Neural networks under SGD

This lecture covers the concept of neural networks trained using Stochastic Gradient Descent (SGD). It explains the relationship between the number of neurons in hidden layers and the number of samples, as well as the parameters involved in the network. The lecture delves into the square loss function and the process of taking small steps to optimize the network. It also discusses the evolution of particles in the network, the interpretation of true risk, and the implications of replacing particles with density. The lecture concludes with an exploration of limits and justifications in the context of neural network training.