Dr. Kfir Dolev

Kfir Dolev grew up in Israel and moved to California with his family when he was 10. He began researching particle dark matter in college, using the Standard Model of particle physics, but eventually became aware of its limitations in describing complex large-scale phenomena. While working on his PhD in Physics at Stanford University, he became interested in how the laws of physics—particularly quantum mechanics and gravity—place fundamental limits on computation. Convinced that the greatest advances in science in the coming decades would be made by artificial rather than human intelligence, Dr. Dolev remained passionate about physics, but felt that the best way to advance physics, and science in general, was by developing artificial intelligence.

For his postdoc in Tel Aviv University’s Department of Computer Science, he works on solving the problem of using deep learning systems in low-data settings.
Training deep learning systems to exhibit intelligent behavior entirely through exploratory reinforcement learning has worked well in narrow domains with clear rules. The most famous example of this is AlphaGo Zero, which attains super-human skills at the game of Go entirely through self-play.

Dr. Dolev wants to expand such learning to the general setting of future time series data prediction by building models capable of deducing underlying computational processes. His approach involves a systematic exploration of the space of programs to find the ones which generate the most generally useful training data. If successful, his system would be a first-of-its-kind reasoning engine that builds its own curriculum. Such an engine could generate and pursue its own structured questions, growing ad-infinitum and potentially helping to break through blocked areas of research by noting unexpected patterns or regularities that humans might miss.