Universal Algorithmic Intelligence

Marta Kryven will present on approximate Bayesian inference for problem-solving domains at the UAI research meeting tomorrow, Monday February 23rd at 3 pm EST. We will use the regular zoom link.

Abstract:

Bayesian Inference and Optimal Experiment Design provide a general case optimal theoretical framework for how to build world models and interpret evidence. However, this framework is intractable for nearly all practical applications, as it assumes that the hypotheses space must be enumerable. While in many cases Bayesian Inference can be solved approximately – the key question remains: how to effectively enumerate hypotheses in contexts where this space is potentially infinite? And how to prioritize hypotheses that are causally interpretable — i.e. not only provide a mapping from evidence to outcomes, but can inform solving other similar tasks?
I will review a test-case where we use Approximate Bayesian Inference to interpret interaction in a puzzle-solving domain, and present several ways how hypothesis generation can be formalized in a generalizable and interpretable form.

Bio:

Marta Kryven is an assistant professor at Dalhousie University. She holds a PhD in computer science from the University of Waterloo, and her postdoctoral studies where in cognitive psychology at MIT. She now works at the intersection of AI and experimental psychology.

Recently, we have worked together on generative modeling for efficient (human-like) planning. I am excited to hear about her recent work, and hope to see you all there!