This project aims to enhance language model interpretability by generating sentences that maximally activate a specific neuron, inspired by the DeepDream technique in image models. We introduce a novel regularization technique that optimizes over a lower-dimensional latent space rather than the full 768-dimensional embedding space, resulting in more coherent and interpretable sentences. Our approach uses an autoencoder and a separate GPT-2 model as an encoder, and a six-layer transformer as a decoder. Despite the current limitation of our autoencoder not fully reconstructing sentences, our work opens up new directions for future research in improving language model interpretability.
Anonymous: Team members hidden
Scott Viteri and Peter Chatain