Posts in category: Google AI

Posted by Piotr Padlewski and Josip Djolonga, Software Engineers, Google Research

Large Language Models (LLMs) like PaLM or GPT-3 showed that scaling transformers to hundreds of billions of parameters improves performance and unlocks emergent abilities. The biggest dense models for image understanding, however, have reached only 4 billion parameters, despite research indicating that promising multimodal models like PaLI continue to benefit from scaling vision models alongside their language counterparts. Motivated by this, and the results from scaling LLMs, we decided to undertake the next step in the journey of scaling the Vision Transformer.

In “Scaling Vision Transformers to 22 Billion Parameters”, we introduce the biggest dense vision model, ViT-22B. It is 5.5x larger than the previous largest vision backbone, ViT-e, which has 4 billion parameters. To enable this scaling, ViT-22B incorporates ideas from scaling text models like PaLM, with improvements to both training stability (using QK normalization) and training efficiency (with a novel approach called asynchronous parallel linear operations). As a result of its modified architecture, efficient sharding recipe, and bespoke implementation, it was able to be trained on Cloud TPUs with a high hardware utilization¹. ViT-22B advances the state of the art on many vision tasks using frozen representations, or with full fine-tuning. Further, the model has also been successfully used in PaLM-e, which showed that a large model combining ViT-22B with a language model can significantly advance the state of the art in robotics tasks.

Architecture

Our work builds on many advances from LLMs, such as PaLM and GPT-3. Compared to the standard Vision Transformer architecture, we use parallel layers, an approach in which attention and MLP blocks are executed in parallel, instead of sequentially as in the standard Transformer. This approach was used in PaLM and reduced training time by 15%.

Secondly, ViT-22B omits biases in the QKV projections, part of the self-attention mechanism, and in the LayerNorms, which increases utilization by 3%. The diagram below shows the modified transformer architecture used in ViT-22B:

ViT-22B transformer encoder architecture uses parallel feed-forward layers, omits biases in QKV and LayerNorm layers and normalizes Query and Key projections.

Models at this scale necessitate “sharding” — distributing the model parameters in different compute devices. Alongside this, we also shard the activations (the intermediate representations of an input). Even something as simple as a matrix multiplication necessitates extra care, as both the input and the matrix itself are distributed across devices. We develop an approach called asynchronous parallel linear operations, whereby communications of activations and weights between devices occur at the same time as computations in the matrix multiply unit (the part of the TPU holding the vast majority of the computational capacity). This asynchronous approach minimizes the time waiting on incoming communication, thus increasing device efficiency. The animation below shows an example computation and communication pattern for a matrix multiplication.

Asynchronized parallel linear operation. The goal is to compute the matrix multiplication y = Ax, but both the matrix A and activation x are distributed across different devices. Here we illustrate how it can be done with overlapping communication and computation across devices. The matrix A is column-sharded across the devices, each holding a contiguous slice, each block represented as Aij. More details are in the paper.

At first, the new model scale resulted in severe training instabilities. The normalization approach of Gilmer et al. (2023, upcoming) resolved these issues, enabling smooth and stable model training; this is illustrated below with example training progressions.

The effect of normalizing the queries and keys (QK normalization) in the self-attention layer on the training dynamics. Without QK normalization (red) gradients become unstable and the training loss diverges.

Results

Here we highlight some results of ViT-22B. Note that in the paper we also explore several other problem domains, like video classification, depth estimation, and semantic segmentation.

To illustrate the richness of the learned representation, we train a text model to produce representations that align text and image representations (using LiT-tuning). Below we show several results for out-of-distribution images generated by Parti and Imagen:

Examples of image+text understanding for ViT-22B paired with a text model. The graph shows normalized probability distribution for each description of an image.

Human object recognition alignment

To find out how aligned ViT-22B classification decisions are with human classification decisions, we evaluated ViT-22B fine-tuned with different resolutions on out-of-distribution (OOD) datasets for which human comparison data is available via the model-vs-human toolbox. This toolbox measures three key metrics: How well do models cope with distortions (accuracy)? How different are human and model accuracies (accuracy difference)? Finally, how similar are human and model error patterns (error consistency)? While not all fine-tuning resolutions perform equally well, ViT-22B variants are state of the art for all three metrics. Furthermore, the ViT-22B models also have the highest ever recorded shape bias in vision models. This means that they mostly use object shape, rather than object texture, to inform classification decisions — a strategy known from human perception (which has a shape bias of 96%). Standard models (e.g., ResNet-50, which has aa ~20–30% shape bias) often classify images like the cat with elephant texture below according to the texture (elephant); models with a high shape bias tend to focus on the shape instead (cat). While there are still many important differences between human and model perception, ViT-22B shows increased similarities to human visual object recognition.

Cat or elephant? Car or clock? Bird or bicycle? Example images with the shape of one object and the texture of a different object, used to measure shape/texture bias.

Shape bias evaluation (higher = more shape-biased). Many vision models have a low shape / high texture bias, whereas ViT-22B fine-tuned on ImageNet (red, green, blue trained on 4B images as indicated by brackets after model names, unless trained on ImageNet only) have the highest shape bias recorded in a ML model to date, bringing them closer to a human-like shape bias.

Out-of-distribution performance

Measuring performance on OOD datasets helps assess generalization. In this experiment we construct label-maps (mappings of labels between datasets) from JFT to ImageNet and also from ImageNet to different out-of-distribution datasets like ObjectNet (results after pre-training on this data shown in the left curve below). Then the models are fully fine-tuned on ImageNet.

We observe that scaling Vision Transformers increases OOD performance: even though ImageNet accuracy saturates, we see a significant increase on ObjectNet from ViT-e to ViT-22B (shown by the three orange dots in the upper right below).

Even though ImageNet accuracy saturates, we see a significant increase in performance on ObjectNet from ViT-e/14 to ViT-22B.

