<ul><li>This paper examines the theoretical boundaries of efficient learnability in multi-index models.</li><li>The focus is on the minimum sample complexity required for weakly recovering their low-dimensional structure.</li><li>The findings uncover conditions for learning trivial subspaces, easy subspaces, and interactions between different directions.</li><li>The theory builds on the optimality of approximate message-passing among first-order iterative methods.</li></ul>

Fundamental computational limits of weak learnability in high-dimensional multi-index models

Discover more