CTN: Brenden Lake

Title: Meta-learning for more powerful behavioral modeling

Abstract: Two modeling paradigms have historically been in tension: Bayesian models provide an elegant way to incorporate prior knowledge, but they make simplifying and constraining assumptions; on the other hand, neural networks provide great modeling flexibility, but they make it difficult to incorporate prior knowledge. Here I describe how to get the best of both approaches through Behaviorally-Informed Meta-Learning (BIML). BIML allows for modeling behavior with flexible Transformers, even with only minimal data, by distilling Bayesian priors into neural networks and then further fine-tuning the networks on behavioral data. I’ll show some initial successes using BIML to model human concept learning, resulting in superior fits by capturing behavioral heuristics and biases that violate simple Bayesian assumptions. At the end, I would love to discuss how to overcome the challenges of interpreting this new class of models.

Zoom: https://columbiauniversity.zoom.us/j/93740145362?pwd=GgoanUbc3Kc4rWdux2doLOiciiAaO2.1
meeting ID: 937 4014 5362
passcode: ctn