, for instance, smaller-scale training on higher-quality (e.g., labeled, curated, etc.) data. Fine-tuning can affect all or a portion of the parameters of a machine-learned model. For example, various portions of the machine-learned model can be “frozen” for certain training stages. For example, parameters associated with an embedding space can be “frozen” during fine-tuning (e.g., to retain information learned from a broader domain(s) than present in the fine-tuning dataset(s)). An example fine-tuning approach includes reinforcement learning. Reinforcement learning can be based on user feedback on model performance during use.Example Machine-Learned Models

[0089] Figure 4 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.

[0090] Machine-learned model(s) 1 can be or include one or multiple machine-learned models or model components. Example machine-learned models can include neural networks (e.g., deep neural networks). Example machine-learned models can include non-linear models or linear models. Example machine-learned models can use other architectures in lieu of or in addition to neural networks. Example machine-learned models can include decision tree based models, support vector machines, hidden Markov models, Bayesian networks, linear regression models, k-means clustering models, etc.

[0091] 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.

[0092] Machine-learned model(s) 1 can include a single or multiple instances of the same model configured to operate on data from input(s) 2. Machine-learned model(s) 1 can include an ensemble of different models that can cooperatively interact to process data from input(s) 2. For example, machine-learned model(s) 1 can employ a mixture-of-experts structure. See, e.g., Zhou et al., Mixture-of-Experts with Expert Choice Routing, ARXIV:2202.09368V2 (Oct. 14, 2022).

[0093] 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.

[0094] 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 measurementvalues 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.Example Machine-Learned Sequence Processing Models

[0095] Figure 5 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-A , etc. obtained from input(s) 2. Sequence processing model 4 can process input sequence 5 using prediction layer(s) 6 to generate an output sequence 7. Output sequence 7 can include one or more output elements 7-1, 7-2, . . . , 7-A, etc. generated based on input sequence 5. The system can generate output(s) 3 based on output sequence 7.

[0096] 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:2010.11929v2 (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.

[0097] 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 sequenceprocessing 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”).

[0098] Sequence processing model(s) 4 can ingest the data from input(s) 2 and parse the data into a sequence of elements to obtain input sequence 5. For example, a portion of input data from input(s) 2 can be broken down into pieces that collectively represent the content of the portion of the input data. The pieces can provide the elements of the sequence.

[0099] 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.

[0100] 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.

[0101] 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 5 can be the tokens or can be the embedded representations thereof.

[0102] Prediction layer(s) 6 can predict one or more output elements 7-1, 7-2, . . . , 1-N based on the input elements. Prediction layer(s) 6 can include one or more machine-learned model architectures, such as one or more layers of learned parameters that manipulate and transform the input(s) to extract higher-order meaning from, and relationships between, input element(s) 5-1, 5-2, . . . , 5-M. In this manner, for instance, example prediction layer(s) 6 can predict new output element(s) in view of the context provided by input sequence 5.

[0103] 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 .” Exampleprediction 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.”

[0104] A transformer is an example architecture that can be used in prediction layer(s) 4. See, e.g., Vaswani et al., Attention Is All You Need, ARXIV: 1706.03762V7 (Aug. 2, 2023). A transformer is an example of a machine-learned model architecture that uses an attention mechanism to compute associations between items within a context window. The context window can include a sequence that contains input sequence 5 and potentially one or more output element(s) 7-1, 7-2, . . . , 1-N. A transformer block can include one or more attention layer(s) and one or more post-attention layer(s) (e.g., feedforward layer(s), such as a multilayer perceptron).

[0105] 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.

[0106] 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.

[0107] 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.

[0108] Output sequence 7 can be generated autoregressively. For instance, for some applications, an output of one or more prediction layer(s) 6 can be passed through one ormore output layers (e.g., softmax layer) to obtain a probability distribution over an output vocabulary (e.g., a textual or symbolic vocabulary) conditioned on a set of input elements in a context window. In this manner, for instance, output sequence 7 can be autoregressively generated by sampling a likely next output element, adding that element to the context window, and re-generating the probability distribution based on the updated context window, and sampling a likely next output element, and so forth.

[0109] Output sequence 7 can also be generated non-autoregressively. For instance, multiple output elements of output sequence 7 can be predicted together without explicit sequential conditioning on each other. See, e.g., Saharia et al., Non-Autoregressive Machine Translation with Latent Alignments, ARXIV:2004.07437V3 (NOV. 16, 2020).

[0110] 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.

[0111] Figure 6 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.

