Research map / Architectures and efficiency / RNNs and normalization
DISCOVER: Making Vision Networks Interpretable via Competition and Dissection
Konstantinos P. Panousis, Sotirios Chatzis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Replace GELU with stochastic LWTA competition using 2 competitors | DeiT-T · ImageNet-1k | accuracy: 72.2 -> 72.5 | Improved |
| Replace GELU with stochastic LWTA competition using 16 competitors | DeiT-T · ImageNet-1k | accuracy: 72.2 -> 71.1 | Hurt |
| Replace GELU with stochastic LWTA competition using 2 competitors | DeiT-S · ImageNet-1k | accuracy: 77.0 -> 77.3 | Improved |
| Replace ReLU with stochastic LWTA competition using 2 competitors | ResNet-18 · Places365 | accuracy: 52.25 -> 53.9 | Improved |
| Use stochastic LWTA competition with 16 competitors instead of GELU | DeiT-T · ImageNet Val, Broden concept set | CLIP cosine similarity: 0.6250 -> 0.6460 | Improved |
| Use CosSimCubed instead of SoftWPMI for concept matching (ablation) | DeiT-T/16 · ImageNet Val, Broden concept set | CLIP cosine similarity: 0.6460 -> 0.6543 | Improved |
| Use stochastic LWTA competition with 16 competitors instead of GELU | DeiT-T · ImageNet Val, 3k concept set | CLIP cosine similarity: 0.6226 -> 0.6650 | Improved |
| Use CosSimCubed instead of SoftWPMI for concept matching (ablation) | DeiT-T/16 · ImageNet Val, 3k concept set | CLIP cosine similarity: 0.6650 -> 0.6699 | Improved |
| Use stochastic LWTA competition with 8 competitors instead of GELU | DeiT-T · CIFAR-100 Train, ImageNet concept set | neuron identification accuracy: 53.00% -> 53.10% | No clear effect |
| Use stochastic LWTA competition with 12 competitors instead of GELU | DeiT-S · CIFAR-100 Train, ImageNet concept set | neuron identification accuracy: 53.00% -> 56.10% | Improved |
| Use stochastic LWTA competition instead of conventional GELU | DeiT-T/16 · Average neuron identification accuracy | neuron identification accuracy: 78.00% -> 78.30% | Improved |
| Use stochastic LWTA competition instead of conventional GELU | DeiT-S/2 · Average neuron identification accuracy | neuron identification accuracy: 78.00% -> 78.70% | Improved |
Closest papers by meaning
- NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning (NeurIPS 2026 · similarity 0.88)
- LINE: LLM-based Iterative Neuron Explanations for Vision Models (NeurIPS 2026 · similarity 0.88)
- A Convolutional Neural Network Interpretable Framework for Human Ventral Visual Pathway Representation (AAAI 2024 · similarity 0.87)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.87)
- Automated Natural Language Explanation of Deep Visual Neurons with Large Models (Student Abstract) (AAAI 2024 · similarity 0.87)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.87)
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.
