# The Loss Surface of Multilayer Networks

**Authors:** Anna Choromanska, Mikael Henaff, Michael Mathieu, Gerard Ben Arous, Yann LeCun  
**arXiv:** [1412.0233](https://arxiv.org/abs/1412.0233)  
**Date:** 2014-11-30  
**Summary page:** https://www.summarizepaper.com/en/paper/1412.0233v1/

## Key points

- Study titled "The Loss Surface of Multilayer Networks" by Choromanska et al.
- Relationship between non-convex loss function and Hamiltonian of spherical spin-glass model
- Key assumptions: variable independence, network parametrization redundancy, uniformity
- Insight into fully decoupled neural networks through random matrix theory
- Findings on critical values of random loss function for large-size decoupled networks
- Layered structure of critical values within a narrow band
- Decrease in number of local minima outside the narrow band as network size increases
- Empirical evidence and computer simulations supporting mathematical model accuracy
- Convergence of simulated annealing and SGD algorithms towards band with high concentration of critical points (local minima)
- Probability differences in recovering poor quality local minima between small-size and large-size networks
- Increasing challenge in retrieving global minimum as network size grows, practical irrelevance due to overfitting issues

## Summary

In their study titled "The Loss Surface of Multilayer Networks," authors Anna Choromanska, Mikael Henaff, Michael Mathieu, Gerard Ben Arous, and Yann LeCun explore the complex relationship between the non-convex loss function of a simplistic model of the fully-connected feed-forward neural network and the Hamiltonian of the spherical spin-glass model. By considering variable independence, network parametrization redundancy, and uniformity as key assumptions, the researchers aim to gain insight into the intricacies of fully decoupled neural networks through the lens of random matrix theory. Their findings reveal that for large-size decoupled networks, the lowest critical values of the random loss function are concentrated within a well-defined narrow band that is lower-bounded by the global minimum. Furthermore, these critical values exhibit a layered structure within this band. The study also demonstrates that as network size increases, the number of local minima outside this narrow band decreases exponentially. Through empirical evidence and computer simulations on real networks with high dependencies, Choromanska et al. show that their mathematical model accurately reflects observed behavior. They suggest that both simulated annealing and Stochastic Gradient Descent (SGD) algorithms converge towards the band containing a high concentration of critical points - all identified as local minima associated with high learning quality based on test error metrics. One significant finding highlighted by their research is how small-size networks have a non-zero probability of recovering poor quality local minima compared to large-size networks where such occurrences are rare. Additionally, they establish that retrieving the global minimum becomes increasingly challenging as network size grows and suggest its practical irrelevance due to its tendency to lead to overfitting issues. Overall, this study provides valuable insights into the intricate landscape of loss functions in multilayer networks and offers implications for optimization algorithms and network design strategies in deep learning applications.