[Paper Review] Deep Learning for Case-Based Reasoning through Prototypes: A Neural Network that Explains Its Predictions

Oscar Li, Hao Liu+2 more|arXiv (Cornell University)|Oct 13, 2017

Explainable Artificial Intelligence (XAI)Computer Science158 citations

TL;DR

Introduces an interpretable neural network that combines an autoencoder with a prototype-based layer to explain predictions via learned prototypes in latent space.

ABSTRACT

Deep neural networks are widely used for classification. These deep models often suffer from a lack of interpretability -- they are particularly difficult to understand because of their non-linear nature. As a result, neural networks are often treated as "black box" models, and in the past, have been trained purely to optimize the accuracy of predictions. In this work, we create a novel network architecture for deep learning that naturally explains its own reasoning for each prediction. This architecture contains an autoencoder and a special prototype layer, where each unit of that layer stores a weight vector that resembles an encoded training input. The encoder of the autoencoder allows us to do comparisons within the latent space, while the decoder allows us to visualize the learned prototypes. The training objective has four terms: an accuracy term, a term that encourages every prototype to be similar to at least one encoded input, a term that encourages every encoded input to be close to at least one prototype, and a term that encourages faithful reconstruction by the autoencoder. The distances computed in the prototype layer are used as part of the classification process. Since the prototypes are learned during training, the learned network naturally comes with explanations for each prediction, and the explanations are loyal to what the network actually computes.

Motivation & Objective

• Motivate the need for interpretable predictions in deep learning and address the lack of interpretability in standard neural networks.
• Propose a neural architecture that integrates an autoencoder with a prototype layer to provide case-based explanations.
• Enable visualization of learned prototypes by decoding latent-space prototypes back to input space.
• Ensure the model maintains competitive predictive performance while enhancing interpretability through dedicated regularization terms.

Proposed method

• Two-component architecture: an autoencoder (encoder f and decoder g) and a prototype classification network h in latent space.
• Prototype layer p computes squared L2 distances between encoded input z=f(x) and m prototypes p1,...,pm in R^q; a fully connected layer W combines distances into class logits, followed by softmax.
• Training objective combines cross-entropy loss E, reconstruction loss R, and two interpretability regularizers R1 and R2, plus a hyperparameterized total loss L = E(h∘f,D) + λR(g∘f,D) + λ1R1(...) + λ2R2(...).
• R1 encourages each prototype to be close to at least one encoded input in latent space; R2 encourages every encoded input to be close to at least one prototype.
• Prototype vectors live in the latent space, enabling visualization by decoding into input space; W can be learned to reflect prototype-to-class relationships.

Experimental results

Research questions

• RQ1 Can a neural network be designed to explain its predictions via case-based reasoning using learned prototypes in a latent space?
• RQ2 Do latent-space prototypes, together with explicit regularizers, yield meaningful and visualizable explanations without sacrificing accuracy?
• RQ3 How do the interpretability terms R1 and R2 influence prototype quality and generalization across datasets?
• RQ4 What is the impact of learning the prototype-to-class weight matrix W on classification behavior and interpretability?
• RQ5 How does the architecture perform on standard image classification benchmarks compared to non-interpretable networks?

Key findings

• The model achieves competitive accuracy on MNIST (train 99.53%, test 99.22%), Fashion-MNIST (89.95%), and Cars dataset while providing intrinsic explanations via prototypes.
• Decoded prototypes visually resemble realistic digits and clothing items, demonstrating meaningful latent-space representations enabled by R1 and R2.
• Ablation studies indicate that removing the prototype layer or decoder yields accuracies similar to non-interpretable baselines, showing interpretability does not drastically reduce performance in these tasks.
• The learned weight matrix W reveals which prototypes most influence each class, offering insight into class relationships and prototype utility.
• Prototype visualizations illustrate intra-class variation (e.g., different handwriting styles for 6 and 3) and cross-class ambiguities, aligning with case-based reasoning.