Linear probe

Linear probe is a technique where a single linear layer is trained on top of a frozen model. Compared to full fine-tuning, this is much cheaper to train and easier to set up. We observed that the linear probe of ViT-22B performance approaches that of state-of-the-art full fine-tuning of smaller models using high-resolution images (training with higher resolution is generally much more expensive, but for many tasks it yields better results). Here are results of a linear probe trained on the ImageNet dataset and evaluated on the ImageNet validation dataset and other OOD ImageNet datasets.

Linear probe results trained on ImageNet, evaluated on Imagenet-ReaL, ImageNet-v2, ObjectNet, ImageNet-R and ImageNet-A datasets. High-resolution fine-tuned ViT-e/14 provided as a reference.

Distillation

The knowledge of the bigger model can be transferred to a smaller model using the distillation method. This is helpful as big models are slower and more expensive to use. We found that ViT-22B knowledge can be transferred to smaller models like ViT-B/16 and ViT-L/16, achieving a new state of the art on ImageNet for those model sizes.

Fairness and bias

ML models can be susceptible to unintended unfair biases, such as picking up spurious correlations (measured using demographic parity) or having performance gaps across subgroups. We show that scaling up the size helps in mitigating such issues.

First, scale offers a more favorable tradeoff frontier — performance improves with scale even when the model is post-processed after training to control its level of demographic parity below a prescribed, tolerable level. Importantly, this holds not only when performance is measured in terms of accuracy, but also other metrics, such as calibration, which is a statistical measure of the truthfulness of the model’s estimated probabilities. Second, classification of all subgroups tends to improve with scale as demonstrated below. Third, ViT-22B reduces the performance gap across subgroups.

Top: Accuracy for each subgroup in CelebA before debiasing. Bottom: The y-axis shows the absolute difference in performance across the two specific subgroups highlighted in this example: females and males. ViT-22B has a small gap in performance compared to smaller ViT architectures.

Conclusions

We have presented ViT-22B, currently the largest vision transformer model at 22 billion parameters. With small but critical changes to the original architecture, we achieved excellent hardware utilization and training stability, yielding a model that advances the state of the art on several benchmarks. Great performance can be achieved using the frozen model to produce embeddings and then training thin layers on top. Our evaluations further show that ViT-22B shows increased similarities to human visual perception when it comes to shape and texture bias, and offers benefits in fairness and robustness, when compared to existing models.

Acknowledgements

This is a joint work of Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin Fathy, Elsayed Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetić, Dustin Tran, Thomas Kipf, Mario Lučić, Xiaohua Zhai, Daniel Keysers Jeremiah Harmsen, and Neil Houlsby

We would like to thank Jasper Uijlings, Jeremy Cohen, Arushi Goel, Radu Soricut, Xingyi Zhou, Lluis Castrejon, Adam Paszke, Joelle Barral, Federico Lebron, Blake Hechtman, and Peter Hawkins. Their expertise and unwavering support played a crucial role in the completion of this paper. We also acknowledge the collaboration and dedication of the talented researchers and engineers at Google Research.

¹Note: ViT-22B has 54.9% model FLOPs utilization (MFU) while PaLM reported
46.2% MFU and we measured 44.0% MFU for ViT-e on the same hardware. ^↩

Read More

Posted by Peter Mattson, Senior Staff Engineer, ML Performance, and Praveen Paritosh, Senior Research Scientist, Google Research, Brain Team

Machine learning (ML) offers tremendous potential, from diagnosing cancer to engineering safe self-driving cars to amplifying human productivity. To realize this potential, however, organizations need ML solutions to be reliable with ML solution development that is predictable and tractable. The key to both is a deeper understanding of ML data — how to engineer training datasets that produce high quality models and test datasets that deliver accurate indicators of how close we are to solving the target problem.

The process of creating high quality datasets is complicated and error-prone, from the initial selection and cleaning of raw data, to labeling the data and splitting it into training and test sets. Some experts believe that the majority of the effort in designing an ML system is actually the sourcing and preparing of data. Each step can introduce issues and biases. Even many of the standard datasets we use today have been shown to have mislabeled data that can destabilize established ML benchmarks. Despite the fundamental importance of data to ML, it’s only now beginning to receive the same level of attention that models and learning algorithms have been enjoying for the past decade.

Towards this goal, we are introducing DataPerf, a set of new data-centric ML challenges to advance the state-of-the-art in data selection, preparation, and acquisition technologies, designed and built through a broad collaboration across industry and academia. The initial version of DataPerf consists of four challenges focused on three common data-centric tasks across three application domains; vision, speech and natural language processing (NLP). In this blogpost, we outline dataset development bottlenecks confronting researchers and discuss the role of benchmarks and leaderboards in incentivizing researchers to address these challenges. We invite innovators in academia and industry who seek to measure and validate breakthroughs in data-centric ML to demonstrate the power of their algorithms and techniques to create and improve datasets through these benchmarks.

Data is the new bottleneck for ML

Data is the new code: it is the training data that determines the maximum possible quality of an ML solution. The model only determines the degree to which that maximum quality is realized; in a sense the model is a lossy compiler for the data. Though high-quality training datasets are vital to continued advancement in the field of ML, much of the data on which the field relies today is nearly a decade old (e.g., ImageNet or LibriSpeech) or scraped from the web with very limited filtering of content (e.g., LAION or The Pile).

Despite the importance of data, ML research to date has been dominated by a focus on models. Before modern deep neural networks (DNNs), there were no ML models sufficient to match human behavior for many simple tasks. This starting condition led to a model-centric paradigm in which (1) the training dataset and test dataset were “frozen” artifacts and the goal was to develop a better model, and (2) the test dataset was selected randomly from the same pool of data as the training set for statistical reasons. Unfortunately, freezing the datasets ignored the ability to improve training accuracy and efficiency with better data, and using test sets drawn from the same pool as training data conflated fitting that data well with actually solving the underlying problem.

Because we are now developing and deploying ML solutions for increasingly sophisticated tasks, we need to engineer test sets that fully capture real world problems and training sets that, in combination with advanced models, deliver effective solutions. We need to shift from today’s model-centric paradigm to a data-centric paradigm in which we recognize that for the majority of ML developers, creating high quality training and test data will be a bottleneck.

