Phase Transitions for Feature Learning in Neural Networks

Andrea Montanari, Zihao Wang

According to a popular viewpoint, neural networks learn from data by first identifying low-dimensional representations, and subsequently fitting the best model in this space. Recen…