The Arnhold Institute hosted its first session of the Technology and Entrepreneurship Speaker Series on Wednesday, March 22, 2017. Our featured guest was Léon Bottou from Facebook AI Research. Below is the abstract from the talk
Abstract: Artificial intelligence algorithms can identify objects in images, but they must learn on more images than a human can see in a lifetime. Algorithms can translate text, but they must learn from more text than a human can read in a lifetime. Algorithms can learn to play Atari games, but learn by playing more games than any teenager can endure. Although we can pin this inefficiency to some properties of our algorithms, there remains a lot of signal in natural data that algorithms do not exploit. I will report on two works in this direction. The first one establishes that something as simple as a collection of static images contains nontrivial information about the causal relations between the objects they represent. The second one, time permitting, shows how an attempt to discover structure in observational data led to a clear improvement of Generative Adversarial Networks.