astroautomata/SymTorch

Deep Learning Interpretability with Symbolic Regression

What it solves

SymTorch addresses the "black box" nature of deep learning models by allowing users to approximate the behavior of specific model components using symbolic equations. This process, known as symbolic distillation, helps make the complex internal workings of neural networks more interpretable.

How it works

The framework leverages PySR (Symbolic Regression) to find mathematical expressions that mimic the behavior of components within deep learning models.

Who it’s for

Researchers and developers who need to understand the mathematical logic behind deep learning model components or those looking to perform symbolic distillation of neural networks.

Highlights

  • Symbolic Approximation: Uses PySR to turn deep learning components into readable symbolic equations.
  • Symbolic Distillation: Provides a framework specifically designed for the distilling of deep neural networks into symbolic forms.
  • PyTorch Integration: Available as a Python package (torch-symbolic) for use within the PyTorch ecosystem.

Related

  • Project
    astroautomata/PySR

    PySR is a high-performance symbolic regression tool that searches for interpretable mathematical expressions to fit datasets or distill neural networks into analytic equations.

  • Project
    pytorch/pytorch

    A Python-based machine learning framework that provides GPU-accelerated tensor computation and a dynamic autograd system for building deep neural networks.

  • Project
    catalyst-team/catalyst

    A PyTorch framework for deep learning R&D that removes training loop boilerplate to accelerate experimentation and reproducibility.

  • Project
    alibaba/TorchEasyRec

    A PyTorch-based recommendation system framework that provides a library of production-ready deep learning models for matching, ranking, and generative recommendation.

  • Project
    tum-pbs/PhiFlow

    A differentiable simulation toolkit for PyTorch, Jax, and TensorFlow that enables the integration of physics simulations into machine learning optimization loops.