[0112] Input sequence 8 can be the same as or different from input sequence 5. Input sequence 8 can be a multimodal input sequence that contains elements that represent data from different modalities using a common dimensional representation. For instance, an embedding space can have P dimensions. Input sequence 8 can be configured to contain a plurality of elements that have / Jdimensions. In this manner, for instance, example implementations can facilitate information extraction and reasoning across diverse data modalities by projecting data into elements in the same embedding space for comparison, combination, or other computations therebetween.

[0113] 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.

[0114] 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.

[0115] 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 beprovided as a data type associated with an input modality and projected along with that input modality (e.g., the input value can be a textual task label that is embedded along with other textual data in the input; the input value can be a pixel-based representation of a task that is embedded along with other image data in the input; etc.). The input value can be provided as a data type that differs from or is at least independent from other input(s). For instance, the input value represented by element 8-0 can be a learned within a continuous embedding space.

[0116] 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).

[0117] 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.).

[0118] Data-to-sequence models 11-1, 11-2, and 11-3 can form part of machine-learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be jointly trained with or trained independently from machine-learned sequence processing model(s) 4. Data-to-sequence models 11-1, 11-2, and 11-3 can be trained end-to-end with machine- learned sequence processing model(s) 4.Example Machine-Learned Model Development Platform

[0119] Figure 7 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.

[0120] 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.

[0121] 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.

[0122] 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.

[0123] 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).

[0124] 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.

[0125] 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.

[0126] Fine-tuning pipelines 17-3 can include a machine-learned model training workflow configured to refine the model parameters of development model 16 with higher-quality data. Fine-tuning pipelines 17-3 can update development model 16 by conducting supervised training with labeled dataset(s) in dataset(s) 17-1. Fine-tuning pipelines 17-3 can update development model 16 by conducting reinforcement learning using reward signals from user feedback signals. Workbench 15 can implement a fine-tuning pipeline 17-3 to finetune development model 16.

[0127] 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.

[0128] 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.

[0129] 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).

[0130] 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.

[0131] 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.

[0132] 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.

[0133] Although various training examples described herein with respect to model development platform 12 refer to “pre-training” and “fine-tuning,” it is to be understood that model alignment toolkit 17 can generally support a wide variety of training techniques adapted for training a wide variety of machine-learned models. Example training techniques can correspond to the example training method 300 described above.

[0134] 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.

[0135] 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”).

[0136] 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.

[0137] 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 implementingtool 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.

[0138] Model plugin toolkit 18 can integrate with prompt libraries 17-4 to build a catalog of available tools for use with development model 16. For instance, a model can receive, in an input, a catalog of available tools, and the model can generate an output that selects a tool from the available tools and initiates a tool call for using the tool.

[0139] 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.

[0140] 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.

[0141] Figure 8 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 anycombination) 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. 8 depicts elements performed in a particular order for purposes of illustration and discussion. Those of ordinary skill in the art, using the disclosures provided herein, will understand that the elements of any of the methods discussed herein can be adapted, rearranged, expanded, omitted, combined, or modified in various ways without deviating from the scope of the present disclosure. FIG. 8 is described with reference to elements / terms described with respect to other systems and figures for exemplary illustrated purposes and is not meant to be limiting. One or more portions of the example training flow can be performed additionally, or alternatively, by other systems.

[0142] 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.

[0143] 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).

[0144] 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.

[0145] 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 feedback26. 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.

[0146] In some implementations, computational optimization operations can be applied before, during, or after each stage. For instance, initialized model 21 can undergo computational optimization 29-1 (e.g., using computational optimization toolkit 19) before pre-training stage 22. Pre-trained model 23 can undergo computational optimization 29-2 (e.g., using computational optimization toolkit 19) before fine-tuning stage 24. Fine-tuned model 25 can undergo computational optimization 29-3 (e.g., using computational optimization toolkit 19) before refinement with user feedback 26. Refined model 27 can undergo computational optimization 29-4 (e.g., using computational optimization toolkit 19) before output to downstream system(s) 28. Computational optimization(s) 29-1, . . . , 29-4 can all be the same, all be different, or include at least some different optimization techniques.Example Machine-Learned Model Inference System

[0147] Figure 9 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.

[0148] 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.

