Language evolution seminar

Speaker: Aaron Courville (Université de Montréal)

Title: Applications of Iterated Learning in Deep Neural Networks

Abstract: In recent years, deep learning has revolutionized our approach to AI and related tasks. There has been tremendous strides in performance on a wide range of domains such as speech recognition, translation, object recognition and reinforcement learning. Yet despite this impressive progress, challenges remain in making our neural network-based predictive models robust to data distribution shifts. Our models simply do not possess the sort of generalization robustness of which humans are routinely capable.

Iterative Learning is a theory of how compositional human language emerged. Surprisingly,  the same process seems to support the emergence of compositional representations in training neural networks. In this talk I will be providing an overview of my group's attempts to apply Iterative Learning to promote better and more robust generalization in deep neural networks. I will discuss an application to Visual Question Answering, an attempt to minimize language drift in training dialog AI agents and an application to improve the robustness of object recognition. I will also discuss our explorations of simplifications of the iterated learning procedure and how it can improve performance in training data-efficient RL agents.

Contact

Seminars are organised by the Centre for Language Evolution

• Elizabeth Pankratz

• Centre for Language Evolution