Occlusion Sensitivity

The core mechanism involves iteratively sliding a gray square (or a baseline patch) across the input image. At each position, the patch replaces the underlying pixels, and the model's output probability for the target class is recorded. The resulting drop in confidence forms a heatmap, where darker regions indicate higher importance. The size of the occluding patch, or kernel size, is a critical hyperparameter that determines the granularity of the explanation.

The raw output is a 2D map of probability deltas. For each patch position (x, y), the method computes P(class|original_image) - P(class|occluded_image). A large positive delta signifies that the occluded region was critical for the prediction. This direct measurement of prediction sensitivity avoids the need for backpropagation or access to internal model weights, making it suitable for any model that outputs a probability score.

The choice of the occluding patch size presents a fundamental trade-off between spatial resolution and computational cost:

• Small Kernels (e.g., 4x4): Provide high-resolution, fine-grained heatmaps but require many forward passes, increasing latency.
• Large Kernels (e.g., 32x32): Are computationally cheap but produce coarse, blocky explanations that may miss small, critical features.
• Adaptive Kernels: Advanced implementations can vary patch size based on local image entropy.

The pixel value used to fill the occluding patch significantly impacts the explanation's fidelity. Common strategies include:

• Zero Padding: Replacing pixels with black (0).
• Mean Pixel Value: Using the dataset's average pixel intensity.
• Blurring: Applying a Gaussian blur to the patch region instead of a solid color.
• Learned Baseline: Using a generative model to inpaint a realistic, neutral background. A poor baseline can introduce out-of-distribution artifacts, causing the model to react to the patch itself rather than the missing information.

Occlusion sensitivity is computationally expensive. For an image of size H x W and a stride S, the number of forward passes is approximately (H/S) * (W/S). This linear scaling with input resolution makes it impractical for real-time applications. However, the process is embarrassingly parallel, as each patch perturbation is an independent inference call that can be batched and distributed across multiple GPUs to reduce wall-clock time.

Occlusion sensitivity is the deterministic precursor to Randomized Input Sampling for Explanation (RISE). While occlusion uses a systematic grid of square masks, RISE uses thousands of random, soft-edged binary masks. RISE produces smoother, more visually appealing heatmaps by averaging over many random perturbations, but occlusion sensitivity offers a more direct, exhaustive measurement of local importance without the stochastic noise inherent in random sampling.