Shifting from today’s model-centric paradigm to a data-centric paradigm enabled by quality datasets and data-centric algorithms like those measured in DataPerf.

Enabling ML developers to create better training and test datasets will require a deeper understanding of ML data quality and the development of algorithms, tools, and methodologies for optimizing it. We can begin by recognizing common challenges in dataset creation and developing performance metrics for algorithms that address those challenges. For instance:

• Data selection: Often, we have a larger pool of available data than we can label or train on effectively. How do we choose the most important data for training our models?
• Data cleaning: Human labelers sometimes make mistakes. ML developers can’t afford to have experts check and correct all labels. How can we select the most likely-to-be-mislabeled data for correction?

We can also create incentives that reward good dataset engineering. We anticipate that high quality training data, which has been carefully selected and labeled, will become a valuable product in many industries but presently lack a way to assess the relative value of different datasets without actually training on the datasets in question. How do we solve this problem and enable quality-driven “data acquisition”?

DataPerf: The first leaderboard for data

We believe good benchmarks and leaderboards can drive rapid progress in data-centric technology. ML benchmarks in academia have been essential to stimulating progress in the field. Consider the following graph which shows progress on popular ML benchmarks (MNIST, ImageNet, SQuAD, GLUE, Switchboard) over time:

Performance over time for popular benchmarks, normalized with initial performance at minus one and human performance at zero. (Source: Douwe, et al. 2021; used with permission.)

Online leaderboards provide official validation of benchmark results and catalyze communities intent on optimizing those benchmarks. For instance, Kaggle has over 10 million registered users. The MLPerf official benchmark results have helped drive an over 16x improvement in training performance on key benchmarks.

DataPerf is the first community and platform to build leaderboards for data benchmarks, and we hope to have an analogous impact on research and development for data-centric ML. The initial version of DataPerf consists of leaderboards for four challenges focused on three data-centric tasks (data selection, cleaning, and acquisition) across three application domains (vision, speech and NLP):

• Training data selection (Vision): Design a data selection strategy that chooses the best training set from a large candidate pool of weakly labeled training images.
• Training data selection (Speech): Design a data selection strategy that chooses the best training set from a large candidate pool of automatically extracted clips of spoken words.
• Training data cleaning (Vision): Design a data cleaning strategy that chooses samples to relabel from a “noisy” training set where some of the labels are incorrect.
• Training dataset evaluation (NLP): Quality datasets can be expensive to construct, and are becoming valuable commodities. Design a data acquisition strategy that chooses which training dataset to “buy” based on limited information about the data.

For each challenge, the DataPerf website provides design documents that define the problem, test model(s), quality target, rules and guidelines on how to run the code and submit. The live leaderboards are hosted on the Dynabench platform, which also provides an online evaluation framework and submission tracker. Dynabench is an open-source project, hosted by the MLCommons Association, focused on enabling data-centric leaderboards for both training and test data and data-centric algorithms.

How to get involved

We are part of a community of ML researchers, data scientists and engineers who strive to improve data quality. We invite innovators in academia and industry to measure and validate data-centric algorithms and techniques to create and improve datasets through the DataPerf benchmarks. The deadline for the first round of challenges is May 26th, 2023.

Acknowledgements

The DataPerf benchmarks were created over the last year by engineers and scientists from: Coactive.ai, Eidgenössische Technische Hochschule (ETH) Zurich, Google, Harvard University, Meta, ML Commons, Stanford University. In addition, this would not have been possible without the support of DataPerf working group members from Carnegie Mellon University, Digital Prism Advisors, Factored, Hugging Face, Institute for Human and Machine Cognition, Landing.ai, San Diego Supercomputing Center, Thomson Reuters Lab, and TU Eindhoven.

Read More

Katie Couric hosts Future Ready by Think with Google, a new interview series featuring experts in AI, cybersecurity and psychology.Read More

Posted by Harsh Mehta, Software Engineer, and Walid Krichene, Research Scientist, Google Research

Large deep learning models are becoming the workhorse of a variety of critical machine learning (ML) tasks. However, it has been shown that without any protection it is plausible for bad actors to attack a variety of models, across modalities, to reveal information from individual training examples. As such, it’s essential to protect against this sort of information leakage.

Differential privacy (DP) provides formal protection against an attacker who aims to extract information about the training data. The most popular method for DP training in deep learning is differentially private stochastic gradient descent (DP-SGD). The core recipe implements a common theme in DP: “fuzzing” an algorithm’s outputs with noise to obscure the contributions of any individual input.

In practice, DP training can be very expensive or even ineffective for very large models. Not only does the computational cost typically increase when requiring privacy guarantees, but the noise also increases proportionally. Given these challenges, there has recently been much interest in developing methods that enable efficient DP training. The goal is to develop simple and practical methods for producing high-quality large-scale private models.

The ImageNet classification benchmark is an effective test bed for this goal because 1) it is a challenging task even in the non-private setting, that requires sufficiently large models to successfully classify large numbers of varied images and 2) it is a public, open-source dataset, which other researchers can access and use for collaboration. With this approach, researchers may simulate a practical situation where a large model is required to train on private data with DP guarantees.

To that end, today we discuss improvements we’ve made in training high-utility, large-scale private models. First, in “Large-Scale Transfer Learning for Differentially Private Image Classification”, we share strong results on the challenging task of image classification on the ImageNet-1k dataset with DP constraints. We show that with a combination of large-scale transfer learning and carefully chosen hyperparameters it is indeed possible to significantly reduce the gap between private and non-private performance even on challenging tasks and high-dimensional models. Then in “Differentially Private Image Classification from Features”, we further show that privately fine-tuning just the last layer of pre-trained model with more advanced optimization algorithms improves the performance even further, leading to new state-of-the-art DP results across a variety of popular image classification benchmarks, including ImageNet-1k. To encourage further development in this direction and enable other researchers to verify our findings, we are also releasing the associated source code.

