What makes neural nets good learners? Lessons from high-dimensional asymptotics of two-layer networks

About these Bruno Loureiro lecture notes

These notes ask why neural networks learn so effectively from data through the lens of high-dimensional asymptotics. They move from empirical risk minimization and kernel limits to feature learning in two-layer and quadratic neural networks.

• Empirical risk minimization and generalization
• High-dimensional learning and kernel limitations
• Two-layer networks and feature learning
• Marchenko–Pastur laws and spectral transitions
• Quadratic networks and spiked random matrices