Attribution Calculation

Aumann–Shapley values are originally designed for evaluating the contribution of a game player based on the marginal difference of the particular output between the case of removing and maintaining this player. It can be applied to calculate attributions in deep networks, if we can achieve feature removing and maintaining. However, we cannot achieve it in neural networks without retraining. As an alternative, a baseline expressing the no-signal state is widely used in attribution methods, i.e., instead of actually removing features, the original feature values are replaced with the baseline values. By recalling the definition of Aumann–Shapley method, we first propose two primal baseline properties for calculating baselines to apply Aumann–Shapley values to deep networks. Then, we design an optimization-based baseline selection method which is slow but theoretically accurate and a quadratic approximation of the optimization-based method which is much faster but slightly worse. The experiments show our methods outperform other alternative baselines and achieve better attributions than other attribution calculation methods.

Calculated attribution scores represent feature importance to the particular output, as shown in the image below.


“Output-targeted baseline for neuron attribution calculation.” Image and Vision Computing. Article GitHub Repo