Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes

Bor-Shiun Wang, Chien-Yi Wang, Wei-Chen Chiu

TL;DR

MCPNet addresses the need for faithful, multi-scale model explanations by learning multi-level concept prototypes across backbone layers using a Centered Kernel Alignment loss and an energy-weighted PCA. Classification is performed via MCP distributions, aligning image concepts with class-specific centroids through a Class-aware Concept Distribution loss, without relying on a final fully connected layer. The approach yields richer, multi-level explanations—from low-level color cues to high-level object concepts—while maintaining competitive accuracy and demonstrating strong generalization in few-shot scenarios. This framework enhances interpretability and trust in deep models with practical implications for transparent AI in visual classification tasks.

Abstract

Recent advancements in post-hoc and inherently interpretable methods have markedly enhanced the explanations of black box classifier models. These methods operate either through post-analysis or by integrating concept learning during model training. Although being effective in bridging the semantic gap between a model's latent space and human interpretation, these explanation methods only partially reveal the model's decision-making process. The outcome is typically limited to high-level semantics derived from the last feature map. We argue that the explanations lacking insights into the decision processes at low and mid-level features are neither fully faithful nor useful. Addressing this gap, we introduce the Multi-Level Concept Prototypes Classifier (MCPNet), an inherently interpretable model. MCPNet autonomously learns meaningful concept prototypes across multiple feature map levels using Centered Kernel Alignment (CKA) loss and an energy-based weighted PCA mechanism, and it does so without reliance on predefined concept labels. Further, we propose a novel classifier paradigm that learns and aligns multi-level concept prototype distributions for classification purposes via Class-aware Concept Distribution (CCD) loss. Our experiments reveal that our proposed MCPNet while being adaptable to various model architectures, offers comprehensive multi-level explanations while maintaining classification accuracy. Additionally, its concept distribution-based classification approach shows improved generalization capabilities in few-shot classification scenarios.

MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes

TL;DR

MCPNet addresses the need for faithful, multi-scale model explanations by learning multi-level concept prototypes across backbone layers using a Centered Kernel Alignment loss and an energy-weighted PCA. Classification is performed via MCP distributions, aligning image concepts with class-specific centroids through a Class-aware Concept Distribution loss, without relying on a final fully connected layer. The approach yields richer, multi-level explanations—from low-level color cues to high-level object concepts—while maintaining competitive accuracy and demonstrating strong generalization in few-shot scenarios. This framework enhances interpretability and trust in deep models with practical implications for transparent AI in visual classification tasks.

Abstract

Recent advancements in post-hoc and inherently interpretable methods have markedly enhanced the explanations of black box classifier models. These methods operate either through post-analysis or by integrating concept learning during model training. Although being effective in bridging the semantic gap between a model's latent space and human interpretation, these explanation methods only partially reveal the model's decision-making process. The outcome is typically limited to high-level semantics derived from the last feature map. We argue that the explanations lacking insights into the decision processes at low and mid-level features are neither fully faithful nor useful. Addressing this gap, we introduce the Multi-Level Concept Prototypes Classifier (MCPNet), an inherently interpretable model. MCPNet autonomously learns meaningful concept prototypes across multiple feature map levels using Centered Kernel Alignment (CKA) loss and an energy-based weighted PCA mechanism, and it does so without reliance on predefined concept labels. Further, we propose a novel classifier paradigm that learns and aligns multi-level concept prototype distributions for classification purposes via Class-aware Concept Distribution (CCD) loss. Our experiments reveal that our proposed MCPNet while being adaptable to various model architectures, offers comprehensive multi-level explanations while maintaining classification accuracy. Additionally, its concept distribution-based classification approach shows improved generalization capabilities in few-shot classification scenarios.
Paper Structure (33 sections, 6 equations, 17 figures, 8 tables, 1 algorithm)

This paper contains 33 sections, 6 equations, 17 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Post-hoc Interpretation Methods
  4. Inherently Interpretable Methods
  5. Method
  6. Overall Framework
  7. Centered Kernel Alignment (CKA) Loss
  8. Multi-level Concept Prototype Extraction
  9. Class-aware Concept Distribution Loss
  10. Multi-Level Concept Prototypes Classifier
  11. Experiments
  12. Implementation Details
  13. Layer Selections.
  14. Training hyper-parameters.
  15. Quantitative Results
  16. ...and 18 more sections

Figures (17)

  • Figure 1: The “Bobcat” was correctly classified by our MCPNet but incorrectly classified as a “Fox” by PIP-Net nauta2023pip. On the right side, we provide an illustration of using our proposed Multi-level Concept Prototype (MCP) distribution to classify and explain the input image. In particular, our concept prototypes are extracted from multiple layers of the classification model (thus having low-level to high-level concepts). In comparison with a recent state-of-the-art baseline, PIP-Net nauta2023pip shown on the left side which only adopts single-level explanations (symbolized as colorful boxes on the bottom portion, they are usually extracted from the last model layer), our proposed MCPNet provides more comprehensive explanations as well as better classification performance.
  • Figure 2: The overall workflow of our MCPNet. To classify the image, we first calculate the class-specific MCP distribution (yellow box) by averaging the MCP distributions from instances of specific classes in the training set (red box). Utilizing the class-specific MCP distribution, images are classified by identifying the most similar class via calculating the Jensen-Shannon (JS) divergence (brown box).
  • Figure 3: Visualization of CKA similarities for each pair of segments from different model layers (model backbone: ResNet50; dataset: AWA2 xian2018zero). The average for the upper triangular portion of each CKA similarity matrix is also provided accordingly. With apply our proposed CKA loss, the similarities between segments are clearly reduced, i.e. leading to more distinct concept segments.
  • Figure 4: Concept prototype examples from MCPNet and PIP-Net nauta2023pip. We show the top-5 responses for the sampled concept prototypes. For MCPNet, the concept prototype from various layers generates explanations in different scales, e.g. color-like explanations in low-layer and object-like explanations in high-layer. On the contrary, PIP-Net nauta2023pip only provides single-scale patch-level explanations.
  • Figure 5: The sampled multi-level concept prototypes learnt by our proposed MCPNet. (Backbone : ResNet50)
  • ...and 12 more figures