Xmaster6y / devinterp

Tools for studying developmental interpretability in neural networks.

DevInterp

A Python Library for Developmental Interpretability Research

DevInterp is a python library for conducting research on developmental interpretability, a novel AI safety research agenda rooted in Singular Learning Theory (SLT). DevInterp proposes tools for detecting, locating, and ultimately controlling the development of structure over training.

Read more about developmental interpretability.

⚠️ This library is still in early development. Don't expect things to work on a first attempt. We are actively working on improving the library and adding new features. If you have any questions or suggestions, please feel free to open an issue or submit a pull request.

Installation

To install devinterp, simply run:

pip install devinterp

Requirements: Python 3.8 or higher.

Getting Started

To see DevInterp in action, check out our example notebooks:

• Introduction
• Normal Crossing Demo
• MNIST Demo
• Deep Linear Networks
• Toy Models of Superposition

Minimal Example

from devinterp.slt import estimate_learning_coeff, estimate_learning_coeff_with_summary
from devinterp.optim import SGLD

# Assuming you have a PyTorch Module and DataLoader
learning_coeff = estimate_learning_coeff(model, trainloader, ...)

# If you want to see mean, std, and learning coeff estimate per chain
learning_coeff_summary = estimate_learning_coeff_with_summary(model, trainloader, ...)

Features

• Estimate the learning coefficient.
  • Supported optimizers:
    • SGLD
    • SGNHT

Contributing

See CONTRIBUTING.md for guidelines on how to contribute.