Explaining Human Comparisons using Alignment-Importance Heatmaps

TL;DR

This paper introduces Alignment Importance Score (AIS) heatmaps to explain human similarity judgments (HSJs) by identifying which DNN feature maps drive alignment with human perception. AIS is computed by mask-based perturbations in the final convolutional layer and selecting a subset of feature maps that maximizes cross-domain predictability of HSJs; heatmaps are then generated to visualize image regions tied to high AIS. Across multiple architectures and datasets, AIS-based pruning improves out-of-sample HSJ prediction and reveals that comparison-relevant information can diverge from visually salient regions. The study also quantifies how AIS heatmaps relate to traditional saliency maps, finding meaningful overlap in some cases (notably Animals) but important divergences in others, and demonstrates broad generalization of AIS across architectures, informing how to build interpretable explanations for human comparisons.

Abstract

We present a computational explainability approach for human comparison tasks, using Alignment Importance Score (AIS) heatmaps derived from deep-vision models. The AIS reflects a feature-map's unique contribution to the alignment between Deep Neural Network's (DNN) representational geometry and that of humans. We first validate the AIS by showing that prediction of out-of-sample human similarity judgments is improved when constructing representations using only higher-scoring AIS feature maps identified from a training set. We then compute image-specific heatmaps that visually indicate the areas that correspond to feature-maps with higher AIS scores. These maps provide an intuitive explanation of which image areas are more important when it is compared to other images in a cohort. We observe a correspondence between these heatmaps and saliency maps produced by a gaze-prediction model. However, in some cases, meaningful differences emerge, as the dimensions relevant for comparison are not necessarily the most visually salient. To conclude, Alignment Importance improves prediction of human similarity judgments from DNN embeddings, and provides interpretable insights into the relevant information in image space.

Figures (18)

• Figure 1: Out-of-sample predictions of human similarity judgments using image embeddings. Full: using all 512 feature maps. Retained: using feature maps identified from an independent training set. The numbers above the second and fourth columns in each group represent averages of feature-map set sizes across 40 folds. Error bars indicate standard errors adjusted for paired-comparisons loftus1994using.
• Figure 2: Heatmaps generated using Alignment Importance Scores of feature maps trained with Ecoset. For each dataset, two images with subjectively higher interpretability (top two rows) and lower interpretability (bottom two rows) were selected.
• Figure 3: Histograms describing statistics of Alignment Importance Score distributions for models trained on Ecoset or ImageNet. The x-axis of (b) and (c) are displayed in e-4 format. A star symbol (*) indicates a significant difference between the two distributions as determined by a KS test.
• Figure 4: AIS heatmaps thresholded at 60th percentile
• Figure 5: AIS heatmaps thresholded at 70th percentile