Enhancing Visual Feature Attribution via Weighted Integrated Gradients
Researchers have introduced Weighted Integrated Gradients (WG), a novel method to improve the reliability of feature attribution in explainable AI, particularly for computer vision models. Unlike existing methods like Expected Gradients (EG) that treat all baseline images equally, WG adaptively selects and weights baselines based on their informativeness for a given input. This approach, which maintains the axiomatic properties of Integrated Gradients, showed up to a 36% improvement in attribution reliability over EG across various convolutional and Transformer architectures on common image datasets. The trade-off for this enhanced fidelity is a slight increase in computational cost due to the baseline suitability evaluation. AI
IMPACT Enhances reliability of AI model explanations, improving understanding and usability of computer vision models.