[Paper Review] Understanding Deep Architectures by Visual Summaries

Marco Carletti, Marco Godi+3 more|arXiv (Cornell University)|Jan 27, 2018

Visual Attention and Saliency Detection23 references4 citations

TL;DR

This paper proposes a novel visualization framework that identifies and clusters salient, crisp image regions across multiple images of the same class to reveal what semantic parts a deep network has learned to recognize. By combining sparse optimization for sharp saliency masks with semantic flow-based clustering, the method generates interpretable visual summaries that correlate with network performance and enable accuracy improvements via summary-driven SVM specialization, increasing top-1 accuracy by 1.08% on ImageNet.

ABSTRACT

In deep learning, visualization techniques extract the salient patterns exploited by deep networks for image classification, focusing on single images; no effort has been spent in investigating whether these patterns are systematically related to precise semantic entities over multiple images belonging to a same class, thus failing to capture the very understanding of the image class the network has realized. This paper goes in this direction, presenting a visualization framework which produces a group of clusters or summaries, each one formed by crisp salient image regions focusing on a particular part that the network has exploited with high regularity to decide for a given class. The approach is based on a sparse optimization step providing sharp image saliency masks that are clustered together by means of a semantic flow similarity measure. The summaries communicate clearly what a network has exploited of a particular image class, and this is proved through automatic image tagging and with a user study. Beyond the deep network understanding, summaries are also useful for many quantitative reasons: their number is correlated with ability of a network to classify (more summaries, better performances), and they can be used to improve the classification accuracy of a network through summary-driven specializations.

Motivation & Objective

• To address the lack of systematic analysis of whether deep networks consistently exploit specific semantic parts across multiple images of the same class.
• To develop a visualization method that generalizes across images of a class to identify discriminative visual patterns rather than analyzing single images in isolation.
• To quantify the relationship between the number of learned visual summaries and network classification performance.
• To enable model improvement by specializing classifiers on the discovered visual summaries.

Proposed method

• Applying sparse optimization to generate crisp, binary saliency masks that highlight regions most responsible for classification decisions.
• Using connected components of these masks as local features for clustering across images of the same class.
• Employing affinity propagation with a semantic flow similarity measure to group spatially and semantically similar regions from different images.
• Generating visual summaries as clusters of salient regions, each representing a distinct semantic part (e.g., nose, tail, legs) of the object class.
• Training linear SVMs on images belonging to each summary to specialize classification for that part.
• Combining the SVM scores with the original network’s softmax output via a convex weighted sum to boost overall accuracy.

Experimental results

Research questions

• RQ1 Do deep networks systematically exploit the same semantic parts (e.g., nose, eyes, tail) across multiple images of the same class?
• RQ2 Can visual summaries derived from multiple images provide a more interpretable and quantitatively measurable explanation of network behavior than single-image saliency maps?
• RQ3 Is there a correlation between the number of visual summaries and the classification accuracy of a deep network?
• RQ4 Can the discovered summaries be used to improve the classification performance of a pre-trained network?

Key findings

• The number of visual summaries correlates positively with network performance: from AlexNet (6 summaries on average) to ResNet (9 summaries), classification accuracy increases in parallel.
• GoogleNet learns 9 summaries on the 'pickup' class, while AlexNet learns only 6, visually explaining GoogleNet’s superior performance through richer part-level understanding.
• User studies show high inter-rater agreement (mean μU = 0.35) on summary tags, confirming that summaries capture meaningful, fine-grained semantic parts that users can reliably identify.
• The method improves top-1 accuracy on ImageNet by 1.08% by combining original network predictions with SVMs trained on visual summaries.
• Images misclassified by AlexNet are correctly reclassified when the summary-based boosting is applied, demonstrating practical utility in failure recovery.
• The use of crisp masks outperforms smooth masks in inducing higher classification loss upon perturbation, validating the effectiveness of sparsity in saliency detection.