Tag Archives: NeurIPS

Cappy: Outperforming and boosting large multi-task language models with a small scorer

Large language model (LLM) advancements have led to a new paradigm that unifies various natural language processing (NLP) tasks within an instruction-following framework. This paradigm is exemplified by recent multi-task LLMs, such as T0, FLAN, and OPT-IML. First, multi-task data is gathered with each task following a task-specific template, where each labeled example is converted into an instruction (e.g., "Put the concepts together to form a sentence: ski, mountain, skier) paired with a corresponding response (e.g., "Skier skis down the mountain"). These instruction-response pairs are used to train the LLM, resulting in a conditional generation model that takes an instruction as input and generates a response. Moreover, multi-task LLMs have exhibited remarkable task-wise generalization capabilities as they can address unseen tasks by understanding and solving brand-new instructions.

The demonstration of the instruction-following pre-training of multi-task LLMs, e.g., FLAN. Pre-training tasks under this paradigm improves the performance for unseen tasks.

Due to the complexity of understanding and solving various tasks solely using instructions, the size of multi-task LLMs typically spans from several billion parameters to hundreds of billions (e.g., FLAN-11B, T0-11B and OPT-IML-175B). As a result, operating such sizable models poses significant challenges because they demand considerable computational power and impose substantial requirements on the memory capacities of GPUs and TPUs, making their training and inference expensive and inefficient. Extensive storage is required to maintain a unique LLM copy for each downstream task. Moreover, the most powerful multi-task LLMs (e.g., FLAN-PaLM-540B) are closed-sourced, making them impossible to be adapted. However, in practical applications, harnessing a single multi-task LLM to manage all conceivable tasks in a zero-shot manner remains difficult, particularly when dealing with complex tasks, personalized tasks and those that cannot be succinctly defined using instructions. On the other hand, the size of downstream training data is usually insufficient to train a model well without incorporating rich prior knowledge. Hence, it is long desired to adapt LLMs with downstream supervision while bypassing storage, memory, and access issues.

Certain parameter-efficient tuning strategies, including prompt tuning and adapters, substantially diminish storage requirements, but they still perform back-propagation through LLM parameters during the tuning process, thereby keeping their memory demands high. Additionally, some in-context learning techniques circumvent parameter tuning by integrating a limited number of supervised examples into the instruction. However, these techniques are constrained by the model's maximum input length, which permits only a few samples to guide task resolution.

In “Cappy: Outperforming and Boosting Large Multi-Task LMs with a Small Scorer”, presented at NeurIPS 2023, we propose a novel approach that enhances the performance and efficiency of multi-task LLMs. We introduce a lightweight pre-trained scorer, Cappy, based on continual pre-training on top of RoBERTa with merely 360 million parameters. Cappy takes in an instruction and a candidate response as input, and produces a score between 0 and 1, indicating an estimated correctness of the response with respect to the instruction. Cappy functions either independently on classification tasks or serves as an auxiliary component for LLMs, boosting their performance. Moreover, Cappy efficiently enables downstream supervision without requiring any finetuning, which avoids the need for back-propagation through LLM parameters and reduces memory requirements. Finally, adaptation with Cappy doesn’t require access to LLM parameters as it is compatible with closed-source multi-task LLMs, such as those only accessible via WebAPIs.

Cappy takes an instruction and response pair as input and outputs a score ranging from 0 to 1, indicating an estimation of the correctness of the response with respect to the instruction.


Pre-training

We begin with the same dataset collection, which includes 39 diverse datasets from PromptSource that were used to train T0. This collection encompasses a wide range of task types, such as question answering, sentiment analysis, and summarization. Each dataset is associated with one or more templates that convert each instance from the original datasets into an instruction paired with its ground truth response.

Cappy's regression modeling requires each pre-training data instance to include an instruction-response pair along with a correctness annotation for the response, so we produce a dataset with correctness annotations that range from 0 to 1. For every instance within a generation task, we leverage an existing multi-task LLM to generate multiple responses by sampling, conditioned on the given instruction. Subsequently, we assign an annotation to the pair formed by the instruction and every response, using the similarity between the response and the ground truth response of the instance. Specifically, we employ Rouge-L, a commonly-used metric for measuring overall multi-task performance that has demonstrated a strong alignment with human evaluation, to calculate this similarity as a form of weak supervision.

As a result, we obtain an effective regression dataset of 160 million instances paired with correctness score annotations. The final Cappy model is the result of continuous pre-training using the regression dataset on top of the RoBERTa model. The pre-training of Cappy is conducted on Google's TPU-v4, with RedCoast, a lightweight toolkit for automating distributed training.

Data augmentation with a multi-task LLM to construct a weakly supervised regression dataset for Cappy’s pre-training and fine-tuning.


Applying Cappy

Cappy solves practical tasks within a candidate-selection mechanism. More specifically, given an instruction and a set of candidate responses, Cappy produces a score for each candidate response. This is achieved by inputting the instruction alongside each individual response, and then assigning the response with the highest score as its prediction. In classification tasks, all candidate responses are inherently predefined. For example, for an instruction of a sentiment classification task (e.g., “Based on this review, would the user recommend this product?: ‘Stunning even for the non-gamer.’”), the candidate responses are “Yes” or “No”. In such scenarios, Cappy functions independently. On the other hand, in generation tasks, candidate responses are not pre-defined, requiring an existing multi-task LLM to yield the candidate responses. In this case, Cappy serves as an auxiliary component of the multi-task LLM, enhancing its decoding.


Adapting multi-task LLMs with Cappy

When there is available downstream training data, Cappy enables effective and efficient adaptation of multi-task LLMs on downstream tasks. Specifically, we fine-tune Cappy to integrate downstream task information into LLM predictions. This process involves creating a separate regression dataset specific to the downstream training data with the same data annotation process used to construct the pre-training data. As a result, the fine-tuned Cappy collaborates with a multi-task LLM, boosting the LLM's performance on the downstream task.

In contrast to other LLM tuning strategies, adapting LLMs with Cappy significantly reduces the high demand for device memory as it avoids the need for back-propagation through LLM parameters for downstream tasks. Moreover, Cappy adaptation does not rely on the access to LLM parameters, making it compatible with closed-source multi-task LLMs, such as the ones only accessible via WebAPIs. Compared with in-context learning approaches, which circumvent model tuning by attaching training examples to the instruction prefix, Cappy is not restricted by the LLM's maximum input length. Thus, Cappy can incorporate an unlimited number of downstream training examples. Cappy can also be applied with other adaptation methods, such as fine-tuning and in-context learning, further boosting their overall performance.

Downstream adaptation comparison between Cappy and approaches that rely on an LLM’s parameters, such as fine-tuning and prompt tuning. Cappy’s application enhances multi-task LLMs.


Results

We assess Cappy’s performance across eleven held-out language understanding classification tasks from PromptSource. We demonstrate that Cappy, with 360M parameters, outperforms OPT-175B and OPT-IML-30B, and matches the accuracy of the best existing multi-task LLMs (T0-11B and OPT-IML-175B). These findings highlight Cappy’s capabilities and parameter efficiency, which can be credited to its scoring-based pre-training strategy that integrates contrastive information by differentiating between high-quality and low-quality responses. On the contrary, previous multi-task LLMs depend exclusively on teacher-forcing training that utilizes only the ground truth responses.

The overall accuracy averaged over eleven test tasks from PromptSource. “RM” refers to a pre-trained RLHF reward model. Cappy matches the best ones among existing multi-task LLMs.

We also examine the adaptation of multi-task LLMs with Cappy on complex tasks from BIG-Bench, a set of manually curated tasks that are considered beyond the capability of many LLMs. We focus on all the 45 generation BIG-Bench tasks, specifically those that do not offer pre-established answer choices. We evaluate the performance using the Rouge-L score (representing the overall similarity between model generations and corresponding ground truths) on every test set, reporting the average score across 45 tests. In this experiment, all variants of FLAN-T5 serve as the backbone LLMs, and the foundational FLAN-T5 models are frozen. These results, shown below, suggest that Cappy enhances the performance of FLAN-T5 models by a large margin, consistently outperforming the most effective baseline achieved through sample selection using self-scoring of the LLM itself.

The averaged Rouge-L score over 45 complex tasks within BIG-Bench. The x-axis refers to FLAN-T5 models of different sizes. Every dashed line represents an approach working on FLAN-T5s. Self-scoring refers to using the cross-entropy of LLM to select responses. Cappy enhances the performance of FLAN-T5 models by a large margin.


Conclusion

We introduce Cappy, a novel approach that enhances the performance and efficiency of multi-task LLMs. In our experiments, we adapt a single LLM to several domains with Cappy. In the future, Cappy as a pre-trained model can potentially be used in other creative ways beyond on single LLMs.


Acknowledgments

Thanks to Bowen Tan, Jindong Chen, Lei Meng, Abhanshu Sharma and Ewa Dominowska for their valuable feedback. We would also like to thank Eric Xing and Zhiting Hu for their suggestions.

Source: Google AI Blog


StyleDrop: Text-to-image generation in any style

Text-to-image models trained on large volumes of image-text pairs have enabled the creation of rich and diverse images encompassing many genres and themes. Moreover, popular styles such as “anime” or “steampunk”, when added to the input text prompt, may translate to specific visual outputs. While many efforts have been put into prompt engineering, a wide range of styles are simply hard to describe in text form due to the nuances of color schemes, illumination, and other characteristics. As an example, “watercolor painting” may refer to various styles, and using a text prompt that simply says “watercolor painting style” may either result in one specific style or an unpredictable mix of several.

When we refer to "watercolor painting style," which do we mean? Instead of specifying the style in natural language, StyleDrop allows the generation of images that are consistent in style by referring to a style reference image*.

In this blog we introduce “StyleDrop: Text-to-Image Generation in Any Style”, a tool that allows a significantly higher level of stylized text-to-image synthesis. Instead of seeking text prompts to describe the style, StyleDrop uses one or more style reference images that describe the style for text-to-image generation. By doing so, StyleDrop enables the generation of images in a style consistent with the reference, while effectively circumventing the burden of text prompt engineering. This is done by efficiently fine-tuning the pre-trained text-to-image generation models via adapter tuning on a few style reference images. Moreover, by iteratively fine-tuning the StyleDrop on a set of images it generated, it achieves the style-consistent image generation from text prompts.


Method overview

StyleDrop is a text-to-image generation model that allows generation of images whose visual styles are consistent with the user-provided style reference images. This is achieved by a couple of iterations of parameter-efficient fine-tuning of pre-trained text-to-image generation models. Specifically, we build StyleDrop on Muse, a text-to-image generative vision transformer.


Muse: text-to-image generative vision transformer

Muse is a state-of-the-art text-to-image generation model based on the masked generative image transformer (MaskGIT). Unlike diffusion models, such as Imagen or Stable Diffusion, Muse represents an image as a sequence of discrete tokens and models their distribution using a transformer architecture. Compared to diffusion models, Muse is known to be faster while achieving competitive generation quality.


Parameter-efficient adapter tuning

StyleDrop is built by fine-tuning the pre-trained Muse model on a few style reference images and their corresponding text prompts. There have been many works on parameter-efficient fine-tuning of transformers, including prompt tuning and Low-Rank Adaptation (LoRA) of large language models. Among those, we opt for adapter tuning, which is shown to be effective at fine-tuning a large transformer network for language and image generation tasks in a parameter-efficient manner. For example, it introduces less than one million trainable parameters to fine-tune a Muse model of 3B parameters, and it requires only 1000 training steps to converge.

Parameter-efficient adapter tuning of Muse.

Iterative training with feedback

While StyleDrop is effective at learning styles from a few style reference images, it is still challenging to learn from a single style reference image. This is because the model may not effectively disentangle the content (i.e., what is in the image) and the style (i.e., how it is being presented), leading to reduced text controllability in generation. For example, as shown below in Step 1 and 2, a generated image of a chihuahua from StyleDrop trained from a single style reference image shows a leakage of content (i.e., the house) from the style reference image. Furthermore, a generated image of a temple looks too similar to the house in the reference image (concept collapse).

We address this issue by training a new StyleDrop model on a subset of synthetic images, chosen by the user or by image-text alignment models (e.g., CLIP), whose images are generated by the first round of the StyleDrop model trained on a single image. By training on multiple synthetic image-text aligned images, the model can easily disentangle the style from the content, thus achieving improved image-text alignment.

Iterative training with feedback*. The first round of StyleDrop may result in reduced text controllability, such as a content leakage or concept collapse, due to the difficulty of content-style disentanglement. Iterative training using synthetic images, generated by the previous rounds of StyleDrop models and chosen by human or image-text alignment models, improves the text adherence of stylized text-to-image generation.

Experiments


StyleDrop gallery

We show the effectiveness of StyleDrop by running experiments on 24 distinct style reference images. As shown below, the images generated by StyleDrop are highly consistent in style with each other and with the style reference image, while depicting various contexts, such as a baby penguin, banana, piano, etc. Moreover, the model can render alphabet images with a consistent style.

Stylized text-to-image generation. Style reference images* are on the left inside the yellow box. Text prompts used are:
First row: a baby penguin, a banana, a bench.
Second row: a butterfly, an F1 race car, a Christmas tree.
Third row: a coffee maker, a hat, a moose.
Fourth row: a robot, a towel, a wood cabin.
Stylized visual character generation. Style reference images* are on the left inside the yellow box. Text prompts used are: (first row) letter 'A', letter 'B', letter 'C', (second row) letter 'E', letter 'F', letter 'G'.

Generating images of my object in my style

Below we show generated images by sampling from two personalized generation distributions, one for an object and another for the style.

Images at the top in the blue border are object reference images from the DreamBooth dataset (teapot, vase, dog and cat), and the image on the left at the bottom in the red border is the style reference image*. Images in the purple border (i.e. the four lower right images) are generated from the style image of the specific object.

Quantitative results

For the quantitative evaluation, we synthesize images from a subset of Parti prompts and measure the image-to-image CLIP score for style consistency and image-to-text CLIP score for text consistency. We study non–fine-tuned models of Muse and Imagen. Among fine-tuned models, we make a comparison to DreamBooth on Imagen, state-of-the-art personalized text-to-image method for subjects. We show two versions of StyleDrop, one trained from a single style reference image, and another, “StyleDrop (HF)”, that is trained iteratively using synthetic images with human feedback as described above. As shown below, StyleDrop (HF) shows significantly improved style consistency score over its non–fine-tuned counterpart (0.694 vs. 0.556), as well as DreamBooth on Imagen (0.694 vs. 0.644). We observe an improved text consistency score with StyleDrop (HF) over StyleDrop (0.322 vs. 0.313). In addition, in a human preference study between DreamBooth on Imagen and StyleDrop on Muse, we found that 86% of the human raters preferred StyleDrop on Muse over DreamBooth on Imagen in terms of consistency to the style reference image.


