In our next session on November 23rd at 3pm – 4pm Eastern time, Kaarel Hänni is scheduled to present his work on circuit priors and generalization in neural net Bayesian learning.
Bio
Kaarel Hänni is a research scientist on the agent foundations team at Resolution, with a background in math and philosophy. He has worked on various conceptual problems related to AI alignment.
Abstract
Kaarel Hänni will present his joint work with Dmitry Vaintrob proving that (even arbitrarily overparametrized) neural network Bayesian learning has an emergent circuit prior, and so good generalization, when learning functions with small circuits (for certain reasonable priors and with some caveats).
Leave a Reply