Spectral-Spatial Interpretation
Spectral-Spatial Interpretation
Spectral-Spatial Interpretation (SSI) released, using symbolic attribution to explain model scores, assigning contributions to each spatial unit and spectral channel. SSI performed best on average across three metrics under three datasets and two mask baselines, for example, the mean baseline score of HLS joint removal was 0.1545 higher than KernelSHAP's 0.2096. SSI outperformed comparative methods in spatial localization, spectral Hit@K, and channel removal.
🔥🔥🔥Which spectral channels matter, and where?
SSI explains a model score with signed attributions for each spatial cell and channel. It keeps the surrounding image fixed while allocating a cell's contribution among its channels using Shapley values.
A channel's contribution can change across an image. SSI asks how the channels within one spatial cell affect the chosen model score, conditioned on the observed values of all other cells.
✅ SSI has the best reported mean for all three metrics on all three datasets under both masking baselines. For example, mean-baseline HLS joint deletion is .1545 for SSI versus .2096 for KernelSHAP, the strongest non-SSI method on that metric.
✅ SSI leads the compared methods in spatial localization, spectral Hit@K, and channel deletion across the evaluated datasets.
Source: Hugging Face Daily Papers · huggingface.co