Conclusion

StyleDrop achieves style consistency at text-to-image generation using a few style reference images. Google’s AI Principles guided our development of Style Drop, and we urge the responsible use of the technology. StyleDrop was adapted to create a custom style model in Vertex AI, and we believe it could be a helpful tool for art directors and graphic designers — who might want to brainstorm or prototype visual assets in their own styles, to improve their productivity and boost their creativity — or businesses that want to generate new media assets that reflect a particular brand. As with other generative AI capabilities, we recommend that practitioners ensure they align with copyrights of any media assets they use. More results are found on our project website and YouTube video.


Acknowledgements

This research was conducted by Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, Yuan Hao, Irfan Essa, Michael Rubinstein, and Dilip Krishnan. We thank owners of images used in our experiments (links for attribution) for sharing their valuable assets.


*See image sources 

Source: Google AI Blog


Accelerating Text Generation with Confident Adaptive Language Modeling (CALM)

Language models (LMs) are the driving force behind many recent breakthroughs in natural language processing. Models like T5, LaMDA, GPT-3, and PaLM have demonstrated impressive performance on various language tasks. While multiple factors can contribute to improving the performance of LMs, some recent studies suggest that scaling up the model’s size is crucial for revealing emergent capabilities. In other words, some instances can be solved by small models, while others seem to benefit from increased scale.

Despite recent efforts that enabled the efficient training of LMs over large amounts of data, trained models can still be slow and costly for practical use. When generating text at inference time, most autoregressive LMs output content similar to how we speak and write (word after word), predicting each new word based on the preceding words. This process cannot be parallelized since LMs need to complete the prediction of one word before starting to compute the next one. Moreover, predicting each word requires significant computation given the model’s billions of parameters.

In “Confident Adaptive Language Modeling”, presented at NeurIPS 2022, we introduce a new method for accelerating the text generation of LMs by improving efficiency at inference time. Our method, named CALM, is motivated by the intuition that some next word predictions are easier than others. When writing a sentence, some continuations are trivial, while others might require more effort. Current LMs devote the same amount of compute power for all predictions. Instead, CALM dynamically distributes the computational effort across generation timesteps. By selectively allocating more computational resources only to harder predictions, CALM generates text faster while preserving output quality.


Confident Adaptive Language Modeling

When possible, CALM skips some compute effort for certain predictions. To demonstrate this, we use the popular encoder-decoder T5 architecture. The encoder reads the input text (e.g., a news article to summarize) and converts the text to dense representations. Then, the decoder outputs the summary by predicting it word by word. Both the encoder and decoder include a long sequence of Transformer layers. Each layer includes attention and feedforward modules with many matrix multiplications. These layers gradually modify the hidden representation that is ultimately used for predicting the next word.

Instead of waiting for all decoder layers to complete, CALM attempts to predict the next word earlier, after some intermediate layer. To decide whether to commit to a certain prediction or to postpone the prediction to a later layer, we measure the model’s confidence in its intermediate prediction. The rest of the computation is skipped only when the model is confident enough that the prediction won’t change. For quantifying what is “confident enough”, we calibrate a threshold that statistically satisfies arbitrary quality guarantees over the full output sequence.

Text generation with a regular language model (top) and with CALM (bottom). CALM attempts to make early predictions. Once confident enough (darker blue tones), it skips ahead and saves time.

Language Models with Early Exits

Enabling this early exit strategy for LMs requires minimal modifications to the training and inference processes. During training, we encourage the model to produce meaningful representations in intermediate layers. Instead of predicting only using the top layer, our learning loss function is a weighted average over the predictions of all layers, assigning higher weight to top layers. Our experiments demonstrate that this significantly improves the intermediate layer predictions while preserving the full model’s performance. In one model variant, we also include a small early-exit classifier trained to classify if the local intermediate layer prediction is consistent with the top layer. We train this classifier in a second quick step where we freeze the rest of the model.

Once the model is trained, we need a method to allow early-exiting. First, we define a local confidence measure for capturing the model’s confidence in its intermediate prediction. We explore three confidence measures (described in the results section below): (1) softmax response, taking the maximum predicted probability out of the softmax distribution; (2) state propagation, the cosine distance between the current hidden representation and the one from the previous layer; and (3) early-exit classifier, the output of a classifier specifically trained for predicting local consistency. We find the softmax response to be statistically strong while being simple and fast to compute. The other two alternatives are lighter in floating point operations (FLOPS).

Another challenge is that the self-attention of each layer depends on hidden-states from previous words. If we exit early for some word predictions, these hidden-states might be missing. Instead, we attend back to the hidden state of the last computed layer.

Finally, we set up the local confidence threshold for exiting early. In the next section, we describe our controlled process for finding good threshold values. As a first step, we simplify this infinite search space by building on a useful observation: mistakes that are made at the beginning of the generation process are more detrimental since they can affect all of the following outputs. Therefore, we start with a higher (more conservative) threshold, and gradually reduce it with time. We use a negative exponent with user-defined temperature to control this decay rate. We find this allows better control over the performance-efficiency tradeoff (the obtained speedup per quality level).


Reliably Controlling the Quality of the Accelerated Model

Early exit decisions have to be local; they need to happen when predicting each word. In practice, however, the final output should be globally consistent or comparable to the original model. For example, if the original full model generated “the concert was wonderful and long”, one would accept CALM switching the order of the adjectives and outputting “the concert was long and wonderful”. However, at the local level, the word “wonderful” was replaced with “long”. Therefore, the two outputs are globally consistent, but include some local inconsistencies. We build on the Learn then Test (LTT) framework to connect local confidence-based decisions to globally consistent outputs.

In CALM, local per-timestep confidence thresholds for early exiting decisions are derived, via LTT calibration, from user-defined consistency constraints over the full output text. Red boxes indicate that CALM used most of the decoder’s layers for that specific prediction. Green boxes indicate that CALM saved time by using only a few Transformer layers. Full sentence shown in the last example of this post.

First, we define and formulate two types of consistency constraints from which to choose:

  1. Textual consistency: We bound the expected textual distance between the outputs of CALM and the outputs of the full model. This doesn’t require any labeled data.
  2. Risk consistency: We bound the expected increase in loss that we allow for CALM compared to the full model. This requires reference outputs against which to compare.

For each of these constraints, we can set the tolerance that we allow and calibrate the confidence threshold to allow early exits while reliably satisfying our defined constraint with an arbitrarily high probability.


CALM Saves Inference Time

We run experiments on three popular generation datasets: CNN/DM for summarization, WMT for machine translation, and SQuAD for question answering. We evaluate each of the three confidence measures (softmax response, state propagation and early-exit classifier) using an 8-layer encoder-decoder model. To evaluate global sequence-level performance, we use the standard Rouge-L, BLEU, and Token-F1 scores that measure distances against human-written references. We show that one can maintain full model performance while using only a third or half of the layers on average. CALM achieves this by dynamically distributing the compute effort across the prediction timesteps.

As an approximate upper bound, we also compute the predictions using a local oracle confidence measure, which enables exiting at the first layer that leads to the same prediction as the top one. On all three tasks, the oracle measure can preserve full model performance when using only 1.5 decoder layers on average. In contrast to CALM, a static baseline uses the same number of layers for all predictions, requiring 3 to 7 layers (depending on the dataset) to preserve its performance. This demonstrates why the dynamic allocation of compute effort is important. Only a small fraction of the predictions require most of the model’s complexity, while for others much less should suffice.

Performance per task against the average number of decoder layers used.

Finally, we also find that CALM enables practical speedups. When benchmarking on TPUs, we saved almost half of the compute time while maintaining the quality of the outputs.

Example of a generated news summary. The top cell presents the reference human-written summary. Below is the prediction of the full model (8 layers) followed by two different CALM output examples. The first CALM output is 2.9x faster and the second output is 3.6x faster than the full model, benchmarked on TPUs.

Conclusion

CALM allows faster text generation with LMs, without reducing the quality of the output text. This is achieved by dynamically modifying the amount of compute per generation timestep, allowing the model to exit the computational sequence early when confident enough.

As language models continue to grow in size, studying how to efficiently use them becomes crucial. CALM is orthogonal and can be combined with many efficiency related efforts, including model quantization, distillation, sparsity, effective partitioning, and distributed control flows.


Acknowledgements

It was an honor and privilege to work on this with Adam Fisch, Ionel Gog, Seungyeon Kim, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Q. Tran, Yi Tay, and Donald Metzler. We also thank Anselm Levskaya, Hyung Won Chung, Tao Wang, Paul Barham, Michael Isard, Orhan Firat, Carlos Riquelme, Aditya Menon, Zhifeng Chen, Sanjiv Kumar, and Jeff Dean for helpful discussions and feedback. Finally, we thank Tom Small for preparing the animation in this blog post.

Source: Google AI Blog


Google at NeurIPS 2022

This week marks the beginning of the 36th annual Conference on Neural Information Processing Systems (NeurIPS 2022), the biggest machine learning conference of the year, which is being held in New Orleans, LA. NeurIPS 2022 will be held in person with additional options for virtual attendees, and includes invited talks, demonstrations and presentations of some of the latest in machine learning research. This year, NeurIPS is also offering a new track, called Spotlight Papers, which will provide opportunities to highlight papers presented in prestigious journals that would otherwise not have been eligible for submission.

Google is proud to be a Diamond level sponsor of NeurIPS this year and will have a significant presence year with more than 175 accepted papers, additionally contributing to and learning from the broader academic research community through numerous talks, posters, workshops, and tutorials. You can learn more about our work being presented in the list below (Google affiliations highlighted in bold).


Organizing Committee

General Chairs includes: Sanmi Koyejo

Program Chairs include: Alekh Agarwal

Workshop Chairs include: Hanie Sedghi

Tutorial Chairs include: Adji Bousso Dieng, Jessica Schrouff

Affinity Workshop Chair: Adji Bousso Dieng, Jessica Schrouff

Program Committee, Senior Area Chairs include: Corinna Cortes, Claudio Gentile, Mohammad Ghavamzadeh, Amir Globerson, Elad Hazan, Katherine Heller, Satyen Kale, Been Kim, Sanjiv Kumar, Hugo Larochelle, Sergey Levine, Yishay Mansour, Mehryar Mohri, Tara Sainath, Dale Schuurmans, Daniel Tarlow

NeurIPS Foundation Board Secretary: Michael Mozer

NeurIPS Foundation Board Members include: Corinna Cortes, Isabelle Guyon, Sanmi Koyejo, Hugo Larochelle

NeurIPS Foundation Advisory Board include: Peter Bartlett, Zoubin Ghahramani, John C. Platt, Fernando Pereira, Dale Schuurmans


Keynote Speakers

The Data-Centric Era: How ML is Becoming an Experimental Science
Isabelle Guyon

The Forward-Forward Algorithm for Training Deep Neural Networks
Geoffrey Hinton


Outstanding Paper Award

Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding
Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, Mohammad Norouzi


EXPO Day Workshops

Graph Neural Networks in Tensorflow: A Practical Guide
Workshop Organizers include: Bryan Perozzi, Sami Abu-el-Haija

A Hands-On Introduction to Tensorflow and Jax
Workshop Organizers include: Josh Gordon


Affinity Workshops

LatinX in AI (LXAI)
Platinum Sponsor
Networking & Social Chairs include: Andres Muñoz Medina
Program Committee includes: Johan Obando Ceron

Queer in AI
Panelists include: Sara Beery, Talia Ringer

Women in Machine Learning (WiML)
Platinum Sponsor
Workshop Organizers and Mentorship Chairs include: Beliz Gunel
Mentors include: Adam Roberts, Eleni Triantafillou, Zelda Mariet, Clara Hu, Rosanne Liu, Alekh Agarwal, Vinod Prabhakaran, Rose Yu, Katherine Heller


Workshops

New in ML
Workshop Organizers include: Isabelle Guyon

AI for Accelerated Materials Design (AI4Mat)
Workshop Organizers include: Benjamin Sanchez-Lengeling

All Things Attention: Bridging Different Perspectives on Attention
Invited Speakers and Panelists include: Vidhya Navalpakkam

Efficient Natural Language and Speech Processing (ENLSP-II): The Future of Pre-trained Models
Invited Speakers include: Tara Sainath, Anna Huang
Invited Panelists include: Mohammad Norouzi
Program Committee includes: Wenhu Chen

Federated Learning: Recent Advances and New Challenges
Program Committee includes: Kallista Bonawitz, Zachary Charles, Wenshuo Guo, Peter Kairouz, Zhaozhuo Xu, Zheng Xu

Gaussian Processes, Spatiotemporal Modeling, and Decision-Making Systems
Workshop Organizers include: Zi Wang
Invited Speakers include: Jasper Snoek, Carolina Osorio
Advisory Board includes: Zoubin Ghahramani

Has it Trained Yet? A Workshop for Algorithmic Efficiency in Practical Neural Network Training
Workshop Organizers include: Zachary Nado, George Dahl, Naman Agarwal, Aakanksha Chowdhery
Invited Speakers include: Aakanksha Chowdhery, Priya Goyal

Human in the Loop Learning (HiLL)
Workshop Organizers include: Fisher Yu, Vittorio Ferrari
Invited Speakers include: Dorsa Singh, Igor Mordatch, Ding Zhao

INTERPOLATE — First Workshop on Interpolation Regularizers and Beyond
Workshop Organizers include: Yann Dauphin
Invited Speakers include: Chelsea Finn
Panelists include: Chelsea Finn, Dustin Tran
Program Committee includes: Wang Chen, Kimin Lee

LaReL: Language and Reinforcement Learning
Invited Speakers include: Dorsa Singh, Igor Mordatch

Medical Imaging Meets NeurIPS
Program Committee includes: Chenyu You

Memory in Artificial and Real Intelligence (MemARI)
Program Committee includes: Benjamin Eysenbach, Otilia Stretcu

