The NeurIPS 2021 has started from December 6 and will go on till December 14 packed with tutorials, conferences and workshops. Meanwhile, NeurIPS 2021 has announced the recipients of the 2021 Outstanding Paper Awards, the Test of Time Award, and the new Datasets and Benchmarks Track Best Paper Awards. 

Let us take a more detailed look at them and understand what made them stand out.

A Universal Law of Robustness via Isoperimetry

By Sébastien Bubeck and Mark Sellke

The researchers said that data interpolation with a parametrised model class is possible if the number of parameters is more than the number of equations to be satisfied. In deep learning, models are trained with many more parameters than what this classical theory would suggest. In this paper, the researchers suggest a theoretical explanation for this that proves for a broad class of data distributions and model classes, over parameterisation is necessary. This is required if one wants to interpolate the data smoothly. They show that smooth interpolation requires d times more parameters than just interpolation (d is the ambient data dimension). The team also shows the universal law of robustness for any smoothly parameterised function class with polynomial-size weights and any covariate distribution verifying isoperimetry.

Read the full paper here. 

On the Expressivity of Markov Reward

By David Abel, Will Dabney, Anna Harutyunyan, Mark K. Ho, Michael Littman, Doina Precup, and Satinder Singh.

The researchers have studied the expressivity of Markov reward functions in finite environments through inspection of what kinds of tasks these functions can express. The paper looks into understanding the expressivity of reward as a way to capture tasks that we would want an agent to perform. The team bases the workaround three new abstract notions of “task” that might be desirable – a set of acceptable behaviours, a partial ordering over behaviours, or a partial ordering over trajectories. 

The results show that though reward can express many of the tasks, yet there exist cases of each task type that no Markov reward function can capture. The researchers also provide a set of polynomial-time algorithms that build a Markov reward function which allows an agent to optimise tasks of each of these three types. It determines correctly no such reward function exists. 

Read the full paper here. 

Deep Reinforcement Learning at the Edge of the Statistical Precipice

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

This paper looks into practical approaches to improve the rigour of deep reinforcement learning algorithm comparison. It looks into 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 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.

MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers

By Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers, John Thickstun, Sean Welleck, Yejin Choi, and Zaid Harchaoui.

The researchers said that measuring how close machine-generated text is to human language is an important issue. Mauve is a comparison measure for open-ended text generation. It compares the learnt distribution from a text generation model to the distribution of human-written text using divergence frontiers. The team said that Mauve identifies known properties of generated text and scales naturally with model size. It correlates with human judgments, with fewer restrictions than existing evaluation metrics.

Read the full paper here. 

 Continuized Accelerations of Deterministic and Stochastic Gradient Descents, and of Gossip Algorithms

By Mathieu Even, Raphaël Berthier, Francis Bach, Nicolas Flammarion, Pierre Gaillard, Hadrien Hendrikx, Laurent Massoulié, and Adrien Taylor.

The paper introduces the “continuized” Nesterov acceleration in which the two separate vector variables evolve jointly in continuous time. It uses the best of the continuous and the discrete frameworks: as a continuous process. They show that discretisation has the same structure as Nesterov acceleration but with random parameters. The team provides continuized Nesterov acceleration under deterministic and stochastic gradients, with either additive or multiplicative noise. In the end, they provide the first rigorous acceleration of asynchronous gossip algorithms by using their continuized framework and expressing the gossip averaging problem as the stochastic minimisation of a certain energy function.

Read the full paper here. 

 Moser Flow: Divergence-based Generative Modeling on Manifolds

By Noam Rozen, Aditya Grover, Maximilian Nickel, and Yaron Lipman.

The paper introduces Moser Flow (MF) which is a class of generative models within the family of continuous normalising flows (CNF). MF also produces a CNF through a solution to the change-of-variable formula. Its model (learned) density is parameterised as the source (prior) density minus the divergence of a neural network. The divergence is a local, linear differential operator, easy to approximate and calculate on manifolds. It does not require invoking or backpropagating through an ODE solver during training. They demonstrate the use of flow models for sampling from general curved surfaces and have achieved significant improvements in density estimation, sample quality, and training complexity.

Read the full paper here.

 Datasets & Benchmarks Best Paper Awards

NeurIPS launched the new Datasets & Benchmarks track this year. The award recipients for this are:

Reduced, Reused and Recycled: The Life of a Dataset in Machine Learning Research ( by Bernard Koch, Emily Denton, Alex Hanna, and Jacob Gates Foster and ATOM3D: Tasks on Molecules in Three Dimensions (by Raphael John Lamarre Townshend, Martin Vögele, Patricia Adriana Suriana, Alexander Derry, Alexander Powers, Yianni Laloudakis, Sidhika Balachandar, Bowen Jing, Brandon M. Anderson, Stephan Eismann, Risi Kondor, Russ Altman, and Ron O. Dror). Read the full papers here-Reduced, Reused and Recycled: The Life of a Dataset in Machine Learning Research and ATOM3D

Test of Time Award

The Test of Time Award went to Online Learning for Latent Dirichlet Allocation by Matthew Hoffman, David Blei, and Francis Bach.

The research team developed a variational Bayes (VB) algorithm for Latent Dirichlet Allocation (LDA). It is based on online stochastic optimisation 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.