Out of Distribution Generalization

Explore a comprehensive lecture on out-of-distribution generalization in machine learning, delivered by Rajesh Ranganath at the Computational Genomics Summer Institute. Delve into key concepts such as nuisance-induced spurious correlations, invariant risk minimization, and distributionally robust neural networks. Examine real-world examples including cow vs. penguin classification, waterbirds vs. landbirds, and pneumonia detection. Investigate techniques like nuisance randomization and uncorrelating representations to address common issues in machine learning models. Analyze the challenges of natural language inference and explore various coupling assumptions. Gain insights into the role of causality in machine learning and the trade-offs between flexibility and estimation costs. Access related research papers for further study on out-of-distribution generalization techniques and their applications in computational genomics.

Classifying cows versus penguins
A common issue
Example: cows v. penguins
Idea 1: nuisance randomization
Doesn't work by itself
Idea 2: Uncorrelating representations
What's the best uncorrelating representation?
A simulation
Waterbirds versus landbirds
Pneumonia classification
Related work
What did I mean by local optima?
Did we lose anything?
Natural language inference
Generic setup not solvable
One assumption is test = train
Invariant coupling
Subgroup Coupling
Common factor coupling
What does good mean?
Simultaneous Optimality?
Why does causality appear?
The added flexibility comes with an estimation cost
Some questions
Research References