Research map / Label-efficient and robust learning / Recognition and long-tail
Finding NEM-U: Explaining unsupervised representation learning through neural network generated explanation masks
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.
Introduces NEM (Neural Explanation Masks), a framework that augments a frozen representation model with a masking U-Net to produce occlusion-based explanations in a single forward pass, instantiated as NEM-U.
Problem
Explaining which input regions matter for a learned vector embedding from unsupervised or supervised representation models; existing methods are either inaccurate or too slow for practical use.
Approach
Train a U-Net masking network on top of a frozen encoder, reusing its multi-level features via skip connections, to predict occlusion masks that preserve the embedding while being sparse and binary; explanations are then produced in parallel with inference rather than by per-sample optimization.
Key results
- NEM-U achieves about 0.002 s per image versus 2.048 s for RELAX (roughly 1000x faster) on an RTX 4090.
- NEM-U has the highest relevance mass on all datasets and models (e.g., 0.823 vs 0.679 for U-RELAX on supervised VOC) and the lowest complexity in most settings.
- RELAX attains the highest relevance rank in most settings (e.g., 0.761 vs 0.738 for NEM-U on DINO VOC), so NEM-U is not best on that locality metric.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NEM-U instead of RELAX | Representation explanations for supervised ResNet50, DINO, SwAV, and SimCLR · VOC and COCO | relevance mass: VOC: 0.599 -> 0.823; 0.642 -> 0.749; 0.589 -> 0.736; 0.612 -> 0.710. COCO: 0.556 -> 0.697; 0.592 -> 0.688; 0.544 -> 0.64 | Helped |
| Use NEM-U instead of RELAX | Representation explanations for supervised ResNet50, DINO, SwAV, and SimCLR · VOC and COCO | relevance rank: VOC: 0.719 -> 0.691; 0.761 -> 0.738; 0.721 -> 0.686; 0.709 -> 0.692. COCO: 0.689 -> 0.644; 0.725 -> 0.638; 0.685 -> 0.58 | Mixed |
| Use NEM-U instead of RELAX | Explanation generation · COCO and VOC evaluation images | time per image (seconds): 2.048 -> 0.002 | Helped |
| Use NEM-U instead of other evaluated explanation methods | Representation explanations · Evaluated models and datasets | complexity | Helped |
Surprising findings
- NEM-U is better at relevance mass but worse at relevance rank than RELAX, suggesting its set-of-features explanations are less informative for pixel ranking.
Limitations
- Requires training a masking network per explained model, so it is efficient only when many explanations are needed.
- Lower relevance rank than RELAX; explanations are set-of-features rather than additive, making comparisons difficult.
- Masks are dataset-biased and denser on COCO than VOC.
Experimental setup
Datasets: COCO VOC
Benchmarks: COCO VOC
Metrics: relevance rank accuracy relevance mass accuracy complexity sparseness runtime
Baselines: RELAX U-RELAX Integrated Gradients Gradient SHAP Saliency
Builds on: RELAX rate-distortion explanation U-Net label-free XAI signal-distractor framework
Closest papers by meaning
- UNR-Explainer: Counterfactual Explanations for Unsupervised Node Representation Learning Models (ICLR 2024 · similarity 0.86)
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.86)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.85)
- Understanding Distributed Representations of Concepts in Deep Neural Networks without Supervision (AAAI 2024 · similarity 0.85)
- CoSy: Evaluating Textual Explanations of Neurons (NeurIPS 2024 · similarity 0.85)
- Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.