Meta-Learning
Workshop Organizers include: Eleni Triantafillou
Invited Speakers include: Lucas Byer, Chelsea Finn
Program Committee includes: Ishita Dasgupta, Praneet Dutta, Benjamin Eysenbach, Maximilian Igl, Louis Kirsch, Parsa Mahmoudieh, Marc Pickett, Eleni Triantafillou

New Frontiers in Graph Learning (GLFrontiers)
Workshop Organizers include: Hanjun Dai

Offline Reinforcement Learning Workshop: Offline RL as a "Launchpad"
Workshop Organizers include: Rishabh Agarwal, Aviral Kumar, George Tucker
Invited Speakers include: Dorsa Sadigh

Score-Based Methods
Invited Speakers include: Mohammad Norouzi
Invited Panelists include: Jascha Sohl-Dickstein

Synthetic Data for Empowering ML Research
Invited Speakers include: Mehryar Mohri
Invited Panelists include: Katrina Ligett
Program Committee includes: Jinsung Yoon

Table Representation Learning
Workshop Organizers include: Pengcheng Yin
Invited Speakers include: Xinyun Chen, Carsten Binnig
Panelists include: Julian Eisenschlos
Program Committee includes: Wenhu Chen, Xinyun Chen, Beliz Gunel

A Causal View on Dynamical Systems
Program Committee includes: Rose Yu

Algorithmic Fairness Through the Lens of Causality and Privacy
Workshop Organizers include: Awa Dieng
Invited Speakers include: Nicolas Papernot
Roundtable Leads include: David Madras, Negar Rostamzadeh, Nyalleng Moroosi
Program Committee includes: Matt Kusner

Broadening Research Collaborations in ML
Workshop Organizers include: Rosanne Liu, Pablo Samuel Castro, Sunipa Dev

Decentralization and Trustworthy Machine Learning in Web3: Methodologies, Platforms, and Applications
Invited Speakers include: Peter Kairouz

Distribution Shifts (DistShift): Connecting Methods and Applications
Workshop Organizers include: Becca Roelofs, Chelsea Finn, Jacob Eisenstein, Pang Wei Koh
Invited Speakers include: Sarah Beery

Foundation Models for Decision Making
Workshop Organizers include: Sherry Yang, Yilun Du, Igor Mordatch, Shixiang Shane Gu,Ofir Nachum
Invited Speakers include: Dorsa Sadigh, Dale Schuurmans, Machel Reid
Program Committee includes: Bo Dai, Aleksandra Faust, Hiroki Furuta, Kati Goshvadi, Izzeddin Gur, Austin Huang, Kimin Lee, Kuang-Huei Lee, Lisa Lee, Yingjie Miao, Jordi Orbay, Ted Xiao

Gaze Meets ML
Program Committee includes: Peter Mattson, Mehdi Moradi

I Can’t Believe It’s Not Better: Understanding Deep Learning Through Empirical Falsification
Workshop Organizers include: Javier Antorán
Panelists include: Kevin Murphy

Interactive Learning for Natural Language Processing
Invited Speakers include: Anca Dragan
Program Committees include: Julia Kreutzer, Shunyu Yao

Machine Learning and the Physical Sciences
Workshop Organizers include: Adji Bousso Dieng
Invited Speakers include: Ekin Doğuş Çubuk

Machine Learning for Systems
Workshop Organizers include: Martin Maas, Azade Nova, Dan Zhang
Invited Speakers include: Jeff Dean
Program Committee includes: Milad Hashemi, Kevin Swersky

Machine Learning in Structural Biology
Invited Speakers include: David Fleet

MATH-AI: Toward Human-Level Mathematical Reasoning
Workshop Organizers include: Swaroop Mishra, Yuhuai Wu
Invited Speakers include: Talia Ringer

OPT 2022: Optimization for Machine Learning
Workshop Organizers include: Courtney Paquette

Reinforcement Learning for Real Life (RL4RealLife)
Workshop Organizers include: Minmin Chen
Invited Panelists include: Pablo Samuel Castro
Program Committee includes: Victor Carbune, Bo Chang, Yinlam Chow, Konstantina Christakopoulou, Bo Dai, Hanjun Dai, Aleksandra Faust, Joshua Greaves‎, Chih-wei Hsu, Rahul Kidambi, Srivatsan Krishnan, Iou-Jen Liu, Cong Lu, Jincheng Mei, Chao Qin

Self-Supervised Learning - Theory and Practice
Invited Speakers include: Mathilde Caron

Symmetry and Geometry in Neural Representations (NeurReps)
Invited Speakers include: Noah Shutty
Program Committee includes: Ondrej Biza, Noah Shutty

Temporal Graph Learning Workshop
Invited Speakers include: Mehran Kazemi

Transfer Learning for Natural Language Processing
Workshop Organizers include: Deepak Ramachandran, Sebastian Ruder
Invited Speakers include: Jonas Pfeiffer
Invited Debaters include: Ellie Pavlick
Program Committee includes: Patrick Fernandes, Jonas Pfeiffer, Jiao Sun, Tu Vu, Xinyi Wang, Xin Xu

Cultures of AI and AI for Culture
Workshop Organizers include: Rida Qadri, Fernando Diaz

Deep Reinforcement Learning Workshop
Workshop Organizers include: Karol Hausman, Ted Xiao, Zeyu Zheng
Invited Speakers include: Igor Mordatch
Advisory Board includes: Chelsea Finn

Empowering Communities: A Participatory Approach to AI for Mental Health
Program Committee includes: Diana Mincu, Subhrajit Roy, Martin Seneviratne

HCAI@NeurIPS 2022, Human Centered AI
Keynote Speaker includes: Fernanda Viegas

Learning Meaningful Representations of Life
Workshop Organizers include: Adji Bousso Dieng

Machine Learning for Creativity and Design
Workshop Organizers include: Yingtao Tian

Machine Learning Safety
Workshop Organizers include: Nicholas Carlini
Invited Speakers include: Dorsa Sadigh

Neuro Causal and Symbolic AI (nCSI)
Workshop Organizers include: Thomas Kipf

Robot Learning Workshop: Trustworthy Robotics
Workshop Organizers include: Alex Bewley, Jonathan Tompson
Invited Speakers include: Karol Hausman, Brian Ichter, Been Kim, Leila Takayama, Andy Zeng
Program Committee includes: Vincent Vanhoucke

The Symbiosis of Deep Learning and Differential Equations II
Workshop Organizers include: Winnie Xu
Invited Speakers include: Rose Yu

Tackling Climate Change with Machine Learning
Workshop Organizers include: Emma Strubell

Trustworthy and Socially Responsible Machine Learning
Invited Speakers include: Been Kim, Dorsa Sadigh, Milind Tambe

Vision Transformers: Theory and Applications
Invited Speakers include: Cordelia Schmid, Ming Hsuan Yang


Tutorials

Advances in Bayesian Optimization
Tutorial Organizers include: Virginia Aglietti

Creative Culture and Machine Learning
Tutorial Organizers include: Negar Rostamzadeh

Fair and Socially Responsible ML for Recommendations: Challenges and Perspectives
Invited Panelists include: Fernando Diaz

Lifelong Learning Machines
Invited Panelists include: Christopher Summerfield

The Role of Meta-learning for Few-Shot Learning
Tutorial Organizers include: Eleni Triantafillou
Invited Panelists include: Neil Houlsby, Priyanka Agrawal


Competitions

NeurIPS 2022 Competition Track: Overview & Results
Invited Speakers include: Isabelle Guyon

Causal Insights for Learning Paths in Education
Competition Organizers include: Zichao (Jack) Wang

IGLU: Interactive Grounded Language Understanding in a Collaborative Environment
Competition Organizers include: Negar Arabzadeh

Cross-Domain MetaDL: Any-Way Any-Shot Learning Competition with Novel Datasets from Practical Domains
Competition Organizers include: Isabelle Guyon

Reconnaissance Blind Chess: An Unsolved Challenge for Multi-Agent Decision Making Under Uncertainty
Competition Organizers include: Bo Li

VisDA 2022 Challenge: Sim2Real Domain Adaptation for Industrial Recycling
Competition Organizers include: Dina Bashkirova


Spotlight Papers

CoPur: Certifiably Robust Collaborative Inference via Feature Purification
Jing Liu, Chulin Xie, Oluwasanmi O Koyejo, Bo Li

Machine Learning on Graphs: A Model and Comprehensive Taxonomy
Ines Chami*, Sami Abu-El-Haija, Bryan Perozzi, Christopher Ré, Kevin Murphy

Sparse Winning Tickets are Data-Efficient Image Recognizers
Mukund Varma T, Xuxi Chen, Zhenyu Zhang, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang

Federated Learning from Pre-trained Models: A Contrastive Learning Approach
Yue Tan, Guodong Long, Jie Ma, Lu Liu, Tianyi Zhou, Jing Jiang

Improving Multi-task Generalization via Regularizing Spurious Correlation
Ziniu Hu*, Zhe Zhao, Xinyang Yi, Tiansheng Yao, Lichan Hong, Yizhou Sun, Ed H. Chi

The Nature of Temporal Difference Errors in Multi-step Distributional Reinforcement Learning
Yunhao Tang, Mark Rowland, Rémi Munos, Bernardo Ávila Pires, Will Dabney, Marc G. Bellemare

Residual Multiplicative Filter Networks for Multiscale Reconstruction
Shayan Shekarforoush, David B. Lindell, David J. Fleet, Marcus A Brubaker

Differentially Private Learning with Margin Guarantees
Raef Bassily, Mehryar Mohri, Ananda Theertha Suresh

Optimal Query Complexities for Dynamic Trace Estimation
David P. Woodruff*, Fred Zhang*, Qiuyi Zhang


Papers

From Gradient Flow on Population Loss to Learning with Stochastic Gradient Descent
Ayush Sekhari, Satyen Kale, Jason D. Lee, Chris De Sa, Karthik Sridharan

On the Global Convergence Rates of Decentralized Softmax Gradient Play in Markov Potential Games
Runyu Zhang, Jincheng Mei, Bo Dai, Dale Schuurmans, Na Li

Matryoshka Representation Learning
Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, Ali Farhadi

Efficient Risk-Averse Reinforcement Learning
Ido Greenberg, Yinlam Chow, Mohammad Ghavamzadeh, Shie Mannor

Operator Splitting Value Iteration
Amin Rakhsha, Andrew Wang, Mohammad Ghavamzadeh, Amir-massoud Farahmand

Cluster Randomized Designs for One-Sided Bipartite Experiments
Jennifer Brennan*, Vahab Mirrokni, Jean Pouget-Abadie

A Unified Sequence Interface for Vision Tasks
Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin*, David J. Fleet, Geoffrey Hinton

Cryptographic Hardness of Learning Halfspaces with Massart Noise
Ilias Diakonikolas, Daniel M. Kane, Pasin Manurangsi, Lisheng Ren

Better Best of Both Worlds Bounds for Bandits with Switching Costs
Idan Amir, Guy Azov, Tomer Koren, Roi Livni

Fast Neural Kernel Embeddings for General Activations
Insu Han, Amir Zandieh, Jaehoon Lee, Roman Novak, Lechao Xiao, Amin Karbasi

Hierarchical Agglomerative Graph Clustering in Poly-Logarithmic Depth
Laxman Dhulipala, David Eisenstat, Jakub Łącki, Vahab Mirronki, Jessica Shi

Improving Zero-Shot Generalization in Offline Reinforcement Learning Using Generalized Similarity Functions
Bogdan Mazoure*, Ilya Kostrikov, Ofir Nachum, Jonathan Tompson

Indicators of Attack Failure: Debugging and Improving Optimization of Adversarial Examples
Maura Pintor, Luca Demetrio, Angelo Sotgiu, Ambra Demontis, Nicholas Carlini, Battista Biggio, Fabio Roli

Learning Energy Networks with Generalized Fenchel-Young Losses
Mathieu Blondel, Felipe Llinares-López, Robert Dadashi, Léonard Hussenot, Matthieu Geist

Learning Robust Dynamics Through Variational Sparse Gating
Arnav Kumar Jain, Shiva Kanth Sujit, Shruti Joshi, Vincent Michalski, Danijar Hafner, Samira Ebrahimi Kahou

Learning to Reason with Neural Networks: Generalization, Unseen Data and Boolean Measures
Arnav Kumar Jain, Shiva Kanth Sujit, Shruti Joshi, Vincent Michalski, Danijar Hafner, Samira Ebrahimi Kahou

So3krates: Equivariant Attention for Interactions on Arbitrary Length-Scales in Molecular Systems
J. Thorben Frank, Oliver T. Unke, Klaus-Robert Müller

Spectral Bias in Practice: The Role of Function Frequency in Generalization
Sara Fridovich-Keil*, Raphael Gontijo-Lopes, Rebecca Roelofs

Delving into Out-of-Distribution Detection with Vision-Language Representations
Yifei Ming, Ziyang Cai, Jiuxiang Gu, Yiyou Sun, Wei Li, Yixuan Li

Path Independent Equilibrium Models Can Better Exploit Test-Time Computation
Cem Anil, Ashwini Pokle, Kaiqu Liang, Johannes Treutlein, Yuhuai Wu, Shaojie Bai, J. Zico Kolter, Roger Grosse

On Optimal Learning Under Targeted Data Poisoning
Steve Hanneke, Amin Karbasi, Mohammad Mahmoody, Idan Mehalel, Shay Moran

Learning With Little Mixing
Ingvar Ziemann, Stephen Tu

Block-Recurrent Transformers
DeLesley Hutchins, Imanol Schlag*, Yuhuai Wu, Ethan Dyer, Behnam Neyshabur

TabNAS: Rejection Sampling for Neural Architecture Search on Tabular Datasets
Chengrun Yang, Gabriel Bender, Hanxiao Liu, Pieter-Jan Kindermans, Madeleine Udell, Yifeng Lu, Quoc Le, Da Huang

Regret Bounds for Multilabel Classification in Sparse Label Regimes
Robert Busa-Fekete, Heejin Choi, Krzysztof Dembczynski, Claudio Gentile, Henry William Reeve, Balazs Szorenyi

Robust Reinforcement Learning Using Offline Data
Kishan Panaganti, Zaiyan Xu, Dileep Kalathil, Mohammad Ghavamzadeh

Contrastive Learning as Goal-Conditioned Reinforcement Learning
Benjamin Eysenbach, Tianjun Zhang, Sergey Levine, Ruslan Salakhutdinov

Beyond Rewards: A Hierarchical Perspective on Offline Multiagent Behavioral Analysis
Shayegan Omidshafiei, Andrei Kapishnikov, Yannick Assogba, Lucas Dixon, Been Kim

