Winter / Hiver 2019

ICLR Lightning Talks

Featuring

• Alexia Jolicoeur-Martineau
• Gauthier Gidel
• Sandeep Subramanian
• Anirudh Goyal
• Yikang Shen

Note: Audio Since Minute 10:35

Meta-transfer learning for factorizing representations and knowledge for AI

Whereas machine learning theory has focused on generalization to examples from the same distribution as the training data, better understanding of the transfer scenarios where the observed distribution changes often in the lifetime of the learning agent is important, both for robust deployment and to achieve a more powerful form of generalization which humans seem able to enjoy and which seem necessary for learning agents. Whereas most machine learning algorithms and architectures can be traced back to assumptions about the training distributions, we also need to explore assumptions about how the observed distribution changes. We propose that sparsity of change in distribution, when knowledge is represented appropriately, is a good assumption for this purpose, and we claim that if that assumption is verified and knowledge represented appropriately, it leads to fast adaptation to changes in distribution, and thus that the speed of adaptation to changes in distribution can be used as a meta-objective which can drive the discovery of knowledge representation compatible with that assumption. We illustrate these ideas in causal discovery: is some variable a direct cause of another? and how to map raw data to a representation space where different dimensions correspond to causal variables for which a clear causal relationship exists? We propose a large research program in which this non-stationarity assumption and meta-transfer objective is combined with other closely related assumptions about the world embodied in a world model, such as the consciousness prior (the causal graph is captured by a sparse factor graph) and the assumption that the causal variables are often those agents can act upon (the independently controllable factors prior), both of which should be useful for agents which plan, imagine and try to find explanations for what they observe.

Are Overparameterized Neural Networks Actually Just Linear Models?

Neural networks define a rich and expressive class of functions whose properties and behaviors are very hard to describe from a theoretical perspective. Nevertheless, when these functions become highly overparameterized, a surprisingly simple characterization emerges. In this talk, I will discuss several perspectives on this characterization: 1) I will examine the prior over functions induced by common weight initialization schemes and show that it corresponds to a Gaussian process with a well-defined compositional kernel; 2) I will show that by tuning initialization hyperparameters, this kernel can be optimized for signal propagation, yielding networks that are trainable to enormous depths (10k+ layers); and 3) I will demonstrate that the learning dynamics of such overparameterized neural networks are governed by a linear model obtained from the first-order Taylor expansion of the network around its initial parameters.

Multi-objective training of Generative Adversarial Networks with multiple discriminators

Recent literature has demonstrated promising results for training Generative Adversarial Networks by employing a set of discriminators, in contrast to the traditional game involving one generator against a single adversary. Such methods perform single-objective optimization on some simple consolidation of the losses, e.g. an average. In this work, we revisit the multiple-discriminator setting by framing the simultaneous minimization of losses provided by different models as a multi-objective optimization problem. Specifically, we evaluate the performance of multiple gradient descent and the hypervolume maximization algorithm on a number of different datasets. Moreover, we argue that the previously proposed methods and hypervolume maximization can all be seen as variations of multiple gradient descent in which the update direction can be computed efficiently. Our results indicate that hypervolume maximization presents a better compromise between sample quality and computational cost than previous methods.

BabyAI: A Platform to Study the Sample Efficiency of Grounded Language Learning

This talk will present BabyAI, a research project based at Mila with the long-term goal of creating agents which we can communicate with and teach new concepts using natural language. We will begin by discussing why language learning is a hard problem, one to which deep learning doesn't yet have a satisfying answer for. We will then introduce the BabyAI platform, created to study the sample efficiency of grounded language learning in the context of embodied agents and instruction following (see paper accepted at ICLR19, https://openreview.net/forum?id=rJeXCo0cYX). Lastly, we will discuss multiple promising research directions we have identified with the goal of improving the sample efficiency of grounded language learning.