Given a neural network (W,f) [1] with network function Ψ(W,f) and an input x ∈ ℝᵈ, compute the knowledge matrix M(W,f)(x). They were introduced in [2] and studied further in [3].

[1] Armenta, M., Jodoin, P-M., The Representation Theory of Neural Networks (2021)

[2] Armenta, M., Brüstle, T., Hassoun, S., Reineke, M., Double framed moduli spaces of quiver representations (2021)

[3] Leblanc, S., Rasolomanana, A., Armenta, M., Hidden Activations Are Not Enough: A General Approach to Neural Network Predictions (2024)

You can download this using

pip install git+https://github.com/samueleblanc/knowledgematrix.git

For an example on how to use, please refer to example.py.

Any contribution is welcomed. Thank you in advance! However, we are primarily interested with help on the following:

• Implement other type of layers or other architecture.
• Improve memory efficiency or speed.

To set up, once the code is downloaded, simply run

pip install -e .

in the root directory. Then, to run a file, for instance the test for AlexNet, run

python extra/tests/alexnet.py

again from the root directory.

This code was written by Marco Armenta and Samuel Leblanc.

MIT