Revisiting Neural Scaling Laws in Language and Vision
Ibrahim Alabdulmohsin, Behnam Neyshabur, Xiaohua Zhai

Polynomial Neural Fields for Subband Decomposition and Manipulation
Guandao Yang*, Sagie Benaim, Varun Jampani, Kyle Genova, Jonathan T. Barron, Thomas Funkhouser, Bharath Hariharan, Serge Belongie

First Is Better Than Last for Language Data Influence
Chih-Kuan Yeh, Ankur Taly, Mukund Sundararajan, Frederick Liu, Pradeep Ravikumar

The Privacy Onion Effect: Memorization Is Relative
Nicholas Carlini, Matthew Jagielski, Chiyuan Zhang, Nicolas Papernot, Andreas Terzis, Florian Tramer

Deep Hierarchical Planning from Pixels (see blog post)
Danijar Hafner, Kuang-Huei Lee, Ian Fischer, Pieter Abbeel

Discovered Policy Optimisation
Chris Lu, Jakub Grudzien Kuba, Alistair Letcher, Luke Metz, Christian Schroeder de Witt, Jakob Foerster

Semi-supervised Active Linear Regression
Fnu Devvrit, Nived Rajaraman, Pranjal Awasthi

Pruning’s Effect on Generalization Through the Lens of Training and Regularization
Tian Jin, Daniel M. Roy, Michael Carbin, Jonathan Frankle, Gintare Karolina Dziugaite

Exploring Length Generalization in Large Language Models
Cem Anil*, Yuhuai Wu, Anders Andreassen, Aitor Lewkowycz, Vedant Misra, Vinay Ramasesh, Ambrose Slone, Guy Gur-Ari, Ethan Dyer, Behnam Neyshabur

Fast Stochastic Composite Minimization and an Accelerated Frank-Wolfe Algorithm Under Parallelization
Benjamin Dubois-Taine, Francis Bach, Quentin Berthet, Adrien Taylor

Global Normalization for Streaming Speech Recognition in a Modular Framework
Ehsan Variani, Ke Wu, Michael Riley, David Rybach, Matt Shannon, Cyril Allauzen

Learning Predictions for Algorithms with Predictions
Mikhail Khodak, Maria-Florina Balcan, Ameet Talwalkar, Sergei Vassilvitskii

Multimodal Contrastive Learning with LIMoE: the Language-Image Mixture of Experts (see blog post)
Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, Neil Houlsby

Incrementality Bidding via Reinforcement Learning Under Mixed and Delayed Rewards
Ashwinkumar Badanidiyuru, Zhe Feng, Tianxi Li, Haifeng Xu*

Solving Quantitative Reasoning Problems with Language Models (see blog post)
Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, Vedant Misra

Anonymized Histograms in Intermediate Privacy Models
Badih Ghazi, Pritish Kamath, Ravi Kumar, Pasin Manurangsi

Efficient and Stable Fully Dynamic Facility Location
Sayan Bhattacharya, Nikos Parotsidis, Silvio Lattanzi

Are All Losses Created Equal: A Neural Collapse Perspective
Jinxin Zhou, Chong You, Xiao Li, Kangning Liu, Sheng Liu, Qing Qu, Zhihui Zhu

Universal Rates for Interactive Learning
Steve Hanneke, Amin Karbasi, Shay Moran, Grigoris Velegkas

Nearly Optimal Algorithms for Linear Contextual Bandits with Adversarial Corruptions
Jiafan He, Dongruo Zhou, Tong Zhang, Quanquan Gu

Multiclass Learnability Beyond the PAC Framework: Universal Rates and Partial Concept Classes
Alkis Kalavasis, Grigoris Velegkas, Amin Karbasi

Temporal Latent Bottleneck: Synthesis of Fast and Slow Processing Mechanisms in Sequence Learning
Cenk Baykal, Nishanth Dikkala, Rina Panigrahy, Cyrus Rashtchian, Xin Wang

Pre-trained Language Models for Interactive Decision-Making
Shuang Li, Xavier Puig, Chris Paxton, Yilun Du, Clinton Wang, Linxi Fan, Tao Chen, De-An Huang, Ekin Akyürek, Anima Anandkumar, Jacob Andreas, Igor Mordatch, Antonio Torralba, Yuke Zhu

Polynomial Neural Fields for Subband Decomposition and Manipulation
Guandao Yang*, Sagie Benaim, Varun Jampani, Kyle Genova, Jonathan T. Barron, Thomas Funkhouser, Bharath Hariharan, Serge Belongie

Submodular Maximization in Clean Linear Time
Wenxin Li, Moran Feldman, Ehsan Kazemi, Amin Karbasi

Reinforcement Learning with Logarithmic Regret and Policy Switches
Grigoris Velegkas, Zhuoran Yang, Amin Karbasi

Algorithms with Prediction Portfolios
Michael Dinitz, Sungjin Im, Thomas Lavastida, Benjamin Moseley, Sergei Vassilvitskii

Understanding and Improving Robustness of Vision Transformers Through Patch-Based Negative Augmentation
Yao Qin, Chiyuan Zhang, Ting Chen, Balaji Lakshminarayanan, Alex Beutel, Xuezhi Wang

Best of Both Worlds Model Selection
Aldo Pacchiano, Christoph Dann, Claudio Gentile

Fair Wrapping for Black-Box Predictions
Alexander Soen, Ibrahim Alabdulmohsin, Sanmi Koyejo, Yishay Mansour, Nyalleng Moorosi, Richard Nock, Ke Sun, Lexing Xie

A Reduction to Binary Approach for Debiasing Multiclass Datasets
Ibrahim Alabdulmohsin, Jessica Schrouff, Oluwasanmi Koyejo

Weighted Distillation with Unlabeled Examples
Fotis Iliopoulos, Vasilis Kontonis, Cenk Baykal, Gaurav Menghani, Khoa Trihn,Erik Vee

A Closer Look at Learned Optimization: Stability, Robustness, and Inductive Biases
James Harrison, Luke Metz, Jascha Sohl-Dickstein

Post-hoc Estimators for Learning to Defer to an Expert
Harikrishna Narasimhan, Wittawat Jitkrittum, Aditya Krishna Menon, Ankit Singh Rawat, Sanjiv Kumar

Model-Based RL with Optimistic Posterior Sampling: Structural Conditions and Sample Complexity
Alekh Agarwal, Tong Zhang

On the Statistical Efficiency of Reward-Free Exploration in Non-Linear RL
Jinglin Chen, Aditya Modi, Akshay Krishnamurthy, Nan Jiang, Alekh Agarwal

Towards Learning Universal Hyperparameter Optimizers with Transformers (see blog post)
Yutian Chen, Xingyou Song, Chansoo Lee, Zi Wang, Qiuyi Zhang, David Dohan, Kazuya Kawakami, Greg Kochanski, Arnaud Doucet, Marc'aurelio Ranzato, Sagi Perel, Nando de Freitas

Reproducibility in Optimization: Theoretical Framework and Limits
Kwangjun Ahn*, Prateek Jain, Ziwei Ji, Satyen Kale, Praneeth Netrapalli, Gil I. Shamir

Confident Adaptive Language Modeling
Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Q. Tran, Yi Tay, Donald Metzler

Reinforcement Learning with Neural Radiance Fields
Danny Driess, Ingmar Schubert, Pete Florence, Yunzhu Li, Marc Toussaint

Invariant and Transportable Representations for Anti-Causal Domain Shifts
Yibo Jiang, Victor Veitch

Simple Mechanisms for Welfare Maximization in Rich Advertising Auctions
Gagan Aggarwal, Kshipra Bhawalkar, Aranyak Mehta, Divyarthi Mohan, Alexandros Psomas

STaR: Bootstrapping Reasoning with Reasoning
Eric Zelikman, Yuhuai Wu, Jesse Mu, Noah D. Goodman

Stochastic Online Learning with Feedback Graphs: Finite-Time and Asymptotic Optimality
Teodor V. Marinov, Mehryar Mohri, Julian Zimmert

The Curse of Unrolling: Rate of Differentiating Through Optimization
Damien Scieur, Quentin Bertrand, Gauthier Gidel, Fabian Pedregosa

Visual Prompting via Image Inpainting
Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, Alexei A Efros

Multi-Class H-Consistency Bounds
Pranjal Awasthi, Anqi Mao, Mehryar Mohri, Yutao Zhong

Anonymous Bandits for Multi-User Systems
Hossein Esfandiari, Vahab Mirrokni, Jon Schneider

Understanding the Eluder Dimension
Gene Li, Pritish Kamath, Dylan J. Foster, Nathan Srebro

Why So Pessimistic? Estimating Uncertainties for Offline RL Through Ensembles, and Why Their Independence Matters
Seyed Kamyar Seyed Ghasemipour, Shixiang Shane Gu, Ofir Nachum

A Best-of-Both-Worlds Algorithm for Bandits with Delayed Feedback
Saeed Masoudian, Julian Zimmert, Yevgeny Seldin

A Theoretical View on Sparsely Activated Networks
Cenk Baykal, Nishanth Dikkala, Rina Panigrahy, Cyrus Rashtchian, Xin Wang

Chain of Thought Prompting Elicits Reasoning in Large Language Models (see blog post)
Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, Denny Zhou

Decoupled Context Processing for Context Augmented Language Modeling
Zonglin Li, Ruiqi Guo, Sanjiv Kumar

Exploring Through Random Curiosity with General Value Functions
Aditya Ramesh, Louis Kirsch, Sjoerd van Steenkiste, Jürgen Schmidhuber

Object Scene Representation Transformer
Mehdi S. M. Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd van Steenkiste, Filip Pavetić, Mario Lučić, Leonidas J. Guibas, Klaus Greff, Thomas Kipf

Joint Model-Policy Optimization of a Lower Bound for Model-Based RL
Benjamin Eysenbach, Alexander Khazatsky, Sergey Levine, Ruslan Salakhutdinov

A Fourier Approach to Mixture Learning
Mingda Qiao*, Guru Guruganesh, Ankit Singh Rawat, Avinava Dubey, Manzil Zaheer

Why Neural Networks Find Simple Solutions: The Many Regularizers of Geometric Complexity
Benoit Dherin, Michael Munn, Mihaela Rosca, David Barrett

Do Current Multi-task Optimization Methods in Deep Learning Even Help?
Derrick Xin, Behrooz Ghorbani, Ankush Garg, Orhan Firat, Justin Gilmer

Associating Objects and Their Effects in Video Through Coordination Games
Erika Lu, Forrester Cole, Weidi Xie, Tali Dekel, William Freeman, Andrew Zisserman, Michael Rubinstein

Increasing Confidence in Adversarial Robustness Evaluations
Roland S. Zimmermann*, Wieland Brendel, Florian Tramèr, Nicholas Carlini

The Role of Baselines in Policy Gradient Optimization
Jincheng Mei, Wesley Chung, Valentin Thomas, Bo Dai, Csaba Szepesvari, Dale Schuurmans

Scaling Multimodal Pre-training via Cross-Modality Gradient Harmonization
Junru Wu, Yi Liang, Feng Han, Hassan Akbari, Zhangyang Wang, Cong Yu*

S3GC: Scalable Self-Supervised Graph Clustering
Fnu Devvrit*, Aditya Sinha, Inderjit Dhillon, Prateek Jain

Algorithms and Hardness for Learning Linear Thresholds from Label Proportions
Rishi Saket

ALMA: Hierarchical Learning for Composite Multi-Agent Tasks
Shariq Iqbal, Robby Costales, Fei Sha

DC-BENCH: Dataset Condensation Benchmark
Justin Cui, Ruochen Wang, Si Si, Cho-Jui Hsieh

Does GNN Pre-training Help Molecular Representation?
Ruoxi Sun, Hanjun Dai, Adams Yu

Drawing Out of Distribution with Neuro-Symbolic Generative Models
Yichao Liang, Joshua B. Tenenbaum, Tuan Anh Le, N. Siddharth

Mixture-of-Experts with Expert Choice Routing (see blog post)
Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew Dai, Zhifeng Chen, Quoc Le, James Laudon

Near-Optimal Regret for Adversarial MDP with Delayed Bandit Feedback
Tiancheng Jin, Tal Lancewicki, Haipeng Luo, Yishay Mansour, Aviv Rosenberg

Precise Learning Curves and Higher-Order Scalings for Dot-Product Kernel Regression
Lechao Xiao, Jeffrey Pennington, Theodor Misiakiewicz, Hong Hu, Yue Lu

Rate-Optimal Online Convex Optimization in Adaptive Linear Control
Asaf Cassel, Alon Cohen, Tomer Koren

Why Neural Networks Find Simple Solutions: The Many Regularizers of Geometric Complexity
Benoit Dherin, Michael Munn, Mihaela Rosca, David G.T. Barrett

Private Isotonic Regression
Badih Ghazi, Pritish Kamath, Ravi Kumar, Pasin Manurangsi

Sketching Based Representations for Robust Image Classification with Provable Guarantees
Nishanth Dikkala, Sankeerth Rao Karingula, Raghu Meka, Jelani Nelson, Rina Panigrahy, Xin Wang

The Role of Baselines in Policy Gradient Optimization
Jincheng Mei, Wesley Chung, Valentin Thomas, Bo Dai, Csaba Szepesvari, Dale Schuurmans

Bringing Image Scene Structure to Video via Frame-Clip Consistency of Object Tokens
Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, Amir Globerson

Near-Optimal Private and Scalable k-Clustering
Vincent Cohen-Addad, Alessandro Epasto, Vahab Mirrokni, Shyam Narayanan*, Peilin Zhong

When Does Differentially Private Learning Not Suffer in High Dimensions?
Xuechen Li, Daogao Liu, Tatsunori Hashimoto, Huseyin A Inan, Janardhan Kulkarni, YinTat Lee, Abhradeep Guha Thakurta

End-to-End Learning to Index and Search in Large Output Spaces
Nilesh Gupta, Patrick H. Chen, Hsiang-Fu, Yu, Cho-Jui Hsieh, Inderjit S. Dhillon

A Boosting Approach to Reinforcement Learning
Nataly Brukhim, Elad Hazan, Karan Singh

FedRolex: Model-Heterogeneous Federated Learning with Rolling Sub-Model Extraction
Samiul Alam, Luyang Liu, Ming Yan, Mi Zhang

