### Current browse context:

math.AG

### Change to browse by:

### References & Citations

# Mathematics > Algebraic Geometry

# Title: Complex Critical Points of Deep Linear Neural Networks

(Submitted on 30 Jan 2023)

Abstract: We extend the work of Mehta, Chen, Tang, and Hauenstein on computing the complex critical points of the loss function of deep linear neutral networks when the activation function is the identity function. For networks with a single hidden layer trained on a single data point we give an improved bound on the number of complex critical points of the loss function. We show that for any number of hidden layers complex critical points with zero coordinates arise in certain patterns which we completely classify for networks with one hidden layer. We report our results of computational experiments with varying network architectures defining small deep linear networks using HomotopyContinuation.jl.

Link back to: arXiv, form interface, contact.