Research map / Label-efficient and robust learning / Recognition and long-tail

MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes

CVPR 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Self-explaining and visual explanation networks · 16 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MCPNet, an inherently interpretable image classifier that learns concept prototypes across multiple layers via CKA loss and energy-weighted PCA, and classifies by Jensen-Shannon matching of multi-level concept prototype distributions.

Problem

Existing inherently interpretable and post-hoc explanation methods (e.g., ProtoPNet variants) only explain a classifier from its last feature map, leaving low- and mid-level decision processes unexplained.

Approach

Partition feature maps from multiple layers into channel segments, decorrelate segments with a CKA loss, extract each segment's concept prototype via weighted PCA, and classify by comparing an image's multi-level concept-response distribution to class centroid distributions (trained with a class-aware JS-divergence loss), removing the FC layer.

Key results

• MCPNet with ResNet50 reaches 93.92% on AWA2, 93.88% on Caltech101 and 80.15% on CUB, matching or beating ProtoPNet-series baselines (e.g., PIP-Net 85.99% on AWA2).
• With ConvNeXt-tiny, MCPNet reaches 95.61% on AWA2 and 83.45% on CUB, exceeding the ConvNeXt-tiny baseline on AWA2 (96.55% baseline is higher on AWA2, so MCPNet is below it there).
• In 5-shot classification on unseen AWA2 classes, MCPNet achieves 73.79% versus 60.55% for the baseline and 19.71%-33.68% for ProtoPNet-series methods.

Measured effects reported in the paper

Surprising findings

• Using only the CCD loss without CKA gives 44.85% on Caltech101, far below baseline, and produces duplicate concept prototypes.
• PIP-Net accuracy with Inception v3 collapses to 43.82% on AWA2 and 6.76% on CUB under the authors' ImageNet-pretrained setup.

Limitations

• Accuracy does not uniformly match the baseline (e.g., ConvNeXt-tiny on AWA2 and Caltech101 are lower than baseline), with a slight compromise noted in the ablation.
• Evaluated only on three datasets and convolution-based backbones.

Experimental setup

Datasets: AWA2 Caltech101 CUB 200 2011

Benchmarks: AWA2 Caltech101 CUB 200 2011

Metrics: classification accuracy 5-shot classification accuracy

Baselines: ProtoTree Deformable ProtoPNet ST-ProtoPNet PIP-Net Baseline FC classifier

Builds on: ProtoPNet Concept Bottleneck Model PIP-Net Centered Kernel Alignment

Closest papers by meaning

• Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.87)
• Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning (AAAI 2024 · similarity 0.86)
• Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.86)
• A Robust Prototype-Based Network with Interpretable RBF Classifier Foundations (AAAI 2025 · similarity 0.85)
• Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.85)
• Beyond Concept Bottleneck Models: How to Make Black Boxes Intervenable? (NeurIPS 2024 · similarity 0.85)