Research map / Architectures and efficiency / CNN design
Surround Modulation: A Bio-inspired Connectivity Structure for Convolutional Neural Networks
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 |
|---|---|---|---|
| add surround modulation kernel to half of first-layer activation maps | baseline CNN · baseline-ImageNet (100 classes) test | top-1 accuracy: 41.0±0.6 -> 43.8±0.7 | Helped |
| add learnable 5x5 conv layer (no ReLU) instead of SM (ablation) | baseline CNN · baseline-ImageNet (100 classes) test | top-1 accuracy: 41.0±0.6 -> 41.3±0.6 | No clear effect |
| add learnable 5x5 conv layer with ReLU instead of SM (ablation) | baseline CNN · baseline-ImageNet (100 classes) test | top-1 accuracy: 41.0±0.6 -> 40.7±0.5 | No clear effect |
| replace SM kernel with random kernel of matching first/second-order statistics (ablation) | SM-CNN · baseline-ImageNet (100 classes) test | top-1 accuracy: 43.8±0.7 -> 41.0±0.9 | Hurt |
| apply SM kernel to input image only instead of first-layer maps (ablation) | SM-CNN · baseline-ImageNet (100 classes) test | top-1 accuracy: 43.8±0.7 -> 40.8±0.4 | Hurt |
| apply SM to all first-layer maps instead of half (ablation) | SM-CNN · baseline-ImageNet (100 classes) test | top-1 accuracy: 43.8±0.7 -> 42.3±0.6 | Hurt |
| apply SM after first max-pooling layer instead of first layer (ablation) | SM-CNN · baseline-ImageNet (100 classes) test | top-1 accuracy: 43.8±0.7 -> 43.2±0.4 | No clear effect |
| add SM to first conv layer | baseline CNN · small NORB unfamiliar lighting (trained light0) | accuracy: 57.4±1.5 -> 80.8±1.2 | Helped |
| add SM instead of L2, Dropout or BatchNorm regularization (ablation) | baseline CNN · small NORB unfamiliar lighting (trained light2) | accuracy: 35.1±2.1 -> 61.4±2.5 | Helped |
| add SM to first conv layer | baseline CNN · MNIST-1000 occluded test | accuracy: 54.8±1.1 -> 67.0±1.1 | Helped |
| add SM to first conv layer | baseline CNN · MNIST-1000 standard test | accuracy: 92.6±0.3 -> 93.2±0.2 | Helped |
| add SM to first conv layer | baseline CNN · baseline-ImageNet conv layer 3 | kurtosis sparsity: 11.6±1.1 -> 28.2±3.3 | Helped |
Closest papers by meaning
- Self-Attention-Based Contextual Modulation Improves Neural System Identification (ICLR 2025 · similarity 0.88)
- A Unified Theory of Early Visual Representations from Retina to Cortex through Anatomically Constrained Deep CNNs (ICLR 2019 · similarity 0.88)
- Deep Learning Models of the Retinal Response to Natural Scenes (NeurIPS 2016 · similarity 0.87)
- Flexible Context-Driven Sensory Processing in Dynamical Vision Models (NeurIPS 2024 · similarity 0.87)
- Convolution Goes Higher-Order: A Biologically Inspired Mechanism Empowers Image Classification (NeurIPS 2025 · similarity 0.85)
- Long-Range Feedback Spiking Network Captures Dynamic and Static Representations of the Visual Cortex under Movie Stimuli (NeurIPS 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.