Research map / Architectures and efficiency / RNNs and normalization
A Framework to Learn with Interpretation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Jointly train the predictor with the interpretation losses | FLINT predictor · MNIST, FashionMNIST, CIFAR10, QuickDraw | accuracy (%): BASE-f: 98.9$\pm$0.1, 90.4$\pm$0.1, 84.7$\pm$0.3, 85.3$\pm$0.2; FLINT-f: 98.9$\pm$0.2, 90.5$\pm$0.2, 84.5$\pm$0.2, 85.7$ | No clear effect |
| Use FLINT interpreter instead of VIBI | FLINT-g versus VIBI · MNIST, FashionMNIST, CIFAR10, QuickDraw | fidelity (%): 96.6$\pm$0.7 -> 98.7$\pm$0.1; 88.4$\pm$0.3 -> 91.5$\pm$0.1; 65.5$\pm$0.3 -> 93.2$\pm$0.2; 78.6$\pm$0.4 -> 90.8$\pm$0.4 | Helped |
| Add entropy-based losses to the conciseness objective (ablation) | FLINT interpreter · QuickDraw | conciseness | Helped |
| Add entropy losses at ℓ1 strength η=1 (ablation) | FLINT interpreter · QuickDraw | fidelity (%): 92.7 -> 91.2 | Hurt |
| Shuffle attribute activations (ablation) | FLINT-g · MNIST, FashionMNIST, CIFAR10, QuickDraw | accuracy drop (%): 87.6; 76.6; 74.4; 74.9 | Hurt |
| Shuffle attribute activations (ablation) | SENN · MNIST, FashionMNIST, CIFAR10, QuickDraw | accuracy drop (%): 0.5; 10.9; 17.5; 0.3 | No clear effect |
| Select the 16th convolutional layer rather than the 9th (ablation) | FLINT with ResNet18 · QuickDraw | fidelity (%): 78.0 -> 96.0 | Helped |
| Reduce the attribute dictionary from J=25 to J=4 (ablation) | FLINT with LeNet · MNIST | fidelity (%): 98.8 -> 87.4 | Hurt |
| Reduce the attribute dictionary from J=24 to J=4 (ablation) | FLINT with ResNet18 · QuickDraw | fidelity (%): 90.8 -> 19.5 | Hurt |
| Introduce output-fidelity and conciseness losses very late (Epoch 24) instead of Epoch 3 (ablation) | FLINT with ResNet18 · CIFAR10 | fidelity (%): 93.5 -> 86.3 | Hurt |
| Introduce output-fidelity and conciseness losses at nearby earlier epochs (ablation) | FLINT with ResNet18 · CIFAR10 | fidelity (%): Epoch 3: 93.5; Epoch 4: 93.4; Epoch 5: 94.2; Epoch 6: 93.1; Epoch 8: 93.7 | No clear effect |
| Remove the autoencoder input-fidelity loss (ablation) | FLINT interpreter · image datasets | interpretability | Hurt |
Closest papers by meaning
- Intrinsic User-Centric Interpretability through Global Mixture of Experts (ICLR 2025 · similarity 0.88)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.87)
- Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants (NeurIPS 2026 · similarity 0.86)
- A Multimodal Automated Interpretability Agent (ICML 2024 · similarity 0.86)
- Measuring Per-Unit Interpretability at Scale Without Humans (NeurIPS 2024 · similarity 0.85)
- Enhancing Performance of Explainable AI Models with Constrained Concept Refinement (ICML 2025 · 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.