TeLLMe what you see: using LLMs to explain neurons in vision models

Abstract

As the role of machine learning models continues to expand across diverse fields, the demand for model interpretability grows. This is particularly crucial for deep learning models, which are often referred to as black boxes, due to their highly nonlinear nature. This paper proposes a novel method for generating and evaluating concise explanations for the behavior of specific neurons in trained vision models. Doing so signifies an important step towards better understanding the decision making in neural networks. Our technique draws inspiration from a recently published framework that utilized GPT-4 for interpretability of language models. Here, we extend and expand the method to vision models, offering interpretations based on both neuron activations and weights in the network. We illustrate our approach using an AlexNet model and ViT trained on ImageNet, generating clear, human-readable explanations. Our method outperforms the current state-of-the-art in both quantitative and qualitative assessments, while also demonstrating superior capacity in capturing polysemic neuron behavior. The findings hold promise for enhancing transparency, trust and understanding in the deployment of deep learning vision models across various domains.