RISE

RISE generates an importance map by creating N binary masks (typically thousands) that randomly occlude different regions of the input. Each mask is upsampled to the input resolution using bilinear interpolation to produce smooth, continuous occlusion patterns. The model's confidence score on each masked input is recorded, and the final saliency map is computed as the weighted sum of all masks, where each mask's weight is the model's prediction score for the target class on that masked input. This process approximates the expected marginal effect of preserving each pixel.

RISE treats the model as a complete black box, requiring no access to internal weights, gradients, or architecture details. It only needs the ability to query the model with perturbed inputs and receive output probabilities. This makes RISE applicable to any model type, including proprietary APIs, ensemble methods, and non-differentiable classifiers where gradient-based methods like Integrated Gradients or Grad-CAM are impossible to use. The trade-off is computational cost, as each explanation requires thousands of forward passes.

As the number of random masks approaches infinity, the RISE importance map converges to the true conditional expectation of the model output given each pixel's visibility. This provides a probabilistic guarantee that the saliency map faithfully represents the model's reliance on each input region. In practice, the method approximates the Shapley value of each pixel under a coalitional game where pixels are randomly included or excluded. The variance of the estimate decreases with the square root of the mask count.

Masks are generated by first creating small binary grids (e.g., 7x7 or 11x11) where each cell is independently set to 1 (preserve) with probability p or 0 (occlude) with probability 1-p. These coarse masks are then upsampled to the full input resolution using bilinear interpolation, producing smooth spatial transitions that avoid sharp occlusion boundaries. The probability p controls the expected fraction of preserved pixels and is typically set between 0.3 and 0.7. Higher p values produce masks that preserve more context, while lower p values isolate smaller regions.

RISE improves upon the classic Occlusion Sensitivity method in two key ways. First, instead of systematically sliding a single occluding patch, RISE randomizes occlusion patterns, avoiding grid artifacts and providing smoother, more visually coherent saliency maps. Second, by averaging over thousands of masks, RISE captures non-linear interactions between occluded regions that a single sliding window cannot detect. The method also avoids the computational inefficiency of evaluating every possible occlusion position, instead sampling strategically from the space of all possible masks.

RISE saliency maps are commonly evaluated using the Deletion Metric and Insertion Metric. For deletion, pixels are progressively removed from most to least important according to the RISE map, and the area under the resulting probability curve (AUC) is measured—a sharp drop indicates faithful attribution. For insertion, pixels are progressively added into a blurred baseline, and a rapid increase in confidence confirms the map correctly identifies salient regions. RISE consistently achieves competitive AUC scores against gradient-based methods like Grad-CAM and Integrated Gradients on standard benchmarks.