Source-linked AI summary
This Looks Like That: Deep Learning for Interpretable Image Recognition
Chaofan Chen, Oscar Li, Chaofan Tao, Alina Jade Barnett, Jonathan Su, Cynthia Rudin
TL;DR
Challenging image classification often relies on comparing image parts with prototypical class examples, but models need a human-understandable reasoning process. ProtoPNet builds those comparisons into a neural network, and achieves accuracy comparable to analogous non-interpretable models, with combined networks reaching performance on par with strong deep models.
Problem
Image-classification models lack a reasoning process that explains predictions through human-like comparisons between image parts and prototypical class examples.
Method
ProtoPNet learns visual prototypes corresponding to prototypical image parts and bases predictions on weighted similarity scores between those prototypes and input-image regions.
Results
ProtoPNet achieves accuracy comparable to its analogous non-interpretable counterpart, while combined ProtoPNet networks reach accuracy on par with some best-performing deep models.
Takeaways & Limitations
Prototype comparisons are integral to ProtoPNet’s transparent reasoning process, providing interpretability that the authors report as absent in other interpretable deep models.
Takeaways & Limitations
The authors do not quantitatively evaluate unit interpretability because existing fine-grained concept-labeling datasets are unsuitable for their main application.
Abstract
from arXiv · showhide
When we are faced with challenging image classification tasks, we often explain our reasoning by dissecting the image, and pointing out prototypical aspects of one class or another. The mounting evidence for each of the classes helps us make our final decision. In this work, we introduce a deep network architecture -- prototypical part network (ProtoPNet), that reasons in a similar way: the network dissects the image by finding prototypical parts, and combines evidence from the prototypes to make a final classification. The model thus reasons in a way that is qualitatively similar to the way ornithologists, physicians, and others would explain to people on how to solve challenging image classification tasks. The network uses only image-level labels for training without any annotations for parts of images. We demonstrate our method on the CUB-200-2011 dataset and the Stanford Cars dataset. Our experiments show that ProtoPNet can achieve comparable accuracy with its analogous non-interpretable counterpart, and when several ProtoPNets are combined into a larger network, it can achieve an accuracy that is on par with some of the best-performing deep models. Moreover, ProtoPNet provides a level of interpretability that is absent in other interpretable deep models.
1 Introduction
ProtoPNet defines an interpretable image-classification approach that mirrors human case-based reasoning by comparing image parts with learned prototypes. Its built-in comparisons support transparent predictions while retaining accuracy comparable to non-interpretable models.
- Motivation: ProtoPNet imitates human classification by comparing image parts with prototypical parts from candidate classes.The motivation draws on difficult identification tasks such as bird classification and medical diagnosis.
- Approach: The network identifies image regions resembling learned class prototypes and combines their similarity scores to make predictions.Prototype comparisons are integral to the model’s reasoning process rather than added afterward.
- Interpretability: ProtoPNet provides built-in case-based reasoning, unlike posthoc visualization methods that do not explain how a CNN actually makes decisions.Attention-based models expose focused regions but do not identify prototypical cases to which those regions are similar.
- Scope: The authors do not quantitatively measure unit interpretability because existing fine-grained concept-labeling resources are unsuitable for their application.Instead, they evaluate whether the network’s reasoning process is qualitatively similar to human reasoning.
- Related work: ProtoPNet uses a specialized neural architecture that jointly learns feature extraction and prototypes end-to-end, unlike traditional bag-of-visual-words systems.Traditional systems use SIFT and learn visual words separately from feature extraction and the final classifier.
2 Case study 1: bird species identification
ProtoPNet is a bird-species classifier that reasons by comparing image parts with learned class prototypes, producing predictions from prototype similarities while exposing the comparison process. Its architecture and training procedure support prototype-based explanations, and experiments show competitive accuracy with added interpretability.
- Case study setup: ProtoPNet was trained and evaluated for 200-species bird identification on CUB-200-2011 using cropped images and image-level class labels.The case study used offline augmentation and dataset-provided bounding boxes.
- Architecture: The network combines a convolutional feature extractor, prototype layer, and bias-free fully connected layer to produce class probabilities.Prototype similarities are multiplied by the final-layer weights and normalized with softmax.
- Architecture: Each prototype represents a latent prototypical image part, while prototype units compare it with same-shaped feature patches and convert distances into similarity maps.Global max pooling yields one score per prototype, and upsampled maps localize the most similar input region.
- Reasoning process: Class-specific prototypes provide evidence for predictions, with each class assigned a predetermined number of prototypes and logits formed from their weighted similarity scores.The reasoning process compares test-image patches with prototypes for each class before producing the final prediction.
- Training: Training alternates SGD for pre-final layers, prototype projection, and convex optimization of the last layer to organize class-relevant latent patches around prototypes.A theorem states that sufficiently small prototype movements preserve confident correct predictions after projection.
- Results and interpretability: ProtoPNet’s cropped-bird accuracy was at most 3.5% lower than the corresponding non-interpretable baseline, while combined ProtoPNets reached 84.8%.The model provides part-level attention together with similar prototypical cases, unlike attention-only explanations.
- Results and interpretability: The comparison shows a remaining accuracy gap from the best state-of-the-art models, although the authors identify richer explanations and transparency as benefits.This interpretation is tied to the model’s prototypes participating directly in classification computation.
3 Case study 2: car model identification
ProtoPNet was applied to Stanford Cars for identifying 196 car models, achieving accuracy comparable to its baseline and strong performance when combining networks.
- ProtoPNet was trained on the Stanford Cars dataset for identifying 196 car models.
- The ProtoPNet test accuracy differed by no more than 3% from the corresponding baseline model.
- 91.4% accuracy was achieved by combining VGG19-, ResNet34-, and DenseNet121-based ProtoPNets.
- The combined network’s accuracy was on par with B-CNN at 91.3%, RA-CNN at 92.5%, and MA-CNN at 92.8%.
4 Conclusion
The paper defines a human-aligned form of interpretability and presents ProtoPNet as a network architecture and training method applied to bird and car identification.
- The paper defines image-processing interpretability as reasoning that agrees with how humans describe classification decisions.
- ProtoPNet accommodates this interpretability through a network architecture and specialized training algorithm.
- The technique was applied to bird species and car model identification.
- Supplementary material and code are available through the project’s GitHub repository.