Research map / Optimization / Optimizers
High-Performance Real-World Optical Computing Trained by in Situ Gradient-Based Model-Free Optimization
Guangyuan Zhao, Xin Jian Shu, Renjie Zhou
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 |
|---|---|---|---|
| use G-MFO instead of HBT for training under misalignment | two-layer optical system · MNIST (two-layer simulation) | test accuracy: 78.4% -> 87.0% | Improved |
| use G-MFO instead of HBT for training under misalignment | two-layer optical system · FMNIST (two-layer simulation) | test accuracy: 59.8% -> 74.1% | Improved |
| use G-MFO instead of HBT on the real single-layer system | single-layer optical system · MNIST (real single-layer experiment) | test accuracy: 72.8% -> 73.6% | Improved |
| use G-MFO instead of HBT on the real single-layer system | single-layer optical system · FMNIST (real single-layer experiment) | test accuracy: 65.8% -> 70.4% | Improved |
| fine-tune the HBT result with G-MFO | single-layer optical system · MNIST (real single-layer experiment) | test accuracy: 72.8% -> 77.7% | Improved |
| fine-tune the HBT result with G-MFO | single-layer optical system · FMNIST (real single-layer experiment) | test accuracy: 65.8% -> 70.2% | No clear effect |
| use HBT instead of SBT (in situ forward pass) | single-layer optical system · MNIST (real single-layer experiment) | test accuracy: 69.3% -> 72.8% | Improved |
| use two-point ZO estimation instead of G-MFO sampling (ablation) | zeroth-order optimizer · FMNIST (simulated single-layer, 100 samples) | train accuracy: 99.0% -> 22.0% | Hurt |
| use G-MFO instead of full-point ZO | zeroth-order optimizer · FMNIST (simulated single-layer, 100 samples) | training time per epoch: 113 min -> 2 min | Improved |
| increase sampling size M from 32 to 128 (ablation) | G-MFO · MNIST simulation (1,000 samples) | training accuracy: +2.2% | Improved |
| increase sampling size M from 128 to 256 (ablation) | G-MFO · MNIST simulation (1,000 samples) | training accuracy: +0.5% | No clear effect |
| increase trainable parameters H beyond 128^2 at M=128 (ablation) | G-MFO · MNIST simulation (1,000 samples) | training accuracy: H > 128^2 with M=128 | Hurt |
Closest papers by meaning
- Optical Diffusion Models for Image Generation (NeurIPS 2024 · similarity 0.83)
- OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit (CVPR 2025 · similarity 0.82)
- Neural Tangent Knowledge Distillation for Optical Convolutional Networks (NeurIPS 2025 · similarity 0.82)
- Layered-Parameter Perturbation for Zeroth-Order Optimization of Optical Neural Networks (AAAI 2025 · similarity 0.82)
- Physics-Constrained Comprehensive Optical Neural Networks (NeurIPS 2024 · similarity 0.81)
- All-Optical Nonlinear Diffractive Deep Network for Ultrafast Image Denoising (CVPR 2025 · 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.