Non-Convex Online Learning via Algorithmic Equivalence
Udaya Ghai, Zhou Lu, Elad Hazan

Is this the Right Neighborhood? Accurate and Query Efficient Model Agnostic Explanations
Amit Dhurandhar, Karthikeyan Natesan Ramamurthy, Karthikeyan Shanmugam

SAVi++: Towards End-to-End Object-Centric Learning from Real-World Videos
Gamaleldin F. Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C. Mozer, Thomas Kipf

UViM: A Unified Modeling Approach for Vision with Learned Guiding Codes
Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, Neil Houlsby

Implicit Regularization or Implicit Conditioning? Exact Risk Trajectories of SGD in High Dimensions
Courtney Paquette, Elliot Paquette, Ben Adlam, Jeffrey Pennington

Multi-game Decision Transformers (see blog post)
Kuang-Huei Lee, Ofir Nachum, Mengjiao Yang, Lisa Lee, Daniel Freeman, Winnie Xu, Sergio Guadarrama, Ian Fischer, Eric Jang, Henryk Michalewski, Igor Mordatch

Subsidiary Prototype Alignment for Universal Domain Adaptation
Jogendra Nath Kundu, Suvaansh Bhambri, Akshay Ravindra Kulkarni, Hiran Sarkar, Varun Jampani, Venkatesh Babu Radhakrishnan

SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections
Mark Boss*, Andreas Engelhardt*, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P. A. Lensch, Varun Jampani

Chefs’ Random Tables: Non-Trigonometric Random Features
Valerii Likhosherstov, Krzysztof Marcin Choromanski, Avinava Dubey, Frederick Liu, Tamas Sarlos, Adrian Weller

Lottery Tickets on a Data Diet: Finding Initializations with Sparse Trainable Networks
Mansheej Paul, Brett W Larsen, Surya Ganguli, Jonathan Frankle, Gintare Karolina Dziugaite

DP-PCA: Statistically Optimal and Differentially Private PCA
Xiyang Liu, Weihao Kong, Prateek Jain, Sewoong Oh

Emergent Communication: Generalization and Overfitting in Lewis Games
Mathieu Rita, Corentin Tallec, Paul Michel, Jean-Bastien Grill, Olivier Pietquin, Emmanuel Dupoux, Florian Strub

Handcrafted Backdoors in Deep Neural Networks
Sanghyun Hong, Nicholas Carlini, Alexey Kurakin

I2DFormer: Learning Image to Document Attention for Zero-Shot Image Classification
Muhammad Ferjad Naeem, Yongqin Xian, Luc Van Gool, Federico Tombari

Improved Differential Privacy for SGD via Optimal Private Linear Operators on Adaptive Streams
Sergey Denisov, Brendan McMahan, Keith Rush, Adam Smith, Abhradeep Guha Thakurta

Optimal Scaling for Locally Balanced Proposals in Discrete Spaces
Haoran Sun*, Hanjun Dai, Dale Schuurmans

Near-Optimal Correlation Clustering with Privacy
Vincent Cohen-Addad, Chenglin Fan, Silvio Lattanzi, Slobodan Mitrović, Ashkan Norouzi-Fard, Nikos Parotsidis, Jakub Tarnawski

Thor: Wielding Hammers to Integrate Language Models and Automated Theorem Provers
Albert Q. Jiang, Wenda Li, Szymon Tworkowski, Konrad Czechowski, Tomasz Odrzygóźdź, Piotr Miłoś, Yuhuai Wu, Mateja Jamnik

TPU-KNN: K Nearest Neighbor Search at Peak FLOP/s
Felix Chern, Blake Hechtman, Andy Davis, Ruiqi Guo, David Majnemer, Sanjiv Kumar

When Does Dough Become a Bagel? Analyzing the Remaining Mistakes on ImageNet
Vijay Vasudevan, Benjamin Caine, Raphael Gontijo-Lopes, Sara Fridovich-Keil, Rebecca Roelofs

DASCO: Dual-Generator Adversarial Support Constrained Offline Reinforcement Learning
Quan Vuong, Aviral Kumar, Sergey Levine, Yevgen Chebotar

A Characterization of Semi-Supervised Adversarially Robust PAC Learnability
Idan Attias, Steve Hanneke, Yishay Mansour

Back Razor: Memory-Efficient Transfer Learning by Self-Sparsified Backpropagation
Ziyu Jiang, Xuxi Chen, Xueqin Huang, Xianzhi Du, Denny Zhou, Zhangyang Wang

Subquadratic Kronecker Regression with Applications to Tensor Decomposition
Matthew Fahrbach, Gang Fu, Mehrdad Ghadiri

Zero-Shot Transfer Learning Within a Heterogeneous Graph via Knowledge Transfer Networks
Minji Yoon*, John Palowitch, Dustin Zelle, Ziniu Hu*, Ruslan Salakhutdinov, Bryan Perozzi

Differentially Private Graph Learning via Sensitivity-Bounded Personalized PageRank
Alessandro Epasto, Vahab Mirrokni, Bryan Perozzi, Anton Tsitsulin, Peilin Zhong

Reincarnating Reinforcement Learning: Reusing Prior Computation to Accelerate Progress (see blog post)
Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, Marc G. Bellemare

Private and Communication-Efficient Algorithms for Entropy Estimation
Gecia Bravo-Hermsdorff, Robert Busa-Fekete, Mohammad Ghavamzadeh, Andres Munoz Medina, Umar Syed

Oracle Inequalities for Model Selection in Offline Reinforcement Learning
Jonathan Lee, George Tucker, Ofir Nachum, Bo Dai, Emma Brunskill

Diagnosing Failures of Fairness Transfer Across Distribution Shift in Real-World Medical Settings
Jessica Schrouff*, Natalie Harris, Oluwasanmi O Koyejo, Ibrahim Alabdulmohsin, Eva Schnider*, Krista Opsahl-Ong, Alexander Brown, Subhrajit Roy, Diana Mincu, Christina Chen, Awa Dieng, Yuan Liu, Vivek Natarajan, Alan Karthikesalingam, Katherine A Heller, Silvia Chiappa, Alexander D'Amour

LASSIE: Learning Articulated Shapes from Sparse Image Ensemble via 3D Part Discovery
Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, Varun Jampani

Patching Open-Vocabulary Models by Interpolating Weights
Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, Ludwig Schmidt

TUSK: Task-Agnostic Unsupervised Keypoints
Yuhe Jin, Weiwei Sun, Jan Hosang, Eduard Trulls, Kwang Moo Yi

Active Learning of Classifiers with Label and Seed Queries
Marco Bressan, Nicolò Cesa-Bianchi, Silvio Lattanzi, Andrea Paudice, Maximilian Thiessen

Autoformalization with Large Language Models
Yuhuai Wu, Albert Q. Jiang, Wenda Li, Markus N. Rabe, Charles Staats, Mateja Jamnik, Christian Szegedy

Benign Underfitting of Stochastic Gradient Descent
Tomer Koren, Roi Livni, Yishay Mansour, Uri Sherman

Chain of Thought Imitation with Procedure Cloning
Mengjiao Yang, Dale Schuurmans, Pieter Abbeel, Ofir Nachum

Efficient and Modular Implicit Differentiation
Mathieu Blondel, Quentin Berthet, Marco Cuturi, Roy Frostig, Stephan Hoyer, Felipe Llinares-López, Fabian Pedregosa, Jean-Philippe Vert

Insights into Pre-training via Simpler Synthetic Tasks
Yuhuai Wu, Felix Li, Percy Liang

Self-Supervised Learning with an Information Maximization Criterion
Serdar Ozsoy, Shadi Hamdan, Sercan Ö. Arik, Deniz Yuret, Alper T. Erdogan

Trimmed Maximum Likelihood Estimation for Robust Generalized Linear Model
Weihao Kong, Rajat Sen, Pranjal Awasthi, Abhimanyu Das

Using Embeddings for Causal Estimation of Peer Influence in Social Networks
Irina Cristali, Victor Veitch

VCT: A Video Compression Transformer
Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, Eirikur Agustsson

Video Diffusion Models
Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, David J. Fleet

Large Language Models are Zero-Shot Reasoners
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, Yusuke Iwasawa

Improved Coresets for Euclidean k-Means
Vincent Cohen-Addad, Kasper Green Larsen, David Saulpic, Chris Schwiegelshohn, Omar Ali Sheikh-Omar

On the Adversarial Robustness of Mixture of Experts
Joan Puigcerver, Rodolphe Jenatton, Carlos Riquelme Ruiz, Pranjal Awasthi, Srinadh Bhojanapalli

Stars: Tera-Scale Graph Building for Clustering and Learning
CJ Carey, Jonathan Halcrow, Rajesh Jayaram, Vahab Mirrokni, Warren Schudy, Peilin Zhong

VER: Scaling On-Policy RL Leads to the Emergence of Navigation in Embodied Rearrangement
Erik Wijmans, Irfan Essa, Dhruv Batra

TaSIL: Taylor Series Imitation Learning
Daniel Pfrommer, Thomas TCK Zhang, Stephen Tu, Nikolai Matni

RNNs of RNNs: Recursive Construction of Stable Assemblies of Recurrent Neural Networks
Leo Kozachkov, Michaela M Ennis, Jean-Jacques Slotine

Integral Probability Metrics PAC-Bayes Bounds
Ron Amit, Baruch Epstein, Shay Moran, Ron Meir

D2NeRF: Self-Supervised Decoupling of Dynamic and Static Objects from a Monocular Video
Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, Cengiz Oztireli

Posted Pricing and Dynamic Prior-Independent Mechanisms with Value Maximizers
Yuan Deng, Vahab Mirrokni, Hanrui Zhang

Transformer Memory as a Differentiable Search Index
Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, Donald Metzler



*Work done while at Google.  

Source: Google AI Blog


Mixture-of-Experts with Expert Choice Routing

The capacity of a neural network to absorb information is limited by the number of its parameters, and as a consequence, finding more effective ways to increase model parameters has become a trend in deep learning research. Mixture-of-experts (MoE), a type of conditional computation where parts of the network are activated on a per-example basis, has been proposed as a way of dramatically increasing model capacity without a proportional increase in computation. In sparsely-activated variants of MoE models (e.g., Switch Transformer, GLaM, V-MoE), a subset of experts is selected on a per-token or per-example basis, thus creating sparsity in the network. Such models have demonstrated better scaling in multiple domains and better retention capability in a continual learning setting (e.g., Expert Gate). However, a poor expert routing strategy can cause certain experts to be under-trained, leading to an expert being under or over-specialized.

In “Mixture-of-Experts with Expert Choice Routing”, presented at NeurIPS 2022, we introduce a novel MoE routing algorithm called Expert Choice (EC). We discuss how this novel approach can achieve optimal load balancing in an MoE system while allowing heterogeneity in token-to-expert mapping. Compared to token-based routing and other routing methods in traditional MoE networks, EC demonstrates very strong training efficiency and downstream task scores. Our method resonates with one of the vision for Pathways, which is to enable heterogeneous mixture-of-experts via Pathways MPMD (multi program, multi data) support.


Overview of MoE Routing

MoE operates by adopting a number of experts, each as a sub-network, and activating only one or a few experts for each input token. A gating network must be chosen and optimized in order to route each token to the most suited expert(s). Depending on how tokens are mapped to experts, MoE can be sparse or dense. Sparse MoE only selects a subset of experts when routing each token, reducing computational cost as compared to a dense MoE. For example, recent work has implemented sparse routing via k-means clustering, linear assignment to maximize token-expert affinities, or hashing. Google also recently announced GLaM and V-MoE, both of which advance the state of the art in natural language processing and computer vision via sparsely gated MoE with top-k token routing, demonstrating better performance scaling with sparsely activated MoE layers. Many of these prior works used a token choice routing strategy in which the routing algorithm picks the best one or two experts for each token.

Token Choice Routing. The routing algorithm picks the top-1 or top-2 experts with highest affinity scores for each token. The affinity scores can be trained together with model parameters.

The independent token choice approach often leads to an imbalanced load of experts and under-utilization. In order to mitigate this, previous sparsely gated networks introduced additional auxiliary losses as regularization to prevent too many tokens being routed to a single expert, but the effectiveness was limited. As a result, token choice routings need to overprovision expert capacity by a significant margin (2x–8x of the calculated capacity) to avoid dropping tokens when there is a buffer overflow.

In addition to load imbalance, most prior works allocate a fixed number of experts to each token using a top-k function, regardless of the relative importance of different tokens. We argue that different tokens should be received by a variable number of experts, conditioned on token importance or difficulty.


Expert Choice Routing

To address the above issues, we propose a heterogeneous MoE that employs the expert choice routing method illustrated below. Instead of having tokens select the top-k experts, the experts with predetermined buffer capacity are assigned to the top-k tokens. This method guarantees even load balancing, allows a variable number of experts for each token, and achieves substantial gains in training efficiency and downstream performance. EC routing speeds up training convergence by over 2x in an 8B/64E (8 billion activated parameters, 64 experts) model, compared to the top-1 and top-2 gating counterparts in Switch Transformer, GShard, and GLaM.

Expert Choice Routing. Experts with predetermined buffer capacity are assigned top-k tokens, thus guaranteeing even load balancing. Each token can be received by a variable number of experts.

In EC routing, we set expert capacity k as the average tokens per expert in a batch of input sequences multiplied by a capacity factor, which determines the average number of experts that can be received by each token. To learn the token-to-expert affinity, our method produces a token-to-expert score matrix that is used to make routing decisions. The score matrix indicates the likelihood of a given token in a batch of input sequences being routed to a given expert.

Similar to Switch Transformer and GShard, we apply an MoE and gating function in the dense feedforward (FFN) layer, as it is the most computationally expensive part of a Transformer-based network. After producing the token-to-expert score matrix, a top-k function is applied along the token dimension for each expert to pick the most relevant tokens. A permutation function is then applied based on the generated indexes of the token, to create a hidden value with an additional expert dimension. The data is split across multiple experts such that all experts can execute the same computational kernel concurrently on a subset of tokens. Because a fixed expert capacity can be determined, we no longer overprovision expert capacity due to load imbalancing, thus significantly reducing training and inference step time by around 20% compared to GLaM.


Evaluation

To illustrate the effectiveness of Expert Choice routing, we first look at training efficiency and convergence. We use EC with a capacity factor of 2 (EC-CF2) to match the activated parameter size and computational cost on a per-token basis to GShard top-2 gating and run both for a fixed number of steps. EC-CF2 reaches the same perplexity as GShard top-2 in less than half the steps and, in addition, we find that each GShard top-2 step is 20% slower than our method.

