RISE

RISE operates without any access to the model's internal architecture, gradients, or parameters. It treats the model as a complete black box, requiring only the ability to query it with inputs and receive outputs. This makes it universally applicable to any model type—convolutional neural networks, vision transformers, or even proprietary cloud APIs—without modification. The method works by observing how the model's prediction for a target class changes when different parts of the input are randomly occluded, building an importance map purely from input-output behavior.

The core mechanism involves generating thousands of binary masks that randomly occlude portions of the input image. Each mask is a grid of 0s and 1s, typically created at a lower resolution than the input and upsampled using bilinear interpolation to produce smooth occlusion boundaries. Key properties:

• Mask count: Usually 4,000–8,000 masks for stable results
• Mask resolution: Smaller grids (e.g., 7×7 or 11×11) upsampled to image size
• Occlusion probability: Each cell has a fixed probability (often 0.5) of being set to 1 (preserved) or 0 (masked)
• Smooth transitions: Bilinear upsampling creates soft edges that avoid grid artifacts in the final saliency map

RISE computes the final saliency map as a weighted sum of all random masks, where each mask's weight is the model's confidence score for the target class when that mask is applied. The formula is:

S = (1 / N) × Σ (M_i × f(I ⊙ M_i))

Where:

• S: Final saliency map
• N: Total number of masks
• M_i: The i-th binary mask
• f(I ⊙ M_i): Model's score for the masked input
• ⊙: Element-wise multiplication

Masks that preserve regions critical to the prediction receive higher weights, causing those regions to accumulate importance across the averaging process.

RISE exchanges computational efficiency for model-agnostic flexibility. Each mask requires a full forward pass through the model, making the method significantly slower than gradient-based approaches like Grad-CAM or Integrated Gradients. Typical resource considerations:

• Inference count: 4,000–8,000 forward passes per explanation
• GPU memory: Only requires storing one masked image at a time
• Latency: Seconds to minutes per image depending on model size and mask count
• Batch processing: Masks can be batched to leverage GPU parallelism, reducing wall-clock time
• Trade-off: Higher mask counts improve explanation fidelity but increase compute linearly

RISE is particularly valuable in medical imaging contexts where model internals may be inaccessible due to proprietary constraints or regulatory requirements. Its black-box nature allows it to explain FDA-cleared SaMD devices without modifying validated software. Key clinical advantages:

• No model modification: Preserves regulatory validation status
• Vendor-agnostic: Works with third-party diagnostic APIs
• Lesion attribution: Can verify that predictions are based on actual pathological regions
• Audit trail compatibility: Deterministic given a fixed random seed, supporting reproducible explanations
• Limitation: The stochastic nature requires careful seed management for regulatory documentation

RISE occupies a distinct position in the explainability landscape compared to gradient-based alternatives:

vs. Grad-CAM:

• RISE: Black-box, no gradient access, slower, produces finer-grained maps
• Grad-CAM: White-box, requires final convolutional layer access, near-instantaneous, coarser localization

vs. Integrated Gradients:

• RISE: Perturbation-based, satisfies no formal axioms, model-agnostic
• Integrated Gradients: Gradient-based, satisfies completeness and sensitivity axioms, requires differentiable model

vs. LIME:

• Both are black-box perturbation methods
• RISE uses random binary masks on raw pixels
• LIME uses superpixel segmentation and fits a local surrogate model