Research map / Learning theory and algorithms / Prediction and decision losses

Rashomon Sets for Prototypical-Part Networks: Editing Interpretable Models in Real-Time

CVPR 2025 · Learning theory and algorithms › Prediction and decision losses · approach: Theory of neural network sensitivity · 4 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.

Proto-RSet approximates the Rashomon set of a ProtoPNet's last linear layer around a fixed reference model, letting users edit prototypes (remove or require them) in seconds with guaranteed training-accuracy and constraint satisfaction.

Problem

Fixing flawed prototypes in ProtoPNets requires slow, uncertain retraining, the so-called "interaction bottleneck" between domain experts and ML practitioners.

Approach

Fix the backbone and prototype layer, then approximate the Rashomon set over the final logistic-regression layer with a second-order Taylor expansion, yielding an ellipsoid whose geometry supports fast sampling, prototype removal (hyperplane intersection), and constraint checks without retraining.

Key results

• Proto-RSet computes a Rashomon set in under 20 minutes across 18 dataset-backbone settings and removes a prototype in seconds, orders of magnitude faster than retraining baselines.
• Across six backbones and three datasets, Proto-RSet maintains test accuracy as up to 100 random prototypes are removed, and guarantees removed prototypes get zero weight.
• In a 31-participant user study on CUB-200 confounders, users corrected models with Proto-RSet in 2.1 minutes versus 93.7 minutes for ProtoPDebug with similar accuracy change (-0.5% vs +0.8%).

Measured effects reported in the paper

Surprising findings

• Ignoring Proto-RSet's warning that a prototype cannot be removed dropped skin cancer model test accuracy from 70.4% to 57.9%.
• Removing a prototype often substantially changes the reasoning of the model; the weight of another prototype roughly doubles to compensate.

Limitations

• Proto-RSet is an approximation, not the entire Rashomon set of ProtoPNets.
• Requiring a prototype to have a minimum coefficient makes the feasible set non-ellipsoidal, complicating further operations.
• Restricted to positive reasoning and a fixed backbone and prototype layer.

Experimental setup

Datasets: CUB-200 Stanford Cars Stanford Dogs HAM10000

Benchmarks: CUB-200 Stanford Cars Stanford Dogs HAM10000

Metrics: test accuracy runtime per prototype removal l1 norm of removed prototype coefficients validation accuracy change time to correct model

Baselines: naive prototype removal naive prototype removal with retraining hard removal ProtoPDebug

Builds on: ProtoPNet ProtoPDebug ProtoPNeXt Rashomon set estimation for GAMs

Closest papers by meaning

• Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.87)
• LucidPPN: Unambiguous Prototypical Parts Network for User-centric Interpretable Computer Vision (ICLR 2025 · similarity 0.87)
• Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.86)
• Post-hoc Part-Prototype Networks (ICML 2024 · similarity 0.85)
• ProtoArgNet: Interpretable Image Classification with Super-Prototypes and Argumentation (AAAI 2025 · similarity 0.85)
• Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning (AAAI 2024 · similarity 0.84)