[Paper Review] This Looks Like Those: Illuminating Prototypical Concepts Using Multiple Visualizations

Chiyu Ma, Brandon Zhao+2 more|arXiv (Cornell University)|Oct 28, 2023

Explainable Artificial Intelligence (XAI)4 citations

TL;DR

This paper introduces ProtoConcepts, a method that enhances interpretability in prototype-based image classification by replacing single-image prototypes with multiple visualizations of the same prototypical concept. By showing shared visual features across several training patches, the model clarifies semantic meaning (e.g., color or shape), leading to significantly more accurate human interpretation while maintaining comparable classification accuracy.

ABSTRACT

We present ProtoConcepts, a method for interpretable image classification combining deep learning and case-based reasoning using prototypical parts. Existing work in prototype-based image classification uses a ``this looks like that'' reasoning process, which dissects a test image by finding prototypical parts and combining evidence from these prototypes to make a final classification. However, all of the existing prototypical part-based image classifiers provide only one-to-one comparisons, where a single training image patch serves as a prototype to compare with a part of our test image. With these single-image comparisons, it can often be difficult to identify the underlying concept being compared (e.g., ``is it comparing the color or the shape?''). Our proposed method modifies the architecture of prototype-based networks to instead learn prototypical concepts which are visualized using multiple image patches. Having multiple visualizations of the same prototype allows us to more easily identify the concept captured by that prototype (e.g., ``the test image and the related training patches are all the same shade of blue''), and allows our model to create richer, more interpretable visual explanations. Our experiments show that our ``this looks like those'' reasoning process can be applied as a modification to a wide range of existing prototypical image classification networks while achieving comparable accuracy on benchmark datasets.

Motivation & Objective

• To address the ambiguity in single-image prototype visualizations that hinders human understanding of model reasoning.
• To improve the interpretability of prototype-based image classifiers by enabling users to infer semantic concepts through shared features across multiple visual examples.
• To maintain high classification accuracy while enhancing the clarity and reliability of visual explanations.
• To evaluate whether multi-visualization prototypes lead to better human performance in interpreting model decisions.
• To provide a generalizable method applicable to a wide range of existing prototype-based networks.

Proposed method

• The method replaces single-image prototypes with multi-patch visualizations, where each prototype is represented by multiple training image patches sharing a common semantic feature.
• Prototypes are trained to be similar to feature patches from the same class and dissimilar to others, but now with multiple exemplars per concept.
• The model uses a modified architecture that aggregates similarity scores from multiple visualizations of the same concept to improve decision transparency.
• Commonalities across the multiple visualizations (e.g., consistent color or shape) help users infer the underlying semantic feature being used for classification.
• The approach is compatible with existing prototype networks like ProtoPNet and can be applied as a modular enhancement.
• User studies were conducted using Amazon Mechanical Turk to compare interpretation performance between ProtoPNet and ProtoConcepts.

Experimental results

Research questions

• RQ1 Can multiple visualizations of a single prototype improve human users’ ability to interpret model decisions?
• RQ2 Does providing shared visual features across multiple exemplars reduce ambiguity in understanding which visual property (e.g., color, shape) a prototype represents?
• RQ3 Can a multi-visualization prototype approach maintain classification accuracy comparable to single-prototype methods?
• RQ4 To what extent does the new method improve non-expert user performance in fine-grained image classification tasks?
• RQ5 How does the interpretability of ProtoConcepts compare to that of ProtoPNet in real-world human evaluation?

Key findings

• Users given visual explanations from ProtoConcepts achieved a mean accuracy of 62.1% ± 5.4, significantly outperforming users with ProtoPNet explanations (51.5% ± 5.2), with a p-value of 0.003.
• Users relying on ProtoPNet explanations did not perform better than random guessing (p = 0.289), while users with ProtoConcepts explanations significantly exceeded random performance (p = 2.85 × 10⁻⁵).
• The user study showed a 98% approval rate on quality checks, with only one response removed after screening for nonsensical answers.
• There was no statistically significant difference in self-rated machine learning experience between the two user groups (p = 1), ensuring fair comparison.
• The method achieves comparable classification accuracy to existing prototype networks while significantly improving the interpretability of visual explanations.
• The multi-visualization approach enables users to infer semantic concepts (e.g., 'same shade of blue') more reliably than single-prototype methods.