Structures of Neural Network Effective Theories

This work provides insights into neural network behavior at initialization, potentially aiding progress in deep learning and field theory simulations, but it is incremental as it builds on existing effective field theory frameworks.

The paper tackled the problem of computing finite-width corrections to neuron statistics in deep neural networks at initialization by developing a diagrammatic approach to effective field theories, revealing that a single condition governs the criticality of all connected correlators of neuron preactivations.