Cosy (concept Synthesis): A Novel Architecture-agnostic Machine Learning Framework To Evaluate The Quality Of Textual Explanations For Latent Neurons

Explainable AI (XAI) is a growing field that aims to shed light on how deep neural networks (DNNs) make decisions in a way that humans can understand. One approach is to label neurons in a network with descriptions of the ideas they have learned to recognize. However, the lack of a standardized evaluation metric for these descriptions has been a major obstacle. To address this, researchers have developed CoSy, a quantitative evaluation approach for open-vocabulary neuron descriptions.