Two Studies on End-to-End Learning Methods for Instruction Following
A key deliverable for artificial intelligence is enabling natural language interface for human-computer interaction. The deep learning approach to this problem is to find a way to provide abundant data to a powerful learning system. In my talk, I will cover two projects that address two questions that are very important in this context: which data, and how much of it would be required to train a deep-learning-based instruction following agent? First, I will talk about AGILE, a method to train an agent to follow instructions using restricted supervision in the form of (instruction, goal-state) pairs. Second, I will present BabyAI, a framework that we have recently built at Mila to support research on grounded instruction following, and in particular, on the sample efficiency of various approaches to this task.
Dzmitry (Dima) Bahdanau is currently a 4th year PhD student at Mila, Université de Montréal, working under the supervision of Yoshua Bengio and Aaron Courville. Prior to starting his PhD, Dzmitry studied applied mathematics at Belarusian State University and did his MSc in computer science at Jacobs University Bremen with Herbert Jaeger. Dzmitry has also interned at Google Brain with Quoc Le and at DeepMind with Edward Grefenstette.