From Pixels to Perception: Interpretable Predictions via Instance-wise Grouped Feature Selection

TL;DR

This work tackles the challenge of interpretable image classification by enforcing inherent interpretability through instance-wise sparsification over semantically meaningful regions. It introduces P2P, which partitions inputs into perceptually meaningful regions, learns binary masks via Gumbel-Softmax, and jointly models inter-region relationships with a region-wise covariance captured by a Logit-Normal embedding. A dynamic sparsity mechanism selects per-instance masking levels by adjusting a threshold to meet a target certainty, yielding interpretable explanations without sacrificing accuracy. Across diverse datasets, P2P maintains strong predictive performance while providing faithful, human-aligned explanations and informative visualizations that reveal region-level decision rationale.

Abstract

Understanding the decision-making process of machine learning models provides valuable insights into the task, the data, and the reasons behind a model's failures. In this work, we propose a method that performs inherently interpretable predictions through the instance-wise sparsification of input images. To align the sparsification with human perception, we learn the masking in the space of semantically meaningful pixel regions rather than on pixel-level. Additionally, we introduce an explicit way to dynamically determine the required level of sparsity for each instance. We show empirically on semi-synthetic and natural image datasets that our inherently interpretable classifier produces more meaningful, human-understandable predictions than state-of-the-art benchmarks.

Figures (14)

• Figure 1: Masking 80% of input under different constraints. All 3 masks lead to similar predictive performance but only P2P provides interpretability by sparsity.
• Figure 2: Schematic overview of P2P. A frozen region proposer partitions the input ${\bm{x}}$ into perceptually meaningful parts $R_{1:D}$, which are assigned a learned selection probability. The mask is then binarized by sampling, leading to ${\bm{x}}_m$ that serves as the input to the classifier.
• Figure 3: Insertion Fidelity, where the most important pixels of the explanation ${\bm{x}}_m$ are iteratively added to a black image, measuring how much information is required until the original prediction is recovered. The faster, i.e. the steeper the curve, the better. Results are reported as averages and standard deviations across ten seeds.
• Figure 4: Ablation study of P2P with a constant $\tau$. In "Matched", we set $\tau$ to the average value of the dynamic P2P, which is $20\%, 20\%, 40\%, 30\%$, for each dataset, respectively.
• Figure 5: Masked inputs ${\bm{x}}_m$ of COCO-10 for selected methods.