Transfer learning and differential privacy

The main idea behind transfer learning is to reuse the knowledge gained from solving one problem and then apply it to a related problem. This is especially useful when there is limited or low-quality data available for the target problem as it allows us to leverage the knowledge gained from a larger and more diverse public dataset.

In the context of DP, transfer learning has emerged as a promising technique to improve the accuracy of private models, by leveraging knowledge learned from pre-training tasks. For example, if a model has already been trained on a large public dataset for a similar privacy-sensitive task, it can be fine-tuned on a smaller and more specific dataset for the target DP task. More specifically, one first pre-trains a model on a large dataset with no privacy concerns, and then privately fine-tunes the model on the sensitive dataset. In our work, we improve the effectiveness of DP transfer learning and illustrate it by simulating private training on publicly available datasets, namely ImageNet-1k, CIFAR-100, and CIFAR-10.

Better pre-training improves DP performance

To start exploring how transfer learning can be effective for differentially private image classification tasks, we carefully examined hyperparameters affecting DP performance. Surprisingly, we found that with carefully chosen hyperparameters (e.g., initializing the last layer to zero and choosing large batch sizes), privately fine-tuning just the last layer of a pre-trained model yields significant improvements over the baseline. Training just the last layer also significantly improves the cost-utility ratio of training a high-quality image classification model with DP.

As shown below, we compare the performance on ImageNet of the best hyperparameter recommendations both with and without privacy and across a variety of model and pre-training dataset sizes. We find that scaling the model and using a larger pre-training dataset decreases the gap in accuracy coming from the addition of the privacy guarantee. Typically, privacy guarantees of a system are characterized by a positive parameter ε, with smaller ε corresponding to better privacy. In the following figure, we use the privacy guarantee of ε = 10.

Comparing our best models with and without privacy on ImageNet across model and pre-training dataset sizes. The X-axis shows the different Vision Transformer models we used for this study in ascending order of model size from left to right. We used JFT-300M to pretrain B/16, L/16 and H/14 models, JFT-4B (a larger version of JFT-3B) to pretrain H/14-4b and JFT-3B to pretrain G/14-3b. We do this in order to study the effectiveness of jointly scaling the model and pre-training dataset (JFT-3B or 4B). The Y-axis shows the Top-1 accuracy on ImageNet-1k test set once the model is finetuned (in the private or non-private way) with the ImageNet-1k training set. We consistently see that the scaling of the model and the pre-training dataset size decreases the gap in accuracy coming from the addition of the privacy guarantee of ε = 10.

Better optimizers improve DP performance

Somewhat surprisingly, we found that privately training just the last layer of a pre-trained model provides the best utility with DP. While past studies [1, 2, 3] largely relied on using first-order differentially private training algorithms like DP-SGD for training large models, in the specific case of privately learning just the last layer from features, we observe that computational burden is often low enough to allow for more sophisticated optimization schemes, including second-order methods (e.g., Newton or Quasi-Newton methods), which can be more accurate but also more computationally expensive.

In “Differentially Private Image Classification from Features”, we systematically explore the effect of loss functions and optimization algorithms. We find that while the commonly used logistic regression performs better than linear regression in the non-private setting, the situation is reversed in the private setting: least-squares linear regression is much more effective than logistic regression from both a privacy and computational standpoint for typical range of ε values ([1, 10]), and even more effective for stricter epsilon values (ε < 1).

We further explore using DP Newton’s method to solve logistic regression. We find that this is still outperformed by DP linear regression in the high privacy regime. Indeed, Newton’s method involves computing a Hessian (a matrix that captures second-order information), and making this matrix differentially private requires adding far more noise in logistic regression than in linear regression, which has a highly structured Hessian.

Building on this observation, we introduce a method that we call differentially private SGD with feature covariance (DP-FC), where we simply replace the Hessian in logistic regression with privatized feature covariance. Since feature covariance only depends on the inputs (and neither on model parameters nor class labels), we are able to share it across classes and training iterations, thus greatly reducing the amount of noise that needs to be added to protect it. This allows us to combine the benefits of using logistic regression with the efficient privacy protection of linear regression, leading to improved privacy-utility trade-off.

With DP-FC, we surpass previous state-of-the-art results considerably on three private image classification benchmarks, namely ImageNet-1k, CIFAR-10 and CIFAR-100, just by performing DP fine-tuning on features extracted from a powerful pre-trained model.

Comparison of top-1 accuracies (Y-axis) with private fine-tuning using DP-FC method on all three datasets across a range of ε (X-axis). We observe that better pre-training helps even more for lower values of ε (stricter privacy guarantee).

Conclusion

We demonstrate that large-scale pre-training on a public dataset is an effective strategy for obtaining good results when fine-tuned privately. Moreover, scaling both model size and pre-training dataset improves performance of the private model and narrows the quality gap compared to the non-private model. We further provide strategies to effectively use transfer learning for DP. Note that this work has several limitations worth considering — most importantly our approach relies on the availability of a large and trustworthy public dataset, which can be challenging to source and vet. We hope that our work is useful for training large models with meaningful privacy guarantees!

Acknowledgements

In addition to the authors of this blogpost, this research was conducted by Abhradeep Thakurta, Alex Kurakin and Ashok Cutkosky. We are also grateful to the developers of Jax, Flax, and Scenic libraries. Specifically, we would like to thank Mostafa Dehghani for helping us with Scenic and high-performance vision baselines and Lucas Beyer for help with deduping the JFT data. We are also grateful to Li Zhang, Emil Praun, Andreas Terzis, Shuang Song, Pierre Tholoniat, Roxana Geambasu, and Steve Chien for stimulating discussions on differential privacy throughout the project. Additionally, we thank anonymous reviewers, Gautam Kamath and Varun Kanade for helpful feedback throughout the publication process. Finally, we would like to thank John Anderson and Corinna Cortes from Google Research, Borja Balle, Soham De, Sam Smith, Leonard Berrada, and Jamie Hayes from DeepMind for generous feedback.

