The Neural Information Processing Systems (NeurIPS) conference is held every year in the month of December. This year, the 32nd edition was held in Vancouver, Canada. The purpose behind this conference is to foster the exchange of research on neural information processing systems in their biological, technological, mathematical, and theoretical aspects.
The conference unveiled the NeurIPS outstanding paper awards for this year. These papers are the most notable accepted papers at the conference. The committee picked up the papers based on some specific criteria from the set of papers which had been selected for verbal presentation. The criteria are:
- Potential to endure
- Insight
- Creativity
- Revolutionary
- Rigour
- Elegance
- Reproducible
- Scientific
This time, the committee also included an additional award called Outstanding New Directions Paper Award in order to highlight work that distinguished itself in setting a novel avenue for future research.
In this article, we list down the outstanding papers awarded at NeurIPS 2019.
1| Distribution-Independent PAC Learning of Halfspaces with Massart Noise
Category: Outstanding Paper Award
About: This paper studies the problem of distribution-independent PAC learning of halfspaces or Linear Threshold Functions (LTFs) for binary classification in the presence of Massart noise. The main contribution of this paper is the first non-trivial learning algorithm for the class of halfspaces (or even disjunctions) in the distribution-free PAC model with Massart noise.
Read the paper here.
2| Uniform Convergence May Be Unable To Explain Generalization In Deep Learning
Category: Outstanding New Directions Paper Award
About: This paper presents examples of overparameterized linear classifiers and neural networks trained by gradient descent (GD) where uniform convergence probably cannot “explain generalisation.” With this research, the researchers tried to understand the goal of a small generalisation bound which shows appropriate dependence on the sample size, width, depth, label noise, and batch size.
Read the paper here.
3| Nonparametric Density Estimation & Convergence Rates for GANs under Besov IPM Losses
Category: Honorable Mention Outstanding Paper Award
About: In this paper, the researchers study the problem of estimating a non-parametric probability density under a large family of losses called Besov IPMs. The paper also shows that the linear distribution estimates, such as the empirical distribution or kernel density estimator, often fail to converge at the optimal rate. Furthermore, the researchers also showed that GANs can strictly outperform the best linear estimator.
Read the paper here.
4| Fast And Accurate Least-Mean-Squares Solvers
Category: Honorable Mention Outstanding Paper Award
About: Least-mean squares (LMS) solvers such as Linear / Ridge / Lasso-Regression, SVD and Elastic-Net not only solve fundamental machine learning problems but are also the building blocks in a variety of other methods, such as decision trees and matrix factorizations. Keeping this in mind, the researchers presented a novel framework which shows how to reduce the computational complexity of Least Mean-Square solvers by one or two orders of magnitude, with no precision loss and improved numerical stability.
Read the paper here.
5| Putting An End to End-to-End: Gradient-Isolated Learning Of Representations
Category: Honorable Mention Outstanding New Directions Paper Award
About: In this paper, the researchers proposed a novel deep learning method for local self-supervised representation learning which does not require labels nor end-to-end backpropagation but exploits the natural order in data instead. The research is done by splitting a deep neural network into a stack of gradient-isolated modules where each module is trained to maximally preserve the information of its inputs using the InfoNCE bound.
Read the paper here.
6| Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations
Category: Honorable Mention Outstanding New Directions Paper Award
About: The researchers at Stanford University proposed Scene Representation Networks (SRNs) which is a continuous, 3D structure-aware scene representation that encodes both geometry and appearance. SRNs represent scenes as continuous functions which has the ability to map world coordinates to a feature representation of local scene properties. The potential of SRNs is further demonstrated by evaluating them for novel view synthesis, few-shot reconstruction, joint shape, and appearance interpolation, and unsupervised discovery of a non-rigid face mode.
Read the paper here.