[0149] Model host 31 can leverage various other resources and tools to augment the inference task. For instance, model host 31 can communicate with tool interfaces 35 to facilitate tool use by model instance(s) 31-1. Tool interfaces 35 can include local or remote APIs. Tool interfaces 35 can include integrated scripts or other software functionality. Model host 31 can engage online learning interface(s) 36 to facilitate ongoing improvements to machine-learned model(s) 1. For instance, online learning interface(s) 36 can be used within reinforcement learning loops to retrieve user feedback on inferences served by model host 31.Model host 31 can access runtime data source(s) 37 for augmenting input(s) 2 with additional contextual information. For instance, runtime data source(s) 37 can include a knowledge graph 37-1 that facilitates structured information retrieval for information associated with input request(s) 33 (e.g., a search engine service). Runtime data source(s) 37 can include public or private, external or local database(s) 37-2 that can store information associated with input request(s) 33 for augmenting input(s) 2. Runtime data source(s) 37 can include account data 37-3 which can be retrieved in association with a user account corresponding to a client 32 for customizing the behavior of model host 31 accordingly.

[0150] 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.

[0151] 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.

[0152] 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.

[0153] 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 aKV 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.

[0154] Compute resource(s) 31-2 can include one or more processors (central processing units, graphical processing units, tensor processing units, machine-learning accelerators, etc.) connected to one or more memory devices. Compute resource(s) 31-2 can include a dynamic pool of available resources shared with other processes. Compute resource(s) 31-2 can include memory devices large enough to fit an entire model instance in 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.

[0155] 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.

[0156] 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.

[0157] 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.

[0158] Online learning interface(s) 36 can facilitate reinforcement learning of machine- learned model(s) 1. Online learning interface(s) 36 can facilitate reinforcement learning withhuman feedback (RLHF). Online learning interface(s) 36 can facilitate federated learning of machine-learned model(s) 1.

[0159] 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.

[0160] In some implementations, the task is a computer vision task. In some cases, input(s) 2 includes pixel data for one or more images and the task is an image processing task. For example, the image processing task can be image classification, where the output is a set of scores, each score corresponding to a different object class and representing the likelihood that the one or more images depict an object belonging to the object class. The image processing task may be object detection, where the image processing output identifies one or more regions in the one or more images and, for each region, a likelihood that region depicts an object of interest. As another example, the image processing task can be image segmentation, where the image processing output defines, for each pixel in the one or more images, a respective likelihood for each category in a predetermined set of categories. For example, the set of categories can be foreground and background. As another example, the set of categories can be object classes. As another example, the image processing task can be depth estimation, where the image processing output defines, for each pixel in the one or more images, a respective depth value. As another example, the image processing task can be motion estimation, where the network input includes multiple images, and the imageprocessing 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.

[0161] 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).

[0162] In some implementations, input(s) 2 can be or otherwise represent speech data (e.g., data describing spoken natural language, such as audio data, textual data, etc.). Machine-learned model(s) 1 can process the speech data to generate an output. As an example, machine-learned model(s) 1 can process the speech data to generate a speech recognition output. As another example, machine-learned model(s) 1 can process the speech data to generate a speech translation output. As another example, machine-learned model(s) 1 can process the speech data to generate a latent embedding output. As another example, machine-learned model(s) 1 can process the speech data to generate an encoded speech output (e.g., an encoded and / or compressed representation of the speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate an upscaled speech output (e.g., speech data that is higher quality than the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a textual representation output (e.g., a textual representation of the input speech data, etc.). As another example, machine-learned model(s) 1 can process the speech data to generate a prediction output.

[0163] 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.

[0164] 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.

[0165] 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.

[0166] 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 orstorage (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.

[0167] 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.

[0168] 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.

[0169] In some implementations, the task can be an instruction following task. Machine- learned model(s) 1 can be configured to process input(s) 2 that represent instructions to perform a function and to generate output(s) 3 that advance a goal of satisfying the instruction function (e.g., at least a step of a multi-step procedure to perform the function). Output(s) 3 can represent data of the same or of a different modality as input(s) 2. For instance, input(s) 2 can represent textual data (e.g., natural language instructions for a task to be performed) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). Input(s) 2 can represent image data (e.g., image-based instructions for a task to be performed, optionally accompanied by textual instructions) and machine-learned model(s) 1 can process input(s) 2 to generate output(s) 3 that represent textual data responsive to the instructions (e.g., natural language responses, programming language responses, machine language responses, etc.). One or more output(s) 3 can be iteratively or recursively generated to sequentially process and accomplishsteps 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.

[0170] 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.

[0171] 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).

[0172] 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).

