In 2019, a group of researchers at Duke University asked a deceptively simple question: what if a neural network could point at a bird photo and say "I think this is a cardinal because this part looks like that part of a known cardinal"? The resulting paper, ProtoPNet, launched an entire subfield of prototype-based interpretable AI. Since then, dozens of teams have tried to improve on that idea - ProtoPool, TesNet, ST-ProtoPNet, Deformable ProtoPNet - each one chipping away at the same fundamental problem. All of them shared one blind spot. They only looked at the final layer.

The "Last Layer" Problem (Or: Reading Only the Last Page of a Novel)

Deep neural networks process images through a hierarchy. Early layers detect edges and textures. Middle layers recognize shapes and parts. Final layers assemble everything into "that's a golden retriever." Every prototype-based model until now has grabbed its explanations exclusively from that last layer - the high-level semantic summary.

This is like asking someone why they enjoyed a movie and only accepting answers about the plot twist. What about the cinematography? The score? The weird lighting choice in act two?

Wang, Wang, and Chiu from National Yang Ming Chiao Tung University and NVIDIA Research argue - correctly, it turns out - that an explanation ignoring low- and mid-level features isn't really a full explanation at all. Their solution: MCPNet++, a Multi-Level Concept Prototypes Classifier that pulls interpretable concepts from every stage of the network.

How MCPNet++ Actually Works (Without the Math Headache)

The model uses something called Centered Kernel Alignment (CKA) loss combined with energy-based weighted PCA to extract concept prototypes at multiple feature map levels. In human terms: it watches what the network pays attention to at each processing stage and crystallizes those patterns into reusable "concept prototypes."

The "++" in MCPNet++ means it works with both CNNs and Vision Transformers - two architecturally different backbone types that process features in fundamentally different ways. Getting one interpretability method to work cleanly across both is a bit like building a universal translator that handles both Mandarin and Morse code.

Here's where it gets clever. The discovered concepts aren't hand-labeled by humans. MCPNet++ finds them autonomously, then uses a large language model to bridge those learned concepts back to terms humans actually understand. The network discovers that a certain mid-level pattern matters for classification, and the LLM tells you "that's the striped wing pattern." The machine explains itself, then another machine translates the explanation into English. We live in interesting times.

Why Should You Care About Interpretable AI?

The EU AI Act is now in effect. The FDA keeps tightening requirements for AI in medical devices. Insurance companies want to know why an algorithm denied a claim. The era of "trust me, the neural network said so" is ending, and it's ending fast.

In medical imaging, a radiologist doesn't want a model that says "malignant, 94% confidence." They want to see which tissue patterns triggered that call, ideally mapped to patterns from known cases. MCPNet++'s multi-level approach means explanations can reference texture abnormalities (low-level), structural irregularities (mid-level), and overall morphology (high-level) simultaneously. That's closer to how pathologists actually think.

The same logic applies to autonomous driving, manufacturing quality control, and any domain where "it just works" isn't a sufficient safety argument. If you've ever used visual AI tools like combb2.io to enhance images, you've interacted with models processing features at multiple levels - MCPNet++'s contribution is making that multi-level processing transparent.

The Results: It Actually Works

The experimental results show MCPNet++ provides more comprehensive explanations without sacrificing classification accuracy. That "without sacrificing" part matters enormously. Interpretable models have historically paid a performance tax - you could have a model that explains itself or one that's accurate, pick one. MCPNet++ suggests that tradeoff might be a false choice, at least for concept-prototype approaches.

The discovered concepts align closely with human understanding, which was validated through user studies. Humans looked at what the model found important and went "yeah, that makes sense." A low bar, perhaps, but one that plenty of explainability methods still trip over.

What's Next for Prototype Networks?

A recent survey by Sacha et al. (arXiv: 2410.08925) catalogs the explosion of prototype formulations in interpretable deep learning, and a comprehensive review of self-interpretable neural networks (arXiv: 2501.15638) places these methods in a broader taxonomy alongside attribution-based and rule-based approaches. The field is maturing rapidly.

MCPNet++ represents something of a milestone: multi-level interpretability, backbone-agnostic design, autonomous concept discovery, and LLM-powered human translation, all in one package. Whether it becomes the standard approach or just a well-executed proof of concept, the core insight - that faithful explanations require looking at more than the final layer - seems obvious in retrospect. The best ideas usually do.

References:

1. Wang, B.-S., Wang, C.-Y., & Chiu, W.-C. (2026). MCPNet++: Interpretable Classification Models via Multi-Level Concept Prototypes. IEEE Transactions on Pattern Analysis and Machine Intelligence. DOI: 10.1109/TPAMI.2026.3680506. PubMed: 41931444

2. Wang, B.-S., Wang, C.-Y., & Chiu, W.-C. (2024). MCPNet: An Interpretable Classifier via Multi-Level Concept Prototypes. CVPR 2024. arXiv: 2404.08968

3. Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., & Su, J. (2019). This Looks Like That: Deep Learning for Interpretable Image Recognition. NeurIPS 2019. arXiv: 1806.10574

4. Sacha, M., et al. (2024). An Overview of Prototype Formulations for Interpretable Deep Learning. arXiv: 2410.08925

5. Räuker, T., et al. (2025). A Comprehensive Survey on Self-Interpretable Neural Networks. arXiv: 2501.15638

6. Rymarczyk, D., et al. (2022). Interpretable Image Classification with Differentiable Prototypes Assignment. ECCV 2022.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.