NN2MV: extracting formulae in many-valued logic from ReLU networks

The NN2MV package extracts formulae in many-valued logic and its extensions from ReLU networks with integer, rational, or real-valued weights. The extraction procedure is carried out in the following three steps.

• Transform the given ReLU network into an equivalent CReLU network.
• Extract logical formulae in each CReLU neuron layer by layer.
• Compose the logical formulae corresponding to the individual CReLU neurons according to the layered structure of the CReLU network.

Several one-dimensional and two-dimensional examples are included.

Reference

Y. Zhang and H. Bölcskei, "Extracting formulae in many-valued logic from deep neural networks," arxiv:2401.12113, Jan. 2024.

Code

Python code of the algorithm:

• Archive (ZIP) containing the code.

Please include a reference to the paper referenced above if you use the code in your research.