We also scale the number of experts while fixing the expert size to 100M parameters for both EC and GShard top-2 methods. We find that both work well in terms of perplexity on the evaluation dataset during pre-training — having more experts consistently improves training perplexity.

Evaluation results on training convergence: EC routing yields 2x faster convergence at 8B/64E scale compared to top-2 gating used in GShard and GLaM (top). EC training perplexity scales better with the scaling of number of experts (bottom).

To validate whether improved perplexity directly translates to better performance in downstream tasks, we perform fine-tuning on 11 selected tasks from GLUE and SuperGLUE. We compare three MoE methods including Switch Transformer top-1 gating (ST Top-1), GShard top-2 gating (GS Top-2) and a version of our method (EC-CF2) that matches the activated parameters and computational cost of GS Top-2. The EC-CF2 method consistently outperforms the related methods and yields an average accuracy increase of more than 2% in a large 8B/64E setting. Comparing our 8B/64E model against its dense counterpart, our method achieves better fine-tuning results, increasing the average score by 3.4 points.

Our empirical results indicate that capping the number of experts for each token hurts the fine-tuning score by 1 point on average. This study confirms that allowing a variable number of experts per token is indeed helpful. On the other hand, we compute statistics on token-to-expert routing, particularly on the ratio of tokens that have been routed to a certain number of experts. We find that a majority of tokens have been routed to one or two experts while 23% have been routed to three or four experts and only about 3% tokens have been routed to more than four experts, thus verifying our hypothesis that expert choice routing learns to allocate a variable number of experts to tokens.


Final Thoughts

We propose a new routing method for sparsely activated mixture-of-experts models. This method addresses load imbalance and under-utilization of experts in conventional MoE methods, and enables the selection of different numbers of experts for each token. Our model demonstrates more than 2x training efficiency improvement when compared to the state-of-the-art GShard and Switch Transformer models, and achieves strong gains when fine-tuning on 11 datasets in the GLUE and SuperGLUE benchmark.

Our approach for expert choice routing enables heterogeneous MoE with straightforward algorithmic innovations. We hope that this may lead to more advances in this space at both the application and system levels.


Acknowledgements

Many collaborators across google research supported this work. We particularly thank Nan Du, Andrew Dai, Yanping Huang, and Zhifeng Chen for the initial ground work on MoE infrastructure and Tarzan datasets. We greatly appreciate Hanxiao Liu and Quoc Le for contributing the initial ideas and discussions. Tao Lei, Vincent Zhao, Da Huang, Chang Lan, Daiyi Peng, and Yifeng Lu contributed significantly on implementations and evaluations. Claire Cui, James Laudon, Martin Abadi, and Jeff Dean provided invaluable feedback and resource support.

Source: Google AI Blog


Multimodal Bottleneck Transformer (MBT): A New Model for Modality Fusion

People interact with the world through multiple sensory streams (e.g., we see objects, hear sounds, read words, feel textures and taste flavors), combining information and forming associations between senses. As real-world data consists of various signals that co-occur, such as video frames and audio tracks, web images and their captions and instructional videos and speech transcripts, it is natural to apply a similar logic when building and designing multimodal machine learning (ML) models.

Effective multimodal models have wide applications — such as multilingual image retrieval, future action prediction, and vision-language navigation — and are important for several reasons; robustness, which is the ability to perform even when one or more modalities is missing or corrupted, and complementarity between modalities, which is the idea that some information may be present only in one modality (e.g., audio stream) and not in the other (e.g., video frames). While the dominant paradigm for multimodal fusion, called late fusion, consists of using separate models to encode each modality, and then simply combining their output representations at the final step, investigating how to effectively and efficiently combine information from different modalities is still understudied.

In “Attention Bottlenecks for Multimodal Fusion”, published at NeurIPS 2021, we introduce a novel transformer-based model for multimodal fusion in video called Multimodal Bottleneck Transformer (MBT). Our model restricts cross-modal attention flow between latent units in two ways: (1) through tight fusion bottlenecks, that force the model to collect and condense the most relevant inputs in each modality (sharing only necessary information with other modalities), and (2) to later layers of the model, allowing early layers to specialize to information from individual modalities. We demonstrate that this approach achieves state-of-the-art results on video classification tasks, with a 50% reduction in FLOPs compared to a vanilla multimodal transformer model. We have also released our code as a tool for researchers to leverage as they expand on multimodal fusion work.

A Vanilla Multimodal Transformer Model
Transformer models consistently obtain state-of-the-art results in ML tasks, including video (ViViT) and audio classification (AST). Both ViViT and AST are built on the Vision Transformer (ViT); in contrast to standard convolutional approaches that process images pixel-by-pixel, ViT treats an image as a sequence of patch tokens (i.e., tokens from a smaller part, or patch, of an image that is made up of multiple pixels). These models then perform self-attention operations across all pairs of patch tokens. However, using transformers for multimodal fusion is challenging because of their high computational cost, with complexity scaling quadratically with input sequence length.

Because transformers effectively process variable length sequences, the simplest way to extend a unimodal transformer, such as ViT, to the multimodal case is to feed the model a sequence of both visual and auditory tokens, with minimal changes to the transformer architecture. We call this a vanilla multimodal transformer model, which allows free attention flow (called vanilla cross-attention) between different spatial and temporal regions in an image, and across frequency and time in audio inputs, represented by spectrograms. However, while easy to implement by concatenating audio and video input tokens, vanilla cross-attention at all layers of the transformer model is unnecessary because audio and visual inputs contain dense, fine-grained information, which may be redundant for the task — increasing complexity.

Restricting Attention Flow
The issue of growing complexity for long sequences in multimodal models can be mitigated by reducing the attention flow. We restrict attention flow using two methods, specifying the fusion layer and adding attention bottlenecks.

  • Fusion layer (early, mid or late fusion): In multimodal models, the layer where cross-modal interactions are introduced is called the fusion layer. The two extreme versions are early fusion (where all layers in the transformer are cross-modal) and late fusion (where all layers are unimodal and no cross-modal information is exchanged in the transformer encoder). Specifying a fusion layer in between leads to mid fusion. This technique builds on a common paradigm in multimodal learning, which is to restrict cross-modal flow to later layers of the network, allowing early layers to specialize in learning and extracting unimodal patterns.
  • Attention bottlenecks: We also introduce a small set of latent units that form an attention bottleneck (shown below in purple), which force the model, within a given layer, to collate and condense information from each modality before sharing it with the other, while still allowing free attention flow within a modality. We demonstrate that this bottlenecked version (MBT), outperforms or matches its unrestricted counterpart with lower computational cost.
The different attention configurations in our model. Unlike late fusion (top left), where no cross-modal information is exchanged in the transformer encoder, we investigate two pathways for the exchange of cross-modal information. Early and mid fusion (top middle, top right) is done via standard pairwise self attention across all hidden units in a layer. For mid fusion, cross-modal attention is applied only to later layers in the model. Bottleneck fusion (bottom left) restricts attention flow within a layer through tight latent units called attention bottlenecks. Bottleneck mid fusion (bottom right) applies both forms of restriction in conjunction for optimal performance.

Bottlenecks and Computation Cost
We apply MBT to the task of sound classification using the AudioSet dataset and investigate its performance for two approaches: (1) vanilla cross-attention, and (2) bottleneck fusion. For both approaches, mid fusion (shown by the middle values of the x-axis below) outperforms both early (fusion layer = 0) and late fusion (fusion layer = 12). This suggests that the model benefits from restricting cross-modal connections to later layers, allowing earlier layers to specialize in learning unimodal features; however, it still benefits from multiple layers of cross-modal information flow. We find that adding attention bottlenecks (bottleneck fusion) outperforms or maintains performance with vanilla cross-attention for all fusion layers, with more prominent improvements at lower fusion layers.

The impact of using attention bottlenecks for fusion on mAP performance (left) and compute (right) at different fusion layers on AudioSet. Attention bottlenecks (red) improve performance over vanilla cross-attention (blue) at lower computational cost. Mid fusion, which is in fusion layers 4-10, outperforms both early (fusion layer = 0) and late (fusion layer = 12) fusion, with best performance at fusion layer 8.

We compare the amount of computation, measured in GFLOPs, for both vanilla cross-attention and bottleneck fusion. Using a small number of attention bottlenecks (four bottleneck tokens used in our experiments) adds negligible extra computation over a late fusion model, with computation remaining largely constant with varying fusion layers. This is in contrast to vanilla cross-attention, which has a non-negligible computational cost for every layer it is applied to. We note that for early fusion, bottleneck fusion outperforms vanilla cross-attention by over 2 mean average precision points (mAP) on audiovisual sound classification, with less than half the computational cost.

Results on Sound Classification and Action Recognition
MBT outperforms previous research on popular video classification tasks — sound classification (AudioSet and VGGSound) and action recognition (Kinetics and Epic-Kitchens). For multiple datasets, late fusion and MBT with mid fusion (both fusing audio and vision) outperform the best single modality baseline, and MBT with mid fusion outperforms late fusion.

Across multiple datasets, fusing audio and vision outperforms the best single modality baseline, and MBT with mid fusion outperforms late fusion. For each dataset we report the widely used primary metric, i.e., Audioset: mAP, Epic-Kitchens: Top-1 action accuracy, VGGSound, Moments-in-Time and Kinetics: Top-1 classification accuracy.

Visualization of Attention Heatmaps
To understand the behavior of MBT, we visualize the attention computed by our network following the attention rollout technique. We compute heat maps of the attention from the output classification tokens to the image input space for a vanilla cross-attention model and MBT on the AudioSet test set. For each video clip, we show the original middle frame on the left with the ground truth labels overlayed at the bottom. We demonstrate that the attention is particularly focused on regions in the images that contain motion and create sound, e.g., the fingertips on the piano, the sewing machine, and the face of the dog. The fusion bottlenecks in MBT further force the attention to be localized to smaller regions of the images, e.g., the mouth of the dog in the top left and the woman singing in the middle right. This provides some evidence that the tight bottlenecks force MBT to focus only on the image patches that are relevant for an audio classification task and that benefit from mid fusion with audio.

Summary
We introduce MBT, a new transformer-based architecture for multimodal fusion, and explore various fusion approaches using cross-attention between bottleneck tokens. We demonstrate that restricting cross-modal attention via a small set of fusion bottlenecks achieves state-of-the-art results on a number of video classification benchmarks while also reducing computational costs compared to vanilla cross-attention models.

Acknowledgements
This research was conducted by Arsha Nagrani, Anurag Arnab, Shan Yang, Aren Jansen, Cordelia Schmid and Chen Sun. The blog post was written by Arsha Nagrani, Anurag Arnab and Chen Sun. Animations were created by Tom Small.


Source: Google AI Blog


Permutation-Invariant Neural Networks for Reinforcement Learning

“The brain is able to use information coming from the skin as if it were coming from the eyes. We don’t see with the eyes or hear with the ears, these are just the receptors, seeing and hearing in fact goes on in the brain.”
Paul Bach-y-Rita1

People have the amazing ability to use one sensory modality (e.g., touch) to supply environmental information normally gathered by another sense (e.g., vision). This adaptive ability, called sensory substitution, is a phenomenon well-known to neuroscience. While difficult adaptations — such as adjusting to seeing things upside-down, learning to ride a “backwards” bicycle, or learning to “see” by interpreting visual information emitted from a grid of electrodes placed on one’s tongue — require anywhere from weeks, months or even years to attain mastery, people are able to eventually adjust to sensory substitutions.

Examples of Sensory Substitution. Left: Tongue Display Unit (Maris and Bach-y-Rita, 2001; Image: Kaczmarek, 2011). Right: “Upside down goggles” initially conceived by Erismann and Kohler in 1931. (Image Wikipedia).

In contrast, most neural networks are not able to adapt to sensory substitutions at all. For instance, most reinforcement learning (RL) agents require their inputs to be in a pre-specified format, or else they will fail. They expect fixed-size inputs and assume that each element of the input carries a precise meaning, such as the pixel intensity at a specified location, or state information, like position or velocity. In popular RL benchmark tasks (e.g., Ant or Cart-pole), an agent trained using current RL algorithms will fail if its sensory inputs are changed or if the agent is fed additional noisy inputs that are unrelated to the task at hand.

In “The Sensory Neuron as a Transformer: Permutation-Invariant Neural Networks for Reinforcement Learning”, a spotlight paper at NeurIPS 2021, we explore permutation invariant neural network agents, which require each of their sensory neurons (receptors that receive sensory inputs from the environment) to figure out the meaning and context of its input signal, rather than explicitly assuming a fixed meaning. Our experiments show that such agents are robust to observations that contain additional redundant or noisy information, and to observations that are corrupt and incomplete.

Permutation invariant reinforcement learning agents adapting to sensory substitutions. Left: The ordering of the ant’s 28 observations are randomly shuffled every 200 time-steps. Unlike the standard policy, our policy is not affected by the suddenly permuted inputs. Right: Cart-pole agent given many redundant noisy inputs (Interactive web-demo).

In addition to adapting to sensory substitutions in state-observation environments (like the ant and cart-pole examples), we show that these agents can also adapt to sensory substitutions in complex visual-observation environments (such as a CarRacing game that uses only pixel observations) and can perform when the stream of input images is constantly being reshuffled:

We partition the visual input from CarRacing into a 2D grid of small patches, and shuffled their ordering. Without any additional training, our agent still performs even when the original training background (left) is replaced with new images (right).

Method
Our approach takes observations from the environment at each time-step and feeds each element of the observation into distinct, but identical neural networks (called “sensory neurons”), each with no fixed relationship with one another. Each sensory neuron integrates over time information from only their particular sensory input channel. Because each sensory neuron receives only a small part of the full picture, they need to self-organize through communication in order for a global coherent behavior to emerge.

Illustration of observation segmentation.We segment each input into elements, which are then fed to independent sensory neurons. For non-vision tasks where the inputs are usually 1D vectors, each element is a scalar. For vision tasks, we crop each input image into non-overlapping patches.

We encourage neurons to communicate with each other by training them to broadcast messages. While receiving information locally, each individual sensory neuron also continually broadcasts an output message at each time-step. These messages are consolidated and combined into an output vector, called the global latent code, using an attention mechanism similar to that applied in the Transformer architecture. A policy network then uses the global latent code to produce the action that the agent will use to interact with the environment. This action is also fed back into each sensory neuron in the next time-step, closing the communication loop.

