Deep Learning Attribution Methods: A Comprehensive Review

Deep learning (DL) has revolutionized various fields, but its lack of interpretability hampers understanding how these models make decisions. This issue is particularly problematic in applications like healthcare, finance, and security, where explanations are crucial for trustworthy decision-making. In this paper, the authors propose MFABA (Multi-Facet Attribution by Backpropagation with Adaptive Sampling), an algorithm that addresses DL interpretability challenges while providing accurate attribution outputs.
MFABA’s novelty lies in its ability to adaptively sample inputs, ensuring the explanations are faithful and efficient. The authors propose two versions of MFABA: sharp and smooth, which cater to different use cases. They also provide theoretical proof for the axiomatic derivation process of MFABA, supporting its sensitivity and implementation invariance.
The authors conduct a large-scale experiment showcasing MFABA’s superior performance compared to other state-of-the-art methods, including Integrated Gradients (IG), saliency maps (SM), smoothed gradients (SG), DeepLift, BIG, and adaptive sampling methods. The results demonstrate that MFABA provides more focused attention on the identified subjects, offering better interpretability than other methods.
The authors also evaluate MFABA’s attribution performance using error rate indicators, insertion/deletion score (Petsiuk et al., 2018), and area under accuracy curve (Kapishnikov et al., 2019). The results demonstrate that MFABA outperforms other methods in terms of accuracy.
In summary, the authors present a novel algorithm, MFABA, to address DL interpretability challenges while providing accurate attribution outputs. By adaptively sampling inputs, MFABA ensures faithful and efficient explanations, supporting its sensitivity and implementation invariance. The experimental results demonstrate MFABA’s superior performance compared to other state-of-the-art methods, making it a valuable tool for practitioners and researchers seeking trustworthy AI.