[0173] In some implementations, the task can be a data generation task. Machine-learned model(s) 1 can be configured to process input(s) 2 that represent context regarding a desired portion of data (e.g., data from various data domains, such as sensor data, image data, multimodal data, statistical data, etc.). The desired data can be, for instance, synthetic data for training other machine-learned models. The context can include arbitrary data type(s). Machine-learned model(s) 1 can be configured to generate output(s) 3 that represent data that aligns with the desired data. For instance, machine-learned model(s) 1 can be configured to generate data values for populating a dataset. Values for the data object(s) can be selected based on the context (e.g., based on a probability determined based on the context).Example Computing Systems and Devices

[0174] Figure 10 depicts a block diagram of an example networked computing system that can perform aspects of example implementations of the present disclosure. The system can include a number of computing devices and systems that are communicatively coupled over a network 49. An example computing device 50 is described to provide an example of a computing device that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). An example server computing system 60 is described as an example of a server computing system that can perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Computing device 50 and server computing system(s) 60 can cooperatively interact (e.g., over network 49) to perform any aspect of the present disclosure (e.g., implementing model host 31, client(s) 32, or both). Model development platform system 70 is an example system that can host or serve model development platform(s) 12 for development of machine-learned models. Third-party system(s) 80 are example system(s) with which any of computing device 50, server computing system(s) 60, or model development platform system(s) 70 can interact in the performance of various aspects of the present disclosure (e.g., engaging third-party tools, accessing third-party databases or other resources, etc.).

[0175] 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 10 can be co-located with, contained by, or otherwise integrated into one or more other devices or systems.

[0176] 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).

[0177] 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.

[0178] 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.

[0179] 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.

[0180] 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.

[0181] 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.

[0182] Server computing system 60 can store or otherwise include one or more machine- learned models 65. Machine-learned model(s) 65 can be the same as or different from machine-learned model(s) 55. Machine-learned models 65 can include one or more machine- learned model(s) 1, such as a sequence processing model 4. Machine-learned models 65 can include one or multiple model instance(s) 31-1. Machine-learned model(s) 65 can be received from computing device 50, model development platform system 70, third party system(s) 80, or developed locally on server computing system(s) 60. Machine-learned model(s) 65 can be loaded into memory 62 and used or otherwise implemented by processor(s) 61. Server computing system(s) 60 can implement multiple parallel instances of machine-learned model(s) 65.

[0183] 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.

[0184] 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.

[0185] 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 tocause 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).

[0186] Figure 10 illustrates one example arrangement of computing systems that can be used to implement the present disclosure. Other computing system configurations can be used as well. For example, in some implementations, one or both of computing system 50 or server computing system(s) 60 can implement all or a portion of the operations of model development platform system 70. For example, computing system 50 or server computing system(s) 60 can implement developer tool(s) 75 (or extensions thereof) to develop, update / train, or refine machine-learned models 1, 4, 16, 20, 55, 65, etc. using one or more techniques described herein with respect to model alignment toolkit 17. In this manner, for instance, computing system 50 or server computing system(s) 60 can develop, update / train, or refine machine-learned models based on local datasets (e.g., for model personalization / customization, as permitted by user data preference selections).

[0187] Figure 11 depicts 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 11, 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.

[0188] Figure 12 depicts 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 usercomputing 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).

[0189] The central intelligence layer can include a number of machine-learned models. For example, as illustrated in Figure 12, 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.

[0190] 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 12, the central device data layer can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, or additional components. In some implementations, the central device data layer can communicate with each device component using an API (e.g., a private API).Additional Disclosure

[0191] 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.

[0192] Various data modalities exist. For example, textual data may include plain text, document files like PDFs and DOCX, structured text such as XML, JSON, or HTML, or embedded text such as textual token embeddings. Numerical data can be found in tables, spreadsheets, and databases, while temporal data can include time series, sequential, or timestamped event data. Spatial data is another modality, including GIS data, geospatial, or map data. Image data can be expressed as pixels and / or vectors for example in the form of bitmaps, vectors, JPEGs, or other image formats. Example images include medical images like X-rays or MRIs. A series of images can be structured as a video. Video data can includes movies, streaming data, or camera-captured footage. Audio data can represent speech, music, and / or ambient sounds. Audio data can be represented as a waveform or using imagery known as spectrographs. Graph data can be seen in networks, social, or semantic graphs, while categorical data consists of enumerations, labels, or codes. Biological data can be classified into genomic, proteomic, and metabolomic data. Sensor data can be represented through accelerometers, temperature, or pressure data. Symbolic data can include of logic symbols, mathematical notation, or code. Time series data can include a sequence of data points, typically consisting of successive measurements made over a time interval. The data points in a time series are usually collected at adjacent time intervals and are associated with timestamps. Tabular data can include structured data organized in rows and columns, similar to a spreadsheet or a relational database table. Each column in tabular data typically represents a variable or feature, and each row represents a single observation or record. The intersection of a row and a column holds the value of the corresponding variable for that observation. The input(s) to example model(s) described herein can include data corresponding to any combination of the modalities described above. The outputs(s) generated by example model(s) described herein can include data corresponding to any combination of the modalities described above.