Read More

AI can be hugely beneficial to society if citizens, educators, academics, civil society and governments unite to shape it responsibly.Read More

Posted by Rahul Goel and Aditya Gupta, Software Engineers, Google Assistant

Virtual assistants are increasingly integrated into our daily routines. They can help with everything from setting alarms to giving map directions and can even assist people with disabilities to more easily manage their homes. As we use these assistants, we are also becoming more accustomed to using natural language to accomplish tasks that we once did by hand.

One of the biggest challenges in building a robust virtual assistant is identifying what a user wants and what information is needed to perform the task at hand. In the natural language processing (NLP) literature, this is mainly framed as a task-oriented dialogue parsing task, where a given dialogue needs to be parsed by a system to understand the user intent and carry out the operation to fulfill that intent. While the academic community has made progress in handling task-oriented dialogue thanks to custom purpose datasets, such as MultiWOZ, TOP, SMCalFlow, etc., progress is limited because these datasets lack typical speech phenomena necessary for model training to optimize language model performance. The resulting models often underperform, leading to dissatisfaction with assistant interactions. Relevant speech patterns might include revisions, disfluencies, code-mixing, and the use of structured context surrounding the user’s environment, which might include the user’s notes, smart home devices, contact lists, etc.

Consider the following dialogue that illustrates a common instance when a user needs to revise their utterance:

A dialogue conversation with a virtual assistant that includes a user revision.

The virtual assistant misunderstands the request and attempts to call the incorrect contact. Hence, the user has to revise their utterance to fix the assistant’s mistake. To parse the last utterance correctly, the assistant would also need to interpret the special context of the user — in this case, it would need to know that the user had a contact list saved in their phone that it should reference.

Another common category of utterance that is challenging for virtual assistants is code-mixing, which occurs when the user switches from one language to another while addressing the assistant. Consider the utterance below:

A dialogue denoting code-mixing between English and German.

In this example, the user switches from English to German, where “vier Uhr” means “four o’clock” in German.

In an effort to advance research in parsing such realistic and complex utterances, we are launching a new dataset called PRESTO, a multilingual dataset for parsing realistic task-oriented dialogues that includes roughly half a million realistic conversations between people and virtual assistants. The dataset spans six different languages and includes multiple conversational phenomena that users may encounter when using an assistant, including user-revisions, disfluencies, and code-mixing. The dataset also includes surrounding structured context, such as users’ contacts and lists associated with each example. The explicit tagging of various phenomena in PRESTO allows us to create different test sets to separately analyze model performance on these speech phenomena. We find that some of these phenomena are easier to model with few-shot examples, while others require much more training data.

Dataset characteristics

1. Conversations by native speakers in six languages
   All conversations in our dataset are provided by native speakers of six languages — English, French, German, Hindi, Japanese, and Spanish. This is in contrast to other datasets, such as MTOP and MASSIVE, that translate utterances only from English to other languages, which does not necessarily reflect the speech patterns of native speakers in non-English languages.
2. Structured context
   Users often rely on the information stored in their devices, such as notes, contacts, and lists, when interacting with virtual assistants. However, this context is often not accessible to the assistant, which can result in parsing errors when processing user utterances. To address this issue, PRESTO includes three types of structured context, notes, lists, and contacts, as well as user utterances and their parses. The lists, notes, and contacts are authored by native speakers of each language during data collection. Having such context allows us to examine how this information can be used to improve performance on parsing task-oriented dialog models.
   

   Each example in PRESTO consists of: Inputs — A user’s virtual state (context), one or more user utterances, and the corresponding virtual assistant responses (dialogue). Output — The semantic parsing of the last user utterance in the dialogue (parse).

3. User revisions
   It is common for a user to revise or correct their own utterances while speaking to a virtual assistant. These revisions happen for a variety of reasons — the assistant could have made a mistake in understanding the utterance or the user might have changed their mind while making an utterance. One such example is in the figure above. Other examples of revisions include canceling one’s request (‘’Don’t add anything.”) or correcting oneself in the same utterance (“Add bread — no, no wait — add wheat bread to my shopping list.”). Roughly 27% of all examples in PRESTO have some type of user revision that is explicitly labeled in the dataset.
4. Code-mixing
   As of 2022, roughly 43% of the world’s population is bilingual. As a result, many users switch languages while speaking to virtual assistants. In building PRESTO, we asked bilingual data contributors to annotate code-mixed utterances, which amounted to roughly 14% of all utterances in the dataset.
   

   Examples of Hindi-English, Spanish-English, and German-English code-switched utterances from PRESTO.

5. Disfluencies
   Disfluencies, like repeated phrases or filler words, are ubiquitous in user utterances due to the spoken nature of the conversations that the virtual assistants receive. Datasets such as DISFL-QA note the lack of such phenomena in existing NLP literature and contribute towards the goal of alleviating that gap. In our work, we include conversations targeting this particular phenomenon across all six languages.
   

   Examples of utterances in English, Japanese, and French with filler words or repetitions.

Key findings

We performed targeted experiments to focus on each of the phenomena described above. We ran mT5-based models trained using the PRESTO dataset and evaluated them using an exact match between the predicted parse and the human annotated parse. Below we show the relative performance improvements as we scale the training data on each of the targeted phenomena — user revisions, disfluencies, and code-mixing.

K-shot results on various linguistic phenomena and the full test set across increasing training data size.

The k-shot results yield the following takeaways:

1. Zero-shot performance on the marked phenomenon is poor, emphasizing the need for such utterances in the dataset to improve performance.
2. Disfluencies and code-mixing have a much better zero-shot performance than user-revisions (over 40 points difference in exact-match accuracy).

We also investigate the difference between training monolingual and multilingual models on the train set and find that with fewer data multilingual models have an advantage over monolingual models, but the gap shrinks as the data size is increased.

Additional details on data quality, data collection methodology, and modeling experiments can be found in our paper.

