Research map / Architectures and efficiency / RNNs and normalization
Sparsifying Neural Network Connections for Face Recognition
Check 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 |
|---|---|---|---|
| Sparsify layers f, 5b, and 4b using correlation-guided selection and dense-model initialization | VGG-like baseline ConvNet · LFW face verification | accuracy: 0.9895 -> 0.9930 | Helped |
| Combine features from 25 sparse ConvNets | Sparse ConvNet · LFW face verification | accuracy: 0.9947 -> 0.9955± 0.0010 | Helped |
| Use a single sparse ConvNet instead of a single DeepID2+ net | Face verification system · YouTube Faces | accuracy: 91.9% -> 92.7% | Helped |
| Replace the DeepID2+ ensemble with 25 sparse ConvNets | Face verification system · YouTube Faces | accuracy: 93.2% -> 93.5% | Helped |
| Add sparsity to layer f at 1/256 (ablation) | VGG-like baseline ConvNet · LFW face verification | accuracy: 0.9895 -> 0.9898 | Helped |
| Additionally sparsify layer 5b to 1/128 (ablation) | Sparse ConvNet (1/256-f) · LFW face verification | accuracy: 0.9898 -> 0.9923 | Helped |
| Sparsify layer 4b to 1/2 after sparsifying f and 5b (ablation) | Sparse ConvNet (1/256-f, 1/128-5b) · LFW face verification | accuracy: 0.9923 -> 0.9930 | Helped |
| Sparsify layer 5a to 1/32 (ablation) | Sparse ConvNet (1/256-f, 1/128-5b) · LFW face verification | accuracy: 0.9895 -> 0.9898 | No clear effect |
| Vary layer 5a sparsity from 1 to 1/32 (ablation) | Sparse ConvNet (1/256-f, 1/128-5b) · LFW face verification | accuracy: 0.9923, 0.9922, 0.9918, 0.9908, 0.9890, 0.9898 | Mixed |
| Select connections randomly rather than by neural correlation (ablation) | Sparse ConvNet (1/256-f) · LFW face verification | accuracy: 0.9898 -> 0.9893 | Hurt |
| Keep only the highest-correlation connections, removing the selected low-correlation portion (ablation) | Sparse ConvNet (1/256-f, 1/128-5b) · LFW face verification | accuracy: 0.9923 -> 0.9902 | Hurt |
| Train the sparse model from random initialization instead of initializing from the previously learned denser model (ablation) | Sparse ConvNet (1/256-f, 1/128-5b) · LFW face verification | accuracy: 0.9923 -> 0.9845 | Hurt |
Closest papers by meaning
- Revisiting Sparse Convolutional Model for Visual Recognition (NeurIPS 2022 · similarity 0.87)
- Sparsity May Cry: Let Us Fail (Current) Sparse Neural Networks Together! (ICLR 2023 · similarity 0.87)
- Sparse-IFT: Sparse Iso-FLOP Transformations for Maximizing Training Efficiency (ICML 2024 · similarity 0.85)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · similarity 0.85)
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (ICLR 2024 · similarity 0.85)
- Sparse Cocktail: Every Sparse Pattern Every Sparse Ratio All At Once (ICML 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.