alexxnica / deepmath

Experiments towards neural network theorem proving

Deepmath

The Deepmath project seeks to improve automated theorem proving using deep learning and other machine learning techniques. Deepmath is a collaboration between Google Research and several universities.

DISCLAIMER:

The source code in this repository is not an official Google product, but is a research collaboration with external research teams.

For now the repository contains only a C++ implementation of the HOL Light kernel, which we have released early in order to faciliate existing collaborations. More to come soon, including neural network models.

Directories

• hol: A C++ implementation of the HOL Light kernel.
• zz: A reimplementation of the HOL Light kernel (with visualizer and other stuff).

Installation

Deepmath depends on TensorFlow, which is included as a submodule. To use, first install all TensorFlow dependencies and run configure to initialize the build system, as described here:

# Install TensorFlow dependencies if necessary
cd deepmath/tensorflow
./configure  # and answer all questions

You can then build hol via

cd deepmath/hol
bazel build ...