Andrea Montanari, Stanford University, USA

Selected Topics in Machine Learning (7.5h)



1. Principal component analysis

The spiked model

The synchronization problem and its semidefinite programming relaxations

Connection with n-vector spin glasses

2. The hidden clique problem

Optimal local algorithms

Sum of squares hierarchy

The statistic/computation gap

Connection with sparse principal component analysis

Statistical applications

3. Non-negative matrix factorization

Applications to feature learning and topic models

Convex and iterative algorithms

Phase transition under the spiked model

4. Neural networks

Back-propagation algorithm

Approximation theory

Convolutional networks

Connections with statistical physics