Overview of the permutation-invariant RL method. We first feed each individual observation (ot) into a particular sensory neuron (along with the agent’s previous action, at-1). Each neuron then produces and broadcasts a message independently, and an attention mechanism summarizes them into a global latent code (mt) that is given to the agent's downstream policy network (?) to produce the agent’s action at.

Why is this system permutation invariant? Each sensory neuron is an identical neural network that is not confined to only process information from one particular sensory input. In fact, in our setup, the inputs to each sensory neuron are not defined. Instead, each neuron must figure out the meaning of its input signal by paying attention to the inputs received by the other sensory neurons, rather than explicitly assuming a fixed meaning. This encourages the agent to process the entire input as an unordered set, making the system to be permutation invariant to its input. Furthermore, in principle, the agent can use as many sensory neurons as required, thus enabling it to process observations of arbitrary length. Both of these properties will help the agent adapt to sensory substitutions.

Results
We demonstrate the robustness and flexibility of this approach in simpler, state-observation environments, where the observations the agent receives as inputs are low-dimensional vectors holding information about the agent’s states, such as the position or velocity of its components. The agent in the popular Ant locomotion task has a total of 28 inputs with information that includes positions and velocities. We shuffle the order of the input vector several times during a trial and show that the agent is rapidly able to adapt and is still able to walk forward.

In cart-pole, the agent’s goal is to swing up a cart-pole mounted at the center of the cart and balance it upright. Normally the agent sees only five inputs, but we modify the cartpole environment to provide 15 shuffled input signals, 10 of which are pure noise, and the remainder of which are the actual observations from the environment. The agent is still able to perform the task, demonstrating the system’s capacity to work with a large number of inputs and attend only to channels it deems useful. Such flexibility may find useful applications for processing a large unspecified number of signals, most of which are noise, from ill-defined systems.

We also apply this approach to high-dimensional vision-based environments where the observation is a stream of pixel images. Here, we investigate screen-shuffled versions of vision-based RL environments, where each observation frame is divided into a grid of patches, and like a puzzle, the agent must process the patches in a shuffled order to determine a course of action to take. To demonstrate our approach on vision-based tasks, we created a shuffled version of Atari Pong.

Shuffled Pong results. Left: Pong agent trained to play using only 30% of the patches matches performance of Atari opponent. Right: Without extra training, when we give the agent more puzzle pieces, its performance increases.

Here the agent’s input is a variable-length list of patches, so unlike typical RL agents, the agent only gets to “see” a subset of patches from the screen. In the puzzle pong experiment, we pass to the agent a random sample of patches across the screen, which are then fixed through the remainder of the game. We find that we can discard 70% of the patches (at these fixed-random locations) and still train the agent to perform well against the built-in Atari opponent. Interestingly, if we then reveal additional information to the agent (e.g., allowing it access to more image patches), its performance increases, even without additional training. When the agent receives all the patches, in shuffled order, it wins 100% of the time, achieving the same result with agents that are trained while seeing the entire screen.

We find that imposing additional difficulty during training by using unordered observations has additional benefits, such as improving generalization to unseen variations of the task, like when the background of the CarRacing training environment is replaced with a novel image.

Shuffled CarRacing results. The agent has learned to focus its attention (indicated by the highlighted patches) on the road boundaries. Left: Training environment. Right: Test environment with new background.

Conclusion
The permutation invariant neural network agents presented here can handle ill-defined, varying observation spaces. Our agents are robust to observations that contain redundant or noisy information, or observations that are corrupt and incomplete. We believe that permutation invariant systems open up numerous possibilities in reinforcement learning.

If you’re interested to learn more about this work, we invite readers to read our interactive article (pdf version) or watch our video. We also released code to reproduce our experiments.



1Quoted in Livewired, by David Eagleman.  

Source: Google AI Blog


RLiable: Towards Reliable Evaluation & Reporting in Reinforcement Learning

Reinforcement learning (RL) is an area of machine learning that focuses on learning from experiences to solve decision making tasks. While the field of RL has made great progress, resulting in impressive empirical results on complex tasks, such as playing video games, flying stratospheric balloons and designing hardware chips, it is becoming increasingly apparent that the current standards for empirical evaluation might give a false sense of fast scientific progress while slowing it down.

To that end, in “Deep RL at the Edge of the Statistical Precipice”, accepted as an oral presentation at NeurIPS 2021, we discuss how statistical uncertainty of results needs to be considered, especially when using only a few training runs, in order for evaluation in deep RL to be reliable. Specifically, the predominant practice of reporting point estimates ignores this uncertainty and hinders reproducibility of results. Related to this, tables with per-task scores, as are commonly reported, can be overwhelming beyond a few tasks and often omit standard deviations. Furthermore, simple performance metrics like the mean can be dominated by a few outlier tasks, while the median score would remain unaffected even if up to half of the tasks had performance scores of zero. Thus, to increase the field's confidence in reported results with a handful of runs, we propose various statistical tools, including stratified bootstrap confidence intervals, performance profiles, and better metrics, such as interquartile mean and probability of improvement. To help researchers incorporate these tools, we also release an easy-to-use Python library RLiable with a quickstart colab.

Statistical Uncertainty in RL Evaluation
Empirical research in RL relies on evaluating performance on a diverse suite of tasks, such as Atari 2600 video games, to assess progress. Published results on deep RL benchmarks typically compare point estimates of the mean and median scores aggregated across tasks. These scores are typically relative to some defined baseline and optimal performance (e.g., random agent and “average” human performance on Atari games, respectively) so as to make scores comparable across different tasks.

In most RL experiments, there is randomness in the scores obtained from different training runs, so reporting only point estimates does not reveal whether similar results would be obtained with new independent runs. A small number of training runs, coupled with the high variability in performance of deep RL algorithms, often leads to large statistical uncertainty in such point estimates.

The distribution of median human normalized scores on the Atari 100k benchmark, which contains 26 games, for five recently published algorithms, DER, OTR, CURL, two variants of DrQ, and SPR. The reported point estimates of median scores based on a few runs in publications, as shown by dashed lines, do not provide information about the variability in median scores and typically overestimate (e.g., CURL, SPR, DrQ) or underestimate (e.g., DER) the expected median, which can result in erroneous conclusions.

As benchmarks become increasingly more complex, evaluating more than a few runs will be increasingly demanding due to the increased compute and data needed to solve such tasks. For example, five runs on 50 Atari games for 200 million frames takes 1000+ GPU days. Thus, evaluating more runs is not a feasible solution for reducing statistical uncertainty on computationally demanding benchmarks. While prior work has recommended statistical significance tests as a solution, such tests are dichotomous in nature (either “significant” or “not significant”), so they often lack the granularity needed to yield meaningful insights and are widely misinterpreted.

Number of runs in RL papers over the years. Beginning with the Arcade Learning Environment (ALE), the shift toward computationally-demanding benchmarks has led to the practice of evaluating only a handful of runs per task, increasing the statistical uncertainty in point estimates.

Tools for Reliable Evaluation
Any aggregate metric based on a finite number of runs is a random variable, so to take this into account, we advocate for reporting stratified bootstrap confidence intervals (CIs), which predict the likely values of aggregate metrics if the same experiment were repeated with different runs. These CIs allow us to understand the statistical uncertainty and reproducibility of results. Such CIs use the scores on combined runs across tasks. For example, evaluating 3 runs each on Atari 100k, which contains 26 tasks, results in 78 sample scores for uncertainty estimation.

In each task, colored balls denote scores on different runs. To compute statified bootstrap CIs using the percentile method, bootstrap samples are created by randomly sampling scores with replacement proportionately from each task. Then, the distribution of aggregate scores on these samples is the bootstrapping distribution, whose spread around the center gives us the confidence interval.

Most deep RL algorithms often perform better on some tasks and training runs, but aggregate performance metrics can conceal this variability, as shown below.

Data with varied appearance but identical aggregate statistics. Source: Same Stats, Different Graphs.

Instead, we recommend performance profiles, which are typically used for comparing solve times of optimization software. These profiles plot the score distribution across all runs and tasks with uncertainty estimates using stratified bootstrap confidence bands. These plots show the total runs across all tasks that obtain a score above a threshold (?) as a function of the threshold.

Performance profiles correspond to the empirical tail distribution of scores on runs combined across all tasks. Shaded regions show 95% stratified bootstrap confidence bands.

Such profiles allow for qualitative comparisons at a glance. For example, the curve for one algorithm above another means that one algorithm is better than the other. We can also read any score percentile, e.g., the profiles intersect y = 0.5 (dotted line above) at the median score. Furthermore, the area under the profile corresponds to the mean score.

While performance profiles are useful for qualitative comparisons, algorithms rarely outperform other algorithms on all tasks and thus their profiles often intersect, so finer quantitative comparisons require aggregate performance metrics. However, existing metrics have limitations: (1) a single high performing task may dominate the task mean score, while (2) the task median is unaffected by zero scores on nearly half of the tasks and requires a large number of training runs for small statistical uncertainty. To address the above limitations, we recommend two alternatives based on robust statistics: the interquartile mean (IQM) and the optimality gap, both of which can be read as areas under the performance profile, below.

IQM (red) corresponds to the area under the performance profile, shown in blue, between the 25 and 75 percentile scores on the x-axis. Optimality gap (yellow) corresponds to the area between the profile and horizontal line at y = 1 (human performance), for scores less than 1.

As an alternative to median and mean, IQM corresponds to the mean score of the middle 50% of the runs combined across all tasks. It is more robust to outliers than mean, a better indicator of overall performance than median, and results in smaller CIs, and so, requires fewer runs to claim improvements. Another alternative to mean, optimality gap measures how far an algorithm is from optimal performance.

IQM discards the lowest 25% and highest 25% of the combined scores (colored balls) and computes the mean of the remaining 50% scores.

For directly comparing two algorithms, another metric to consider is the average probability of improvement, which describes how likely an improvement over baseline is, regardless of its size. This metric is computed using the Mann-Whitney U-statistic, averaged across tasks.

Re-evaluating Evaluation
Using the above tools for evaluation, we revisit performance evaluations of existing algorithms on widely used RL benchmarks, revealing inconsistencies in prior evaluation. For example, in the Arcade Learning Environment (ALE), a widely recognized RL benchmark, the performance ranking of algorithms changes depending on the choice of aggregate metric. Since performance profiles capture the full picture, they often illustrate why such inconsistencies exist.

Median (left) and IQM (right) human normalized scores on the ALE as a function of the number of environment frames seen during training. IQM results in significantly smaller CIs than median scores.

On DM Control, a popular continuous control benchmark, there are large overlaps in 95% CIs of mean normalized scores for most algorithms.

DM Control Suite results, averaged across six tasks, on the 100k and 500k step benchmark. Since scores are normalized using maximum performance, mean scores correspond to one minus the optimality gap. The ordering of the algorithms is based on their claimed relative performance — all algorithms except Dreamer claimed improvement over at least one algorithm placed below them. Shaded regions show 95% CIs.

Finally, on Procgen, a benchmark for evaluating generalization in RL, the average probability of improvement shows that some claimed improvements are only 50-70% likely, suggesting that some reported improvements could be spurious.

Each row shows the probability that the algorithm X on the left outperforms algorithm Y on the right, given that X was claimed to be better than Y. Shaded region denotes 95% stratified bootstrap CIs.

Conclusion
Our findings on widely-used deep RL benchmarks show that statistical issues can have a large influence on previously reported results. In this work, we take a fresh look at evaluation to improve the interpretation of reported results and standardize experimental reporting. We’d like to emphasize the importance of published papers providing results for all runs to allow for future statistical analyses. To build confidence in your results, please check out our open-source library RLiable and the quickstart colab.

Acknowledgments
This work was done in collaboration with Max Schwarzer, Aaron Courville and Marc G. Bellemare. We’d like to thank Tom Small for an animated figure used in this post. We are also grateful for feedback by several members of the Google Research, Brain Team and DeepMind.

Source: Google AI Blog


Constructing Transformers For Longer Sequences with Sparse Attention Methods

Natural language processing (NLP) models based on Transformers, such as BERT, RoBERTa, T5, or GPT3, are successful for a wide variety of tasks and a mainstay of modern NLP research. The versatility and robustness of Transformers are the primary drivers behind their wide-scale adoption, leading them to be easily adapted for a diverse range of sequence-based tasks — as a seq2seq model for translation, summarization, generation, and others, or as a standalone encoder for sentiment analysis, POS tagging, machine reading comprehension, etc. The key innovation in Transformers is the introduction of a self-attention mechanism, which computes similarity scores for all pairs of positions in an input sequence, and can be evaluated in parallel for each token of the input sequence, avoiding the sequential dependency of recurrent neural networks, and enabling Transformers to vastly outperform previous sequence models like LSTM.

A limitation of existing Transformer models and their derivatives, however, is that the full self-attention mechanism has computational and memory requirements that are quadratic with the input sequence length. With commonly available current hardware and model sizes, this typically limits the input sequence to roughly 512 tokens, and prevents Transformers from being directly applicable to tasks that require larger context, like question answering, document summarization or genome fragment classification. Two natural questions arise: 1) Can we achieve the empirical benefits of quadratic full Transformers using sparse models with computational and memory requirements that scale linearly with the input sequence length? 2) Is it possible to show theoretically that these linear Transformers preserve the expressivity and flexibility of the quadratic full Transformers?

We address both of these questions in a recent pair of papers. In “ETC: Encoding Long and Structured Inputs in Transformers”, presented at EMNLP 2020, we present the Extended Transformer Construction (ETC), which is a novel method for sparse attention, in which one uses structural information to limit the number of computed pairs of similarity scores. This reduces the quadratic dependency on input length to linear and yields strong empirical results in the NLP domain. Then, in “Big Bird: Transformers for Longer Sequences”, presented at NeurIPS 2020, we introduce another sparse attention method, called BigBird that extends ETC to more generic scenarios where prerequisite domain knowledge about structure present in the source data may be unavailable. Moreover, we also show that theoretically our proposed sparse attention mechanism preserves the expressivity and flexibility of the quadratic full Transformers. Our proposed methods achieve a new state of the art on challenging long-sequence tasks, including question answering, document summarization and genome fragment classification.

