Google At NeurIPS 2021: Gets 177 Papers Accepted

The 35th edition of the Neural Information Processing Systems conference 2021 (NeurIPS 2021) commenced on December 6, 2021. The nine day conference is packed with a series of tutorials, workshops, and presentations. Over 9,000 papers were submitted at the conference this year, of which 2,344 papers were accepted; this the highest number of papers accepted since 2013. 

The annual NeurIPS conference is the most awaited and well attended machine learning events of the year. Leading companies and academic institutions like Google, Microsoft, Meta, DeepMind, Stanford, and Carnegie Mellon University participate in great number. Among tech companies, Google had the highest number of accepted papers – 177.

Let us look at some of the winning papers from Google at NeurIPS  2021. 

Deep Reinforcement Learning at the Edge of the Statistical Precipice (Outstanding paper award)

By Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, and Marc G. Bellemare.

This paper outlines practical approaches to improve the rigour of deep reinforcement learning algorithm evaluation. The focus is especially the evaluation of new algorithms that should provide stratified bootstrap confidence intervals, performance profiles across tasks and runs, and interquartile means. The researchers also show that standard approaches for reporting results in deep RL across many tasks and multiple runs can make it difficult to assess if a new algorithm represents a consistent advancements over past methods. The performance summaries are designed to be able to compute with a small number of runs per task.

Read the full paper here.

Do Vision Transformers See Like Convolutional Neural Networks?

By Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, Alexey Dosovitskiy

Convolutional neural networks (CNNs) have been the preferred model for visual data, but recently it has been shown that vision transformer models (ViT) can achieve comparable or even superior performance on image classification tasks. This paper analyses the internal representation structure of ViTs and CNNs on image classification benchmarks. The team found out the differences between the two architectures ( eg: ViT having more uniform representations across all layers). The research also looks at important roles played by self-attention, the ramifications for spatial localization, and the effect of (pretraining) dataset scale on intermediate features and transfer learning.

Read the full paper here.

The Difficulty of Passive Learning in Deep Reinforcement Learning

By George Ostrovski, Pablo Samuel Castro, Will Dabney

Learning to act from observational data without active environmental interaction in reinforcement learning is a challenge. Approached like constraints on the learned policy, preventing strong deviations from the state-action distribution of the dataset are being used to overcome this challenge. But, these methods are evaluated using non-linear function approximation. This paper proposes “tandem learning” experimental paradigm which facilitates the empirical analysis of the difficulties in offline reinforcement learning. The team has also identified approximation in conjunction with fixed data distributions as the strongest factor. This observation challenges the hypothesis that is stated in past work. The team offers insights for offline deep reinforcement learning and sheds light on phenomena observed in online case of learning control.

Read the full paper here.

Datasets and Benchmarks Accepted Papers

NeurIPS launched the new Datasets & Benchmarks track this year. It went to the paper titled, “Reduced, Reused and Recycled: The Life of a Dataset in Machine Learning Research” by Bernard Koch, Emily Denton, Alex Hanna, Jacob G. Foster.

Benchmark datasets play an important role in machine learning research – they serve as a measure of progress towards the goal. But not much attention has been given to the dynamics of benchmark dataset use and reuse, within machine learning communities.

This paper studies how dataset usage patterns differed across machine learning subcommunities between 2015 and 2020. They found increasing concentration on fewer and fewer datasets within task communities, significant adoption of datasets from other tasks, as well as concentration across the field on datasets that have been introduced by researchers situated within a small number of elite institutions. These results have an impact on AI ethics, and equity and access within the field.

Read the full paper here.

 Test of Time Award

 The Test of Time Award went to Online Learning for Latent Dirichlet Allocation by Matthew Hoffman (senior scientist at Google), David Blei, and Francis Bach.

The research team developed a variational Bayes (VB) algorithm for Latent Dirichlet AI -location (LDA). It is based on online stochastic optimization with a natural gradient step and converges to a local optimum of the VB objective function. It comes with the capability of analysing large amounts of document collections, including those arriving in a stream.The paper studies the performance of online LDA by fitting a 100-topic topic model to 3.3 million articles from Wikipedia in a single pass.They show that online LDA finds topic models as good or better than those found with batch VB and does this in a fraction of the time.

‘Read the full paper here.