Conclusion

We created PRESTO, a multilingual dataset for parsing task-oriented dialogues that includes realistic conversations representing a variety of pain points that users often face in their daily conversations with virtual assistants that are lacking in existing datasets in the NLP community. PRESTO includes roughly half a million utterances that are contributed by native speakers of six languages — English, French, German, Hindi, Japanese, and Spanish. We created dedicated test sets to focus on each targeted phenomenon — user revisions, disfluencies, code-mixing, and structured context. Our results indicate that the zero-shot performance is poor when the targeted phenomenon is not included in the training set, indicating a need for such utterances to improve performance. We notice that user revisions and disfluencies are easier to model with more data as opposed to code-mixed utterances, which are harder to model, even with a high number of examples. With the release of this dataset, we open more questions than we answer and we hope the research community makes progress on utterances that are more in line with what users are facing every day.

Acknowledgements

It was a privilege to collaborate on this work with Waleed Ammar, Siddharth Vashishtha, Motoki Sano, Faiz Surani, Max Chang, HyunJeong Choe, David Greene, Kyle He, Rattima Nitisaroj, Anna Trukhina, Shachi Paul, Pararth Shah, Rushin Shah, and Zhou Yu. We’d also like to thank Tom Small for the animations in this blog post. Finally, a huge thanks to all the expert linguists and data annotators for making this a reality.

Read More

Posted by Boris Babenko, Software Engineer, and Akib Uddin, Product Manager, Google Research

Last year we presented results demonstrating that a deep learning system (DLS) can be trained to analyze external eye photos and predict a person’s diabetic retinal disease status and elevated glycated hemoglobin (or HbA1c, a biomarker that indicates the three-month average level of blood glucose). It was previously unknown that external eye photos contained signals for these conditions. This exciting finding suggested the potential to reduce the need for specialized equipment since such photos can be captured using smartphones and other consumer devices. Encouraged by these findings, we set out to discover what other biomarkers can be found in this imaging modality.

In “A deep learning model for novel systemic biomarkers in photos of the external eye: a retrospective study”, published in Lancet Digital Health, we show that a number of systemic biomarkers spanning several organ systems (e.g., kidney, blood, liver) can be predicted from external eye photos with an accuracy surpassing that of a baseline logistic regression model that uses only clinicodemographic variables, such as age and years with diabetes. The comparison with a clinicodemographic baseline is useful because risk for some diseases could also be assessed using a simple questionnaire, and we seek to understand if the model interpreting images is doing better. This work is in the early stages, but it has the potential to increase access to disease detection and monitoring through new non-invasive care pathways.

A model generating predictions for an external eye photo.

Model development and evaluation

To develop our model, we worked with partners at EyePACS and the Los Angeles County Department of Health Services to create a retrospective de-identified dataset of external eye photos and measurements in the form of laboratory tests and vital signs (e.g., blood pressure). We filtered down to 31 lab tests and vitals that were more commonly available in this dataset and then trained a multi-task DLS with a classification “head” for each lab and vital to predict abnormalities in these measurements.

Importantly, evaluating the performance of many abnormalities in parallel can be problematic because of a higher chance of finding a spurious and erroneous result (i.e., due to the multiple comparisons problem). To mitigate this, we first evaluated the model on a portion of our development dataset. Then, we narrowed the list down to the nine most promising prediction tasks and evaluated the model on our test datasets while correcting for multiple comparisons. Specifically, these nine tasks, their associated anatomy, and their significance for associated diseases are listed in the table below.

Key results

As in our previous work, we compared our external eye model to a baseline model (a logistic regression model taking clinicodemographic variables as input) by computing the area under the receiver operator curve (AUC). The AUC ranges from 0 to 100%, with 50% indicating random performance and higher values indicating better performance. For all but one of the nine prediction tasks, our model statistically outperformed the baseline model. In terms of absolute performance, the model’s AUCs ranged from 62% to 88%. While these levels of accuracy are likely insufficient for diagnostic applications, it is in line with other initial screening tools, like mammography and pre-screening for diabetes, used to help identify individuals who may benefit from additional testing. And as a non-invasive accessible modality, taking photographs of the external eye may offer the potential to help screen and triage patients for confirmatory blood tests or other clinical follow-up.

Results on the EyePACS test set, showing AUC performance of our DLS compared to a baseline model. The variable “n” refers to the total number of datapoints, and “N” refers to the number of positives. Error bars show 95% confidence intervals computed using the DeLong method. †Indicates that the target was pre-specified as secondary analysis; all others were pre-specified as primary analysis.

The external eye photos used in both this and the prior study were collected using table top cameras that include a head rest for patient stabilization and produce high quality images with good lighting. Since image quality may be worse in other settings, we wanted to explore to what extent the DLS model is robust to quality changes, starting with image resolution. Specifically, we scaled the images in the dataset down to a range of sizes, and measured performance of the DLS when retrained to handle the downsampled images.

Below we show a selection of the results of this experiment (see the paper for more complete results). These results demonstrate that the DLS is fairly robust and, in most cases, outperforms the baseline model even if the images are scaled down to 150×150 pixels. This pixel count is under 0.1 megapixels, much smaller than the typical smartphone camera.

Conclusion and future directions

Our previous research demonstrated the promise of the external eye modality. In this work, we performed a more exhaustive search to identify the possible systemic biomarkers that can be predicted from these photos. Though these results are promising, many steps remain to determine whether technology like this can help patients in the real world. In particular, as we mention above, the imagery in our studies were collected using large tabletop cameras in a setting that controlled factors such as lighting and head positioning. Furthermore, the datasets used in this work consist primarily of patients with diabetes and did not have sufficient representation of a number of important subgroups – more focused data collection for DLS refinement and evaluation on a more general population and across subgroups will be needed before considering clinical use.

We are excited to explore how these models generalize to smartphone imagery given the potential reach and scale that this enables for the technology. To this end, we are continuing to work with our co-authors at partner institutions like Chang Gung Memorial Hospital in Taiwan, Aravind Eye Hospital in India, and EyePACS in the United States to collect datasets of imagery captured on smartphones. Our early results are promising and we look forward to sharing more in the future.

