Welcome to torch-foresight’s documentation!

This package provides a collection of modules useful for characterizing and predicting the dynamics and performance of neural nets. These consist mostly of novel metrics, derived from fields like theoretical neuroscience and information theory, aimed at helping researchers to better understand how neural networks work. The repository is meant to advance a new “Science of AI” or “Science of Deep Learning” (see neuralnet.science). It currently includes modules for computing:

• Effective Information (paper)

With the following under development:

• Gradient Noise Scale (paper)
• Information Bottleneck (paper)

Modules

• Effective Information
  • Theory
  • Code Documentation
• Gradient Noise Scale
  • Theory
  • Code Documentation

Indices and tables

• Index
• Module Index
• Search Page