Attention as a Graph
The attention module used in Transformer models computes similarity scores for all pairs of positions in an input sequence. It is useful to think of the attention mechanism as a directed graph, with tokens represented by nodes and the similarity score computed between a pair of tokens represented by an edge. In this view, the full attention model is a complete graph. The core idea behind our approach is to carefully design sparse graphs, such that one only computes a linear number of similarity scores.

Full attention can be viewed as a complete graph.

Extended Transformer Construction (ETC)
On NLP tasks that require long and structured inputs, we propose a structured sparse attention mechanism, which we call Extended Transformer Construction (ETC). To achieve structured sparsification of self attention, we developed the global-local attention mechanism. Here the input to the Transformer is split into two parts: a global input where tokens have unrestricted attention, and a long input where tokens can only attend to either the global input or to a local neighborhood. This achieves linear scaling of attention, which allows ETC to significantly scale input length.

In order to further exploit the structure of long documents, ETC combines additional ideas: representing the positional information of the tokens in a relative way, rather than using their absolute position in the sequence; using an additional training objective beyond the usual masked language model (MLM) used in models like BERT; and flexible masking of tokens to control which tokens can attend to which other tokens. For example, given a long selection of text, a global token is applied to each sentence, which connects to all tokens within the sentence, and a global token is also applied to each paragraph, which connects to all tokens within the same paragraph.

An example of document structure based sparse attention of ETC model. The global variables are denoted by C (in blue) for paragraph, S (yellow) for sentence while the local variables are denoted by X (grey) for tokens corresponding to the long input.

With this approach, we report state-of-the-art results in five challenging NLP datasets requiring long or structured inputs: TriviaQA, Natural Questions (NQ), HotpotQA, WikiHop, and OpenKP.

Test set result on Question Answering. For both verified TriviaQA and WikiHop, using ETC achieved a new state of the art.

BigBird
Extending the work of ETC, we propose BigBird — a sparse attention mechanism that is also linear in the number of tokens and is a generic replacement for the attention mechanism used in Transformers. In contrast to ETC, BigBird doesn’t require any prerequisite knowledge about structure present in the source data. Sparse attention in the BigBird model consists of three main parts:

  • A set of global tokens attending to all parts of the input sequence
  • All tokens attending to a set of local neighboring tokens
  • All tokens attending to a set of random tokens
BigBird sparse attention can be seen as adding few global tokens on Watts-Strogatz graph.

In the BigBird paper, we explain why sparse attention is sufficient to approximate quadratic attention, partially explaining why ETC was successful. A crucial observation is that there is an inherent tension between how few similarity scores one computes and the flow of information between different nodes (i.e., the ability of one token to influence each other). Global tokens serve as a conduit for information flow and we prove that sparse attention mechanisms with global tokens can be as powerful as the full attention model. In particular, we show that BigBird is as expressive as the original Transformer, is computationally universal (following the work of Yun et al. and Perez et al.), and is a universal approximator of continuous functions. Furthermore, our proof suggests that the use of random graphs can further help ease the flow of information — motivating the use of the random attention component.

This design scales to much longer sequence lengths for both structured and unstructured tasks. Further scaling can be achieved by using gradient checkpointing by trading off training time for sequence length. This lets us extend our efficient sparse transformers to include generative tasks that require an encoder and a decoder, such as long document summarization, on which we achieve a new state of the art.

Summarization ROUGE score for long documents. Both for BigPatent and ArXiv datasets, we achieve a new state of the art result.

Moreover, the fact that BigBird is a generic replacement also allows it to be extended to new domains without pre-existing domain knowledge. In particular, we introduce a novel application of Transformer-based models where long contexts are beneficial — extracting contextual representations of genomic sequences (DNA). With longer masked language model pre-training, BigBird achieves state-of-the-art performance on downstream tasks, such as promoter-region prediction and chromatin profile prediction.

On multiple genomics tasks, such as promoter region prediction (PRP), chromatin-profile prediction including transcription factors (TF), histone-mark (HM) and DNase I hypersensitive (DHS) detection, we outperform baselines. Moreover our results show that Transformer models can be applied to multiple genomics tasks that are currently underexplored.

Main Implementation Idea
One of the main impediments to the large scale adoption of sparse attention is the fact that sparse operations are quite inefficient in modern hardware. Behind both ETC and BigBird, one of our key innovations is to make an efficient implementation of the sparse attention mechanism. As modern hardware accelerators like GPUs and TPUs excel using coalesced memory operations, which load blocks of contiguous bytes at once, it is not efficient to have small sporadic look-ups caused by a sliding window (for local attention) or random element queries (random attention). Instead we transform the sparse local and random attention into dense tensor operations to take full advantage of modern single instruction, multiple data (SIMD) hardware.

To do this, we first “blockify” the attention mechanism to better leverage GPUs/TPUs, which are designed to operate on blocks. Then we convert the sparse attention mechanism computation into a dense tensor product through a series of simple matrix operations such as reshape, roll, and gather, as illustrated in the animation below.

Illustration of how sparse window attention is efficiently computed using roll and reshape, and without small sporadic look-ups.

Recently, “Long Range Arena: A Benchmark for Efficient Transformers“ provided a benchmark of six tasks that require longer context, and performed experiments to benchmark all existing long range transformers. The results show that the BigBird model, unlike its counterparts, clearly reduces memory consumption without sacrificing performance.

Conclusion
We show that carefully designed sparse attention can be as expressive and flexible as the original full attention model. Along with theoretical guarantees, we provide a very efficient implementation which allows us to scale to much longer inputs. As a consequence, we achieve state-of-the-art results for question answering, document summarization and genome fragment classification. Given the generic nature of our sparse attention, the approach should be applicable to many other tasks like program synthesis and long form open domain question answering. We have open sourced the code for both ETC (github) and BigBird (github), both of which run efficiently for long sequences on both GPUs and TPUs.

Acknowledgements
This research resulted as a collaboration with Amr Ahmed, Joshua Ainslie, Chris Alberti, Vaclav Cvicek, Avinava Dubey, Zachary Fisher, Guru Guruganesh, Santiago Ontañón, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, Li Yang, Manzil Zaheer, who co-authored EMNLP and NeurIPS papers.

Source: Google AI Blog


Constructing Transformers For Longer Sequences with Sparse Attention Methods

Natural language processing (NLP) models based on Transformers, such as BERT, RoBERTa, T5, or GPT3, are successful for a wide variety of tasks and a mainstay of modern NLP research. The versatility and robustness of Transformers are the primary drivers behind their wide-scale adoption, leading them to be easily adapted for a diverse range of sequence-based tasks — as a seq2seq model for translation, summarization, generation, and others, or as a standalone encoder for sentiment analysis, POS tagging, machine reading comprehension, etc. The key innovation in Transformers is the introduction of a self-attention mechanism, which computes similarity scores for all pairs of positions in an input sequence, and can be evaluated in parallel for each token of the input sequence, avoiding the sequential dependency of recurrent neural networks, and enabling Transformers to vastly outperform previous sequence models like LSTM.

A limitation of existing Transformer models and their derivatives, however, is that the full self-attention mechanism has computational and memory requirements that are quadratic with the input sequence length. With commonly available current hardware and model sizes, this typically limits the input sequence to roughly 512 tokens, and prevents Transformers from being directly applicable to tasks that require larger context, like question answering, document summarization or genome fragment classification. Two natural questions arise: 1) Can we achieve the empirical benefits of quadratic full Transformers using sparse models with computational and memory requirements that scale linearly with the input sequence length? 2) Is it possible to show theoretically that these linear Transformers preserve the expressivity and flexibility of the quadratic full Transformers?

We address both of these questions in a recent pair of papers. In “ETC: Encoding Long and Structured Inputs in Transformers”, presented at EMNLP 2020, we present the Extended Transformer Construction (ETC), which is a novel method for sparse attention, in which one uses structural information to limit the number of computed pairs of similarity scores. This reduces the quadratic dependency on input length to linear and yields strong empirical results in the NLP domain. Then, in “Big Bird: Transformers for Longer Sequences”, presented at NeurIPS 2020, we introduce another sparse attention method, called BigBird that extends ETC to more generic scenarios where prerequisite domain knowledge about structure present in the source data may be unavailable. Moreover, we also show that theoretically our proposed sparse attention mechanism preserves the expressivity and flexibility of the quadratic full Transformers. Our proposed methods achieve a new state of the art on challenging long-sequence tasks, including question answering, document summarization and genome fragment classification.

Attention as a Graph
The attention module used in Transformer models computes similarity scores for all pairs of positions in an input sequence. It is useful to think of the attention mechanism as a directed graph, with tokens represented by nodes and the similarity score computed between a pair of tokens represented by an edge. In this view, the full attention model is a complete graph. The core idea behind our approach is to carefully design sparse graphs, such that one only computes a linear number of similarity scores.

Full attention can be viewed as a complete graph.

Extended Transformer Construction (ETC)
On NLP tasks that require long and structured inputs, we propose a structured sparse attention mechanism, which we call Extended Transformer Construction (ETC). To achieve structured sparsification of self attention, we developed the global-local attention mechanism. Here the input to the Transformer is split into two parts: a global input where tokens have unrestricted attention, and a long input where tokens can only attend to either the global input or to a local neighborhood. This achieves linear scaling of attention, which allows ETC to significantly scale input length.

In order to further exploit the structure of long documents, ETC combines additional ideas: representing the positional information of the tokens in a relative way, rather than using their absolute position in the sequence; using an additional training objective beyond the usual masked language model (MLM) used in models like BERT; and flexible masking of tokens to control which tokens can attend to which other tokens. For example, given a long selection of text, a global token is applied to each sentence, which connects to all tokens within the sentence, and a global token is also applied to each paragraph, which connects to all tokens within the same paragraph.

An example of document structure based sparse attention of ETC model. The global variables are denoted by C (in blue) for paragraph, S (yellow) for sentence while the local variables are denoted by X (grey) for tokens corresponding to the long input.

With this approach, we report state-of-the-art results in five challenging NLP datasets requiring long or structured inputs: TriviaQA, Natural Questions (NQ), HotpotQA, WikiHop, and OpenKP.

Test set result on Question Answering. For both verified TriviaQA and WikiHop, using ETC achieved a new state of the art.

BigBird
Extending the work of ETC, we propose BigBird — a sparse attention mechanism that is also linear in the number of tokens and is a generic replacement for the attention mechanism used in Transformers. In contrast to ETC, BigBird doesn’t require any prerequisite knowledge about structure present in the source data. Sparse attention in the BigBird model consists of three main parts:

  • A set of global tokens attending to all parts of the input sequence
  • All tokens attending to a set of local neighboring tokens
  • All tokens attending to a set of random tokens
BigBird sparse attention can be seen as adding few global tokens on Watts-Strogatz graph.

In the BigBird paper, we explain why sparse attention is sufficient to approximate quadratic attention, partially explaining why ETC was successful. A crucial observation is that there is an inherent tension between how few similarity scores one computes and the flow of information between different nodes (i.e., the ability of one token to influence each other). Global tokens serve as a conduit for information flow and we prove that sparse attention mechanisms with global tokens can be as powerful as the full attention model. In particular, we show that BigBird is as expressive as the original Transformer, is computationally universal (following the work of Yun et al. and Perez et al.), and is a universal approximator of continuous functions. Furthermore, our proof suggests that the use of random graphs can further help ease the flow of information — motivating the use of the random attention component.

This design scales to much longer sequence lengths for both structured and unstructured tasks. Further scaling can be achieved by using gradient checkpointing by trading off training time for sequence length. This lets us extend our efficient sparse transformers to include generative tasks that require an encoder and a decoder, such as long document summarization, on which we achieve a new state of the art.

Summarization ROUGE score for long documents. Both for BigPatent and ArXiv datasets, we achieve a new state of the art result.

Moreover, the fact that BigBird is a generic replacement also allows it to be extended to new domains without pre-existing domain knowledge. In particular, we introduce a novel application of Transformer-based models where long contexts are beneficial — extracting contextual representations of genomic sequences (DNA). With longer masked language model pre-training, BigBird achieves state-of-the-art performance on downstream tasks, such as promoter-region prediction and chromatin profile prediction.

On multiple genomics tasks, such as promoter region prediction (PRP), chromatin-profile prediction including transcription factors (TF), histone-mark (HM) and DNase I hypersensitive (DHS) detection, we outperform baselines. Moreover our results show that Transformer models can be applied to multiple genomics tasks that are currently underexplored.

Main Implementation Idea
One of the main impediments to the large scale adoption of sparse attention is the fact that sparse operations are quite inefficient in modern hardware. Behind both ETC and BigBird, one of our key innovations is to make an efficient implementation of the sparse attention mechanism. As modern hardware accelerators like GPUs and TPUs excel using coalesced memory operations, which load blocks of contiguous bytes at once, it is not efficient to have small sporadic look-ups caused by a sliding window (for local attention) or random element queries (random attention). Instead we transform the sparse local and random attention into dense tensor operations to take full advantage of modern single instruction, multiple data (SIMD) hardware.

To do this, we first “blockify” the attention mechanism to better leverage GPUs/TPUs, which are designed to operate on blocks. Then we convert the sparse attention mechanism computation into a dense tensor product through a series of simple matrix operations such as reshape, roll, and gather, as illustrated in the animation below.

Illustration of how sparse window attention is efficiently computed using roll and reshape, and without small sporadic look-ups.

Recently, “Long Range Arena: A Benchmark for Efficient Transformers“ provided a benchmark of six tasks that require longer context, and performed experiments to benchmark all existing long range transformers. The results show that the BigBird model, unlike its counterparts, clearly reduces memory consumption without sacrificing performance.

Conclusion
We show that carefully designed sparse attention can be as expressive and flexible as the original full attention model. Along with theoretical guarantees, we provide a very efficient implementation which allows us to scale to much longer inputs. As a consequence, we achieve state-of-the-art results for question answering, document summarization and genome fragment classification. Given the generic nature of our sparse attention, the approach should be applicable to many other tasks like program synthesis and long form open domain question answering. We have open sourced the code for both ETC (github) and BigBird (github), both of which run efficiently for long sequences on both GPUs and TPUs.

Acknowledgements
This research resulted as a collaboration with Amr Ahmed, Joshua Ainslie, Chris Alberti, Vaclav Cvicek, Avinava Dubey, Zachary Fisher, Guru Guruganesh, Santiago Ontañón, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, Li Yang, Manzil Zaheer, who co-authored EMNLP and NeurIPS papers.

Source: Google AI Blog