A Fiber Criterion for Representation Identifiability in Supervised Learning
A new paper introduces a "fiber criterion" to address the challenge of representation identifiability in supervised learning. This criterion helps determine when specific properties of a learned representation can be reliably inferred from the model's input-output behavior. The research highlights that claims about representations require assumptions beyond just predictive performance, using examples like Waterbirds models to illustrate how different constraints can lead to distinct representations with similar outcomes. AI
IMPACT Clarifies theoretical underpinnings of representation learning, potentially guiding future model interpretability and safety research.