Acquisition of Chess Knowledge in Leela Chess Zero (Lc0)

By Alex Wilson

This work reproduces parts of the Acquisition of Chess Knowledge in AlphaZero paper using open source models and code. The visualization below describes concepts from Leela Chess Zero’s ‘representational space’ (internal activations) and mirrors the original paper’s Online Supplement. 36 factors are extracted from the activations of each of the network’s 15 ResNet blocks using a non-negative matrix factorization-based approach.

Block
Factor