[0193] 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 stillfurther embodiment. Thus, it is intended that the present disclosure cover such alterations, variations, and equivalents.

[0194] 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.”

[0195] 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.

[0196] The term “may” should be understood as referring to a possibility of a feature in various implementations and not as prescribing an ability that is necessarily present in every implementation. For example, the phrase “X may perform Y” should be understood as indicating that, in various implementations, X has the potential to be configured to perform Y, and not as indicating that in every instance X must always be able to perform Y. It should be understood that, in various implementations, X might be unable to perform Y and remain within the scope of the present disclosure.

Claims

WHAT IS CLAIMED IS:

1. A computing system configured to perform processing of multimodal data comprising both time-aligned and time-unaligned data, the computing system comprising: one or more processors; and one or more non-transitory computer-readable media that collectively store: a machine-learned multimodal processing model configured to receive and process a multimodal input data, the machine-learned multimodal processing model comprising: a time-aligned processing model configured to receive and jointly process a first portion of the multimodal input data and a second portion of the multimodal input data to autoregressively generate a sequence of combined feature representations, wherein the first portion of the multimodal input data is expressed in a first modality and the second portion of the multimodal input data expressed in a second, different modality, wherein the first portion of the multimodal input data is time-aligned with the second portion of the multimodal input data; and a time-unaligned processing model configured to receive and process a third portion of the multimodal input data to autoregressively generate a sequence of context feature representations, wherein the third portion of the multimodal input data is time- unaligned with the first portion and the second portion of the multimodal input data; wherein one or both of: the time-aligned processing model receives the sequence of context feature representations as an input; or the time-unaligned processing model receives the sequence of combined feature representations as an input.

2. The computing system of claim 1, wherein the machine-learned multimodal processing model further comprises a prediction model configured to generate a model prediction based on one or both of: the sequence of combined feature representations and the sequence of context feature representations.

3. The computing system of any preceding claim, wherein the time-aligned processing model receives the sequence of context feature representations as an input.

4. The computing system of any preceding claim, wherein the time-unaligned processing model receives the sequence of combined feature representations as an input.

5. The computing system of claim 4, wherein the time-unaligned processing model is configured to perform cross-attention on the sequence of combined feature representations.

6. The computing system of any preceding claim, wherein the time-aligned processing model comprises a first modality encoder, a second modality encoder, a combiner model, and an autoregressive prediction model.

7. The computing system of claim 6, wherein one or both of the first modality encoder and the second modality encoder comprise a vision transformer.

8. The computing system of claim 6 or 7, wherein: for each of a plurality of time chunks of the first portion and the second portion of the multimodal input data: the first modality encoder is configured to encode the first portion of the multimodal input data included in the time chunk to one or more first features; the second modality encoder is configured to encode the second portion of the multimodal input data included in the time chunk to one or more second features; the combiner model is configured to combine the one or more first features and the one or more second features into one or more intermediate combined features; and the autoregressive prediction model is configured to predict a next combined feature representation in the sequence of combined feature representations based on the one or more intermediate combined features generated for the time chunk and one or more prior combined feature representations in the sequence of combined feature representations.

9. The computing system of any of claims 6-8, wherein the combiner model comprises a token turing machine combiner.

10. The computing system of any of claims 6-8, wherein the combiner model comprises a causal transformer combiner.

11. The computing system of any preceding claim, wherein the first input modality comprises imagery, the second input modality comprises audio, and the third input modality comprises text.

12. The computing system of any preceding claim, wherein the machine-learned multimodal processing model has been trained using a latent space reconstruction loss.

13. The computing system of any preceding claim, wherein the machine-learned multimodal processing model has been trained using a video reconstruction loss.

14. The computing system of any preceding claim, wherein the machine-learned multimodal processing model has been trained using an unaligned text cross entropy loss.

15. One or more non -transitory computer-readable media that store the machine- learned multimodal processing model described in any of claims 1-14.

16. A computer-implemented method comprising: obtaining a multimodal input; processing the multimodal input with the machine-learned multimodal processing model described in any of claims 1-14 to generate a model output; and providing the model output as an output.

Images