Acknowledgements

This work involved the efforts of a multidisciplinary team of software engineers, researchers, clinicians and cross functional contributors. Key contributors to this project include: Boris Babenko, Ilana Traynis, Christina Chen, Preeti Singh, Akib Uddin, Jorge Cuadros, Lauren P. Daskivich, April Y. Maa, Ramasamy Kim, Eugene Yu-Chuan Kang, Yossi Matias, Greg S. Corrado, Lily Peng, Dale R. Webster, Christopher Semturs, Jonathan Krause, Avinash V Varadarajan, Naama Hammel and Yun Liu. We also thank Dave Steiner, Yuan Liu, and Michael Howell for their feedback on the manuscript; Amit Talreja for reviewing code for the paper; Elvia Figueroa and the Los Angeles County Department of Health Services Teleretinal Diabetic Retinopathy Screening program staff for data collection and program support; Andrea Limon and Nikhil Kookkiri for EyePACS data collection and support; Dr. Charles Demosthenes for extracting the data and Peter Kuzmak for getting images for the VA data. Last but not least, a special thanks to Tom Small for the animation used in this blog post.

Read More

Posted by Oier Mees, PhD Student, University of Freiburg, and Andy Zeng, Research Scientist, Robotics at Google

People are excellent navigators of the physical world, due in part to their remarkable ability to build cognitive maps that form the basis of spatial memory — from localizing landmarks at varying ontological levels (like a book on a shelf in the living room) to determining whether a layout permits navigation from point A to point B. Building robots that are proficient at navigation requires an interconnected understanding of (a) vision and natural language (to associate landmarks or follow instructions), and (b) spatial reasoning (to connect a map representing an environment to the true spatial distribution of objects). While there have been many recent advances in training joint visual-language models on Internet-scale data, figuring out how to best connect them to a spatial representation of the physical world that can be used by robots remains an open research question.

To explore this, we collaborated with researchers at the University of Freiburg and Nuremberg to develop Visual Language Maps (VLMaps), a map representation that directly fuses pre-trained visual-language embeddings into a 3D reconstruction of the environment. VLMaps, which is set to appear at ICRA 2023, is a simple approach that allows robots to (1) index visual landmarks in the map using natural language descriptions, (2) employ Code as Policies to navigate to spatial goals, such as “go in between the sofa and TV” or “move three meters to the right of the chair”, and (3) generate open-vocabulary obstacle maps — allowing multiple robots with different morphologies (mobile manipulators vs. drones, for example) to use the same VLMap for path planning. VLMaps can be used out-of-the-box without additional labeled data or model fine-tuning, and outperforms other zero-shot methods by over 17% on challenging object-goal and spatial-goal navigation tasks in Habitat and Matterport3D. We are also releasing the code used for our experiments along with an interactive simulated robot demo.

Classic 3D maps with a modern multimodal twist

VLMaps combines the geometric structure of classic 3D reconstructions with the expression of modern visual-language models pre-trained on Internet-scale data. As the robot moves around, VLMaps uses a pre-trained visual-language model to compute dense per-pixel embeddings from posed RGB camera views, and integrates them into a large map-sized 3D tensor aligned with an existing 3D reconstruction of the physical world. This representation allows the system to localize landmarks given their natural language descriptions (such as “a book on a shelf in the living room”) by comparing their text embeddings to all locations in the tensor and finding the closest match. Querying these target locations can be used directly as goal coordinates for language-conditioned navigation, as primitive API function calls for Code as Policies to process spatial goals (e.g., code-writing models interpret “in between” as arithmetic between two locations), or to sequence multiple navigation goals for long-horizon instructions.

# move first to the left side of the counter, then move between the sink and the oven, then move back and forth to the sofa and the table twice.
robot.move_to_left('counter')
robot.move_in_between('sink', 'oven')
pos1 = robot.get_pos('sofa')
pos2 = robot.get_pos('table')
for i in range(2):
   robot.move_to(pos1)
   robot.move_to(pos2)
# move 2 meters north of the laptop, then move 3 meters rightward.
robot.move_north('laptop')
robot.face('laptop')
robot.turn(180)
robot.move_forward(2)
robot.turn(90)
robot.move_forward(3)

Results

We evaluate VLMaps on challenging zero-shot object-goal and spatial-goal navigation tasks in Habitat and Matterport3D, without additional training or fine-tuning. The robot is asked to navigate to four subgoals sequentially specified in natural language. We observe that VLMaps significantly outperforms strong baselines (including CoW and LM-Nav) by up to 17% due to its improved visuo-lingual grounding.

A key advantage of VLMaps is its ability to understand spatial goals, such as “go in between the sofa and TV” or “move three meters to the right of the chair”. Experiments for long-horizon spatial-goal navigation show an improvement by up to 29%. To gain more insights into the regions in the map that are activated for different language queries, we visualize the heatmaps for the object type “chair”.

The improved vision and language grounding capabilities of VLMaps, which contains significantly fewer false positives than competing approaches, enable it to navigate zero-shot to landmarks using language descriptions.

Open-vocabulary obstacle maps

A single VLMap of the same environment can also be used to build open-vocabulary obstacle maps for path planning. This is done by taking the union of binary-thresholded detection maps over a list of landmark categories that the robot can or cannot traverse (such as “tables”, “chairs”, “walls”, etc.). This is useful since robots with different morphologies may move around in the same environment differently. For example, “tables” are obstacles for a large mobile robot, but may be traversable for a drone. We observe that using VLMaps to create multiple robot-specific obstacle maps improves navigation efficiency by up to 4% (measured in terms of task success rates weighted by path length) over using a single shared obstacle map for each robot. See the paper for more details.

Experiments with a mobile robot (LoCoBot) and drone in AI2THOR simulated environments. Left: Top-down view of an environment. Middle columns: Agents’ observations during navigation. Right: Obstacle maps generated for different embodiments with corresponding navigation paths.

