A toy model of universality: Reverse engineering how networks learn group operations
2023.07
A representation is a homomorphism (weight matrix) which maps input vector to the output vector.
The paper deals the universality Universality: whether different models have similar features with one-layer transformer as Transformers learn group theoretic automata.
Acquisition of chess knowledge in alphazero
2022.11
Emergent world representations: Exploring a sequence model trained on a synthetic task