Physical problem-solving in minds and machines @SimonsInstituteTOC

Simons Institute | Physical problem-solving in minds and machines @SimonsInstituteTOC | Uploaded 3 months ago | Updated 1 hour ago
Kelsey Allen (DeepMind)
https://simons.berkeley.edu/talks/kelsey-allen-deepmind-2024-06-05
Understanding Lower-Level Intelligence from AI, Psychology, and Neuroscience Perspectives

The world is structured in countless ways. When cognitive and machine models respect these structures, by factorizing their modules and parameters, they can achieve remarkable efficiency and generalization. In this talk, I will discuss my work investigating the factorizations of objects, relations, and physics to support flexible physical problem-solving in both minds and machines. My research suggests that these ingredients can explain complex cognitive phenomena such as how people effortlessly learn to use new tools, and complex behaviours in machines such as highly realistic simulation and tool innovation. By taking better advantage of problem structure, and combining it with general-purpose methods for statistical learning, we can develop more robust and data-efficient machine agents while also better explaining how natural intelligence learns so much from so little.

New models of games with imperfect information

Ramsey Quantifiers in First-Order Logic: Complexity and Applications to Verification

$Towards Practical Distribution Testing Yash Pote (National University of Singapore) https://simons.berkeley.edu/talks/yash-pote-national-university-singapore-2024-08-07 Workshop on Local Algorithms (WoLA) As systems that employ samplers are deployed in safety-critical software, there is a need for tests that can verify the samplers statistical correctness. This raises the question: given a sampler P and a target distribution Q, can we practically test whether P samples from a distribution close to Q? In the high-dimensional setting, where the domain is {0,1}^N for a large N, black-box testing (sample access) is well-known to be intractable; hence, richer grey-box models, such as conditional sampling, have emerged as promising alternatives for the testing problem. In this talk, I will present our work in developing grey-box algorithms that are fast in theory and practice, and I will focus on the first polynomial query algorithm for TV distance estimation, in the conditional sampling model.$

Through a glass, darkly: Approximations, hacks, and workarounds in intuitive physics and imagination

Stochastic games with neural perception mechanisms

Symbolic Finite- and Infinite-state Synthesis

Robot learning, with inspiration from child development

Synthesizing distributed protocols from global session types

An overview of classical robust statistics and generalizations to the future

Overview of Statistical Learning Theory Part 2