Research map / Architectures and efficiency / RNNs and normalization
L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization
Jiaqi Gu, Hanqing Zhu, Chenghao Feng, Zixuan Jiang, Ray Chen, David Z. Pan
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 |
|---|---|---|---|
| Add feedback sampling (ablation) | L2ight-SL · CIFAR-10, VGG-8 | normalized steps, total: 130.14 -> 73.89 | Improved |
| Add feedback sampling (ablation) | L2ight-SL · CIFAR-10, ResNet-18 | accuracy (%): 92.37$_{\pm0.08}$ -> 91.35$_{\pm0.03}$ | Hurt |
| Add feedback sampling (ablation) | L2ight-SL · CIFAR-10, ResNet-18 | normalized steps, total: 1,969.48 (1.00) -> 1,237.85 (1.59) | Improved |
| Add column sampling after feedback sampling (ablation) | L2ight-SL · CIFAR-10, VGG-8 | normalized PTC energy, total: 29.13 (1.17) -> 19.12 (1.78) | Improved |
| Add column sampling after feedback sampling (ablation) | L2ight-SL · CIFAR-10, VGG-8 | accuracy (%): 86.41$_{\pm0.25}$ -> 85.58$_{\pm0.01}$ | Hurt |
| Add data sampling after feedback and column sampling (ablation) | L2ight-SL · CIFAR-10, VGG-8 | normalized steps, total: 73.07 (1.78) -> 36.54 (3.56) | Improved |
| Add data sampling after feedback and column sampling (ablation) | L2ight-SL · CIFAR-10, VGG-8 | accuracy (%): 85.58$_{\pm0.01}$ -> 84.45$_{\pm0.45}$ | Hurt |
| Use full calibration, mapping, and sparse subspace learning flow (ablation) | L2ight-SL baseline · CIFAR-10, VGG-8 | accuracy (%): 86.66$_{\pm0.13}$ -> 90.20$_{\pm0.05}$ | Improved |
| Use full calibration, mapping, and sparse subspace learning flow (ablation) | L2ight-SL baseline · CIFAR-10, VGG-8 | normalized PTC energy, total: 34.08 (1.00) -> 0.96 (35.64) | Improved |
| Use spatial sampling (RAD) (ablation) | L2ight-SL · CIFAR-10, VGG-8 | normalized steps, total: 130.14 (1.00) -> 130.14 (1.00) | No clear effect |
| Use SWAT-U weight and spatial sampling (ablation) | L2ight-SL · CIFAR-10, VGG-8 | accuracy (%): 86.66$_{\pm0.13}$ -> 73.91$_{\pm0.27}$ | Hurt |
| Use balanced top-K feedback sampling rather than uniform or top-K sampling (ablation) | L2ight sparse subspace learning · feedback gradient approximation | gradient angular similarity | Improved |
Closest papers by meaning
- Neural Tangent Knowledge Distillation for Optical Convolutional Networks (NeurIPS 2025 · similarity 0.87)
- Physics-Constrained Comprehensive Optical Neural Networks (NeurIPS 2024 · similarity 0.85)
- PACE: Pacing Operator Learning to Accurate Optical Field Simulation for Complicated Photonic Devices (NeurIPS 2024 · similarity 0.85)
- Layered-Parameter Perturbation for Zeroth-Order Optimization of Optical Neural Networks (AAAI 2025 · similarity 0.85)
- All-Optical Nonlinear Diffractive Deep Network for Ultrafast Image Denoising (CVPR 2025 · similarity 0.81)
- Towards training digitally-tied analog blocks via hybrid gradient computation (NeurIPS 2024 · similarity 0.81)
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.
