Generalization Under Distribution Shift: PhD Thesis Defense - Aviv Netanyahu

Year
2025
Tags
distribution shift
out-of-distribution generalization
covariate shift
label shift
few-shot learning
task inference
inverse modeling
generative models
diffusion models
demonstrations
behavior cloning
in-context learning
task embeddings
T5 embeddings
compositionality
skill composition
long-horizon control
robotics
autonomous driving
tabletop manipulation
motion capture
bilinear transduction
transductive learning
analogy-based prediction
matrix completion
matrix factorization
low-rank structure
coverage assumption
zero-shot extrapolation
pick-and-place
target generalization
3D grasp prediction
SE(3) equivariance
scale shift
molecules
composition-based features
precision-recall
MIT
Summary

This defense addresses generalization under distribution shift with limited data. It introduces (1) diffusion-based inverse task inference that recovers a task embedding from a few demonstrations, enabling compositional generation without fine-tuning; and (2) bilinear transduction that converts out-of-support inputs into out-of-combination problems, yielding zero-shot extrapolation in robotics and property prediction.