Conclusion

VLMaps takes an initial step towards grounding pre-trained visual-language information onto spatial map representations that can be used by robots for navigation. Experiments in simulated and real environments show that VLMaps can enable language-using robots to (i) index landmarks (or spatial locations relative to them) given their natural language descriptions, and (ii) generate open-vocabulary obstacle maps for path planning. Extending VLMaps to handle more dynamic environments (e.g., with moving people) is an interesting avenue for future work.

Open-source release

We have released the code needed to reproduce our experiments and an interactive simulated robot demo on the project website, which also contains additional videos and code to benchmark agents in simulation.

Acknowledgments

We would like to thank the co-authors of this research: Chenguang Huang and Wolfram Burgard.

Read More

We’re starting to open access to Bard, an early experiment that lets you collaborate with generative AI. We’re beginning with the U.S. and the U.K., and will expand to m…Read More

Posted by Antoine Yang, Student Researcher, and Arsha Nagrani, Research Scientist, Google Research, Perception team

Videos have become an increasingly important part of our daily lives, spanning fields such as entertainment, education, and communication. Understanding the content of videos, however, is a challenging task as videos often contain multiple events occurring at different time scales. For example, a video of a musher hitching up dogs to a dog sled before they all race away involves a long event (the dogs pulling the sled) and a short event (the dogs being hitched to the sled). One way to spur research in video understanding is via the task of dense video captioning, which consists of temporally localizing and describing all events in a minutes-long video. This differs from single image captioning and standard video captioning, which consists of describing short videos with a single sentence.

Dense video captioning systems have wide applications, such as making videos accessible to people with visual or auditory impairments, automatically generating chapters for videos, or improving the search of video moments in large databases. Current dense video captioning approaches, however, have several limitations — for example, they often contain highly specialized task-specific components, which make it challenging to integrate them into powerful foundation models. Furthermore, they are often trained exclusively on manually annotated datasets, which are very difficult to obtain and hence are not a scalable solution.

In this post, we introduce “Vid2Seq: Large-Scale Pretraining of a Visual Language Model for Dense Video Captioning”, to appear at CVPR 2023. The Vid2Seq architecture augments a language model with special time tokens, allowing it to seamlessly predict event boundaries and textual descriptions in the same output sequence. In order to pre-train this unified model, we leverage unlabeled narrated videos by reformulating sentence boundaries of transcribed speech as pseudo-event boundaries, and using the transcribed speech sentences as pseudo-event captions. The resulting Vid2Seq model pre-trained on millions of narrated videos improves the state of the art on a variety of dense video captioning benchmarks including YouCook2, ViTT and ActivityNet Captions. Vid2Seq also generalizes well to the few-shot dense video captioning setting, the video paragraph captioning task, and the standard video captioning task. Finally, we have also released the code for Vid2Seq here.

Vid2Seq is a visual language model that predicts dense event captions together with their temporal grounding in a video by generating a single sequence of tokens.

A visual language model for dense video captioning

Multimodal transformer architectures have improved the state of the art on a wide range of video tasks, such as action recognition. However it is not straightforward to adapt such an architecture to the complex task of jointly localizing and captioning events in minutes-long videos.

For a general overview of how we achieve this, we augment a visual language model with special time tokens (like text tokens) that represent discretized timestamps in the video, similar to Pix2Seq in the spatial domain. Given visual inputs, the resulting Vid2Seq model can both take as input and generate sequences of text and time tokens. First, this enables the Vid2Seq model to understand the temporal information of the transcribed speech input, which is cast as a single sequence of tokens. Second, this allows Vid2Seq to jointly predict dense event captions and temporally ground them in the video while generating a single sequence of tokens.

The Vid2Seq architecture includes a visual encoder and a text encoder, which encode the video frames and the transcribed speech input, respectively. The resulting encodings are then forwarded to a text decoder, which autoregressively predicts the output sequence of dense event captions together with their temporal localization in the video. The architecture is initialized with a powerful visual backbone and a strong language model.

Vid2Seq model overview: We formulate dense event captioning as a sequence-to-sequence problem, using special time tokens to allow the model to seamlessly understand and generate sequences of tokens containing both textual semantic information and temporal localization information grounding each text sentence in the video.

Large-scale pre-training on untrimmed narrated videos

Due to the dense nature of the task, the manual collection of annotations for dense video captioning is particularly expensive. Hence we pre-train the Vid2Seq model using unlabeled narrated videos, which are easily available at scale. In particular, we use the YT-Temporal-1B dataset, which includes 18 million narrated videos covering a wide range of domains.

We use transcribed speech sentences and their corresponding timestamps as supervision, which are cast as a single sequence of tokens. We pre-train Vid2Seq with a generative objective that teaches the decoder to predict the transcribed speech sequence given visual inputs only, and a denoising objective that encourages multimodal learning by requiring the model to predict masked tokens given a noisy transcribed speech sequence and visual inputs. In particular, noise is added to the speech sequence by randomly masking out spans of tokens.

Vid2Seq is pre-trained on unlabeled narrated videos with a generative objective (top) and a denoising objective (bottom).

Results on downstream dense video captioning benchmarks

The resulting pre-trained Vid2Seq model can be fine-tuned on downstream tasks with a simple maximum likelihood objective using teacher forcing (i.e., predicting the next token given previous ground-truth tokens). After fine-tuning, Vid2Seq notably improves the state of the art on three standard downstream dense video captioning benchmarks (ActivityNet Captions, YouCook2 and ViTT) and two video clip captioning benchmarks (MSR-VTT, MSVD). In our paper we provide additional ablation studies, qualitative results, as well as results in the few-shot settings and in the video paragraph captioning task.

Comparison to state-of-the-art methods for dense video captioning (left) and for video clip captioning (right), on the CIDEr metric (higher is better).

Conclusion

We introduce Vid2Seq, a novel visual language model for dense video captioning that simply predicts all event boundaries and captions as a single sequence of tokens. Vid2Seq can be effectively pretrained on unlabeled narrated videos at scale, and achieves state-of-the-art results on various downstream dense video captioning benchmarks. Learn more from the paper and grab the code here.

Acknowledgements

This research was conducted by Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic and Cordelia Schmid.

Read More