WheaCha: An Explanation Method for Models of Code
Explore a groundbreaking explanation method for code models in this 17-minute conference talk from OOPSLA2 2023. Discover WheaCha, a novel approach that separates input programs into 'wheat' (defining features) and 'chaff' to explain model predictions. Learn how the HuoYan tool implements WheaCha to explain four prominent code models: code2vec, seq-GNN, GGNN, and CodeBERT. Examine the efficiency of HuoYan, taking less than twenty seconds on average to compute wheat for an input program. Understand how the wheat used by models for predictions primarily consists of simple syntactic or lexical properties. Compare WheaCha's effectiveness to other explainability methods like SIVAND, CounterFactual Explanations, Integrated Gradients, and SHAP. Explore the practical applications of WheaCha in helping users identify defective code models trained on mislabeled or biased data. Gain insights into the latest advancements in explainable AI for code models and their potential impact on software development and analysis. Read more