HOLMES: HOLonym-MEronym based Semantic inspection for Convolutional Image Classifiers

TL;DR

The paper addresses the lack of interpretable explanations for CNN-based image classifiers by decomposing a holonym into meronyms and explaining decisions at the part level. HOLMES constructs meronym detectors by transferring knowledge from the holonym CNN, sharing its feature extractor, and training a dedicated classifier for each meronym set. Per-part Grad-CAM maps are thresholded, ablated, and combined into a global heatmap $G = \sum_i z_i x^{(p_i)}$, where $z_i$ are normalized score drops and parts are selected using thresholds on score drop and per-part $F1$. Experiments on PASCAL-Part and ImageNet show that HOLMES delivers interpretable part-based explanations with performance comparable to Grad-CAM on fidelity metrics, while enabling semantic explanation of model decisions.

Abstract

Convolutional Neural Networks (CNNs) are nowadays the model of choice in Computer Vision, thanks to their ability to automatize the feature extraction process in visual tasks. However, the knowledge acquired during training is fully subsymbolic, and hence difficult to understand and explain to end users. In this paper, we propose a new technique called HOLMES (HOLonym-MEronym based Semantic inspection) that decomposes a label into a set of related concepts, and provides component-level explanations for an image classification model. Specifically, HOLMES leverages ontologies, web scraping and transfer learning to automatically construct meronym (parts)-based detectors for a given holonym (class). Then, it produces heatmaps at the meronym level and finally, by probing the holonym CNN with occluded images, it highlights the importance of each part on the classification output. Compared to state-of-the-art saliency methods, HOLMES takes a step further and provides information about both where and what the holonym CNN is looking at, without relying on densely annotated datasets and without forcing concepts to be associated to single computational units. Extensive experimental evaluation on different categories of objects (animals, tools and vehicles) shows the feasibility of our approach. On average, HOLMES explanations include at least two meronyms, and the ablation of a single meronym roughly halves the holonym model confidence. The resulting heatmaps were quantitatively evaluated using the deletion/insertion/preservation curves. All metrics were comparable to those achieved by GradCAM, while offering the advantage of further decomposing the heatmap in human-understandable concepts, thus highlighting both the relevance of meronyms to object classification, as well as HOLMES ability to capture it. The code is available at https://github.com/FrancesC0de/HOLMES.

Figures (13)

• Figure 1: HOLMES pipeline. Given an input image of class $c$, its parts (meronyms) are extracted from a Knowledge Base (a). Images depicting each part are either extracted from a densely annotated dataset or collected through Web scraping (b), and then used to train a meronym model by exploiting, through transfer learning, the implicit knowledge embedded in the original holonym model (c). The meronym model then produces part-based explanations, highlighting the most relevant parts for the class prediction (d).
• Figure 2: HOLMES Explanation example for the horse class -- PASCAL-Part (a) and sorrel class -- ImageNet (b). For each part, the corresponding ablation mask (grey), the per-part calibrated F1-score and the holonym score drop are shown. For PASCAL-Part, the ablation masks are compared against the ground truth bounding boxes (green). The final heatmap(s) show the part-based explanations. Two and one part are included in the explanations for examples (a) and (b), respectively, as they exceed both the holonym score drop threshold $T_s$ (0.1) and the calibrated F1-score threshold $T_{F1}$ (0.7).
• Figure 3: Distribution (violin plot) of the average per-part calibrated F1-score.
• Figure 4: Average per-part calibrated score as a function of the number of parts per holonym class (colored dots represent a holonym, blue line is the mean average per-part F1-score).
• Figure 5: Distribution (violin plots) of the average score drop and maximum score drop (in percentages) per image on the PASCAL-Part (a) and ImageNet (b) validation sets. The score drop is calculated for each image and meronym by ablating the corresponding mask; then, the average and maximum score drop are computed over all meronyms appearing in an image.