AI Tasting Seminar
Date:
Attribution methods are essential for understanding how deep networks make decisions, yet prediction-based approaches routinely outperform classic gradient-based ones. We show that the key difference lies in their frequency content: gradient-based attribution maps contain excessive high-frequency noise, while prediction-based maps do not. By analyzing gradients across multiple CNN classifiers, we trace this noise to aliasing introduced during downsampling operations. Applying an optimal low-pass filter removes this high-frequency contamination and dramatically improves the performance of gradient-based methods, reshaping the ranking of state-of-the-art attribution techniques. Our results highlight a simple principle—filtering out high-frequency noise restores the faithfulness of gradients, and point toward a renewed appreciation of efficient, interpretable gradient-based explanations.