CommAI: Taking the first small steps towards general intelligence

Abstract: It is by now clear that statistical algorithms can successfully extract knowledge from large amounts of data (distributed models of word meaning, object classification in images, etc.). However, one of the core characteristics of human intelligence is the ability to flexibly adapt to new tasks after seeing very little relevant data, or by just relying on a linguistic description of the intended goal. I will introduce FAIR's CommAI initiative. CommAI defines a set of intuitively simple character manipulation tasks whose solution requires fast adaption and learning through linguistic interaction. We conjecture that such tasks are beyond the scope of current state-of-the-art computational models, and we hope their availability will spur research into more powerful methods displaying human-like learning-to-learn capabilities.

Collaborators: Rahma Chaabouni, Germán Kruszewski, Allan Jabri, Armand Joulin, Klemen Simonic and Tomas Mikolov.