Occlusion
A perturbation-based post-hoc explanation method that measures importance by systematically occluding different parts of the input and observing changes in model output. It is model-agnostic but computationally expensive.
Timeline (last 90 days)
The evaluation results show that only a few methods maintain consistent performance across all evaluation criteria, among which Occlusion performs best overall on the two dimensions of faithfulness and interpretability.
Occlusion is a class of perturbation-based post-hoc explanation methods that measures importance by systematically occluding different parts of the input and observing changes in the model's output probabilities.
Occlusion's advantages are being model-agnostic and conceptually intuitive; its disadvantages are that computational cost grows with the number of occlusion units and that it is sensitive to the choice of occlusion granularity.
All Facts (3)
The evaluation results show that only a few methods maintain consistent performance across all evaluation criteria, among which Occlusion performs best overall on the two dimensions of faithfulness and interpretability.
50%UnverifiedOcclusion is a class of perturbation-based post-hoc explanation methods that measures importance by systematically occluding different parts of the input and observing changes in the model's output probabilities.
50%UnverifiedOcclusion's advantages are being model-agnostic and conceptually intuitive; its disadvantages are that computational cost grows with the number of occlusion units and that it is sensitive to the choice of occlusion granularity.
50%