Generalization insights from actual cognition
summary
Opens by making malicious fun of out-of-distribution and argues generalization is about structure that maps past to future, not tiny distribution shifts. Uses movement experiments and normative models: saccadic gain adaptation with ~1,000 trials shows an overnight reversion (errors return after dark-room overnight), lesion evidence that cerebellum is required, and a human study where two weeks of mouse use turns local generalizers into global ones. Models are Kalman-like with a continuum of time scales to capture predictable drift and long versus short term components.
AI-written summary · about
chapters (24)
- 0:00Introduction
- 1:51Why generalization matters
- 3:40Generalization beyond small shifts
- 6:11Defining generalization
- 8:47Time-aware models
- 10:11Generalization in movement science
- 12:23Cup-moving experiment
- 15:54Saccadic gain adaptation
- 18:42Normative models of generalization
- 21:40Vocabulary learning
- 23:26Global versus local generalization
- 25:12Generalization complexity
- 26:34Time horizons and uncertainty
- 28:02Curiosity and causal inference
- 29:52Critique of beautiful math
- 31:43Embracing ugliness
- 32:32Audience questions on generalization frameworks
- 37:14Predictability and context shifts
- 40:10Human prediction limits
- 43:53Innate versus learned generalization
- 46:33Research directions
- 49:26Mathematical rigor and time
- 51:15Levels of learning
- 53:06Closing remarks
related talks
01h 10mModern paradigms of generalization, the heliocentric model of Aristarchus,...
10mModern Paradigms in Generalization
01h 16mGeneralization in natural language processing
36mThe Limits of Post-Selection Generalization
01h 05mAdaptive Generalization: The Role of Dynamic Evaluation in Low-Resource Settings
41mDistribution shift in ecological data: generalization vs. specialization,
01h 07mDoes generalization imply accuracy on the line? A new look at robust generalization
01h 03mGeneralization from the behavioral perspective