More Accurate Behavioral Predictions with Hybrid Bayesian-Transformer Models

Watch a research lecture from NYU's Brenden Lake exploring the integration of Bayesian models and Transformers through Behaviorally-Informed Meta-Learning (BIML). Discover how BIML combines the prior knowledge incorporation strengths of Bayesian approaches with the flexibility of neural networks to better model human concept learning and behavior. Learn about the methodology of distilling Bayesian priors into neural networks and subsequent fine-tuning on behavioral data, resulting in models that can capture human heuristics and biases beyond simple Bayesian assumptions. Engage with discussions on the interpretability challenges of these hybrid models and their implications for cognitive science, linguistics, and neuroscience research.