Multimodal Occlusion Sensitivity

The core mechanism involves iteratively occluding parts of the input space and observing the output delta. For images, this typically involves sliding a gray square over patches; for text, it involves replacing tokens with a [MASK] token. The resulting prediction confidence drop is recorded for each perturbation, creating a direct causal map of feature importance without requiring access to internal model gradients or weights.

Unlike unimodal occlusion, this technique measures how information in one modality depends on another. A critical diagnostic involves occluding an image region and measuring the impact on text-based predictions, or vice versa. This reveals cross-modal grounding failures, such as when a model correctly identifies a "dog" in text but fails to visually locate it, indicating brittle alignment between the vision and language encoders.

The method serves as a ground-truth benchmark for evaluating other explainability techniques. Multimodal Faithfulness is quantified by comparing the model's output change when the top-k features identified by an explanation are removed versus when random features are removed. A faithful explanation will cause a significantly larger drop in confidence, validating that the highlighted features are causally relevant to the prediction.

The granularity of the occlusion grid presents a key hyperparameter tension. A small occlusion patch provides high-resolution importance maps but requires many forward passes, increasing computational cost. A large patch is faster but suffers from the boundary artifact problem, where importance is smeared across adjacent features. Adaptive strategies that dynamically resize the occlusion window based on initial coarse-grained results are often employed to balance fidelity and efficiency.

This technique is often used in conjunction with Modality Ablation to distinguish between global and local modality reliance. While full modality dropout removes an entire data stream to measure its overall contribution, occlusion sensitivity pinpoints the specific spatial regions or semantic tokens within that modality that drive the decision. The combination reveals whether a model uses a broad, holistic integration strategy or relies on a few highly specific cross-modal correlations.

Occlusion maps can expose brittle, non-robust features. If occluding a small, semantically irrelevant background patch causes a dramatic prediction flip, it indicates the model relies on spurious correlations rather than robust concepts. This diagnostic is critical for identifying susceptibility to adversarial attacks, where an imperceptible perturbation in one modality can hijack the cross-modal reasoning chain and produce an incorrect output.