This repository contains the code used for training and analyzing the Single-Photon-Detection Neural Networks (SPDNNs). These networks have been detailed in our study "Quantum-noise-limited optical neural networks operating at a few quanta per activation".
In essence, SPDNNs are uniquely designed to detect only a few photons with each dot product or neuron activation during inference. The models trained here can be integrated on various Optical Neural Network (ONN) platforms equipped with single-photon detectors (SPDs). In order to include the stochastic single-photon detection process during training, we employ "physics-aware stochastic training" to incorporate the actual stochastic physical process forward pass, while eliminate the stochasicity in the backward pass. This is inspired by the training methods for binary stochastic neurons (see Bengio et al. (2013)).
This repository includes several directories, each containing specific resources:
The 'src' directory contains Python scripts that define the photon-detection activation function and the SPDNN models.
In the 'training' directory, you will find Jupyter notebooks that provide examples of training SPDNN models.
Trained models from the notebooks are stored in the 'models' directory.
In the 'test' directory, you will find Jupyter notebooks to test the performance of the trained SPDNN models.
Processed test accuracies and intermediate data are saved in the 'results' directory.
The 'data' directory houses external data used for training.
All generated plots are stored in the 'plots' directory.
Other utility functions and files can be found in the 'misc' directory.
Refer to the requirements.txt file for necessary Python packages. Use the following commands to set up the environment with conda:
conda create --name spdnn python=3.9 pip
conda activate spdnn
pip3 install -r requirements.txt
784-
- MNIST test accuracies for SPDNNs using coherent optical matrix-vector multiplers where the information is encoded in optical amplitude, and directly detected by SPDs.
| Model | |||||||
|---|---|---|---|---|---|---|---|
| 784-10-10 | 80.13±0.32% | 85.65±0.23% | 87.52±0.20% | 88.86±0.18% | 89.40±0.15% | 89.87±0.15% | 90.70±0.00% |
| 784-20-10 | 89.32±0.21% | 92.46±0.17% | 93.35±0.17% | 94.09±0.15% | 94.40±0.12% | 94.62±0.11% | 95.05±0.00% |
| 784-25-10 | 91.50±0.22% | 94.14±0.15% | 94.89±0.14% | 95.48±0.10% | 95.69±0.10% | 95.86±0.10% | 96.10±0.00% |
| 784-30-10 | 92.79±0.20% | 94.92±0.13% | 95.56±0.12% | 96.06±0.11% | 96.26±0.09% | 96.40±0.10% | 96.74±0.00% |
| 784-50-10 | 95.51±0.15% | 96.89±0.11% | 97.25±0.09% | 97.55±0.08% | 97.66±0.08% | 97.75±0.07% | 97.93±0.00% |
| 784-100-10 | 97.40±0.12% | 98.15±0.10% | 98.36±0.07% | 98.52±0.07% | 98.58±0.06% | 98.62±0.06% | 98.70±0.00% |
| 784-200-10 | 98.32±0.10% | 98.77±0.07% | 98.89±0.06% | 98.97±0.06% | 99.01±0.05% | 99.03±0.05% | 99.12±0.00% |
| 784-400-10 | 98.64±0.08% | 98.95±0.06% | 99.03±0.05% | 99.10±0.05% | 99.12±0.05% | 99.13±0.04% | 99.19±0.00% |
| 784-400-400-10 | 98.94±0.07% | 99.22±0.06% | 99.29±0.05% | 99.33±0.04% | 99.36±0.04% | 99.37±0.04% | 99.40±0.00% |
| 784-C16-400-10 | 99.33±0.06% | 99.45±0.04% | 99.47±0.04% | 99.50±0.04% | 99.51±0.03% | 99.52±0.03% | 99.54±0.00% |
- MNIST test accuracies for SPDNNs using incoherent optical matrix-vector multiplers where the information is encoded in light intensity (non-negative).
| Model | |||||||
|---|---|---|---|---|---|---|---|
| 784-10-10 | 78.04±0.29% | 83.23±0.23% | 84.78±0.22% | 86.12±0.16% | 86.64±0.17% | 87.09±0.14% | 87.91±0.00% |
| 784-20-10 | 86.73±0.23% | 89.97±0.19% | 90.97±0.14% | 91.69±0.14% | 92.01±0.12% | 92.21±0.13% | 92.66±0.00% |
| 784-50-10 | 93.01±0.16% | 94.47±0.13% | 94.93±0.13% | 95.23±0.09% | 95.38±0.09% | 95.48±0.07% | 95.73±0.00% |
| 784-100-10 | 95.18±0.15% | 96.22±0.10% | 96.53±0.09% | 96.77±0.09% | 96.84±0.08% | 96.90±0.08% | 97.02±0.00% |
| 784-200-10 | 96.61±0.13% | 97.33±0.09% | 97.54±0.09% | 97.70±0.08% | 97.76±0.07% | 97.80±0.05% | 97.98±0.00% |
| 784-300-10 | 96.99±0.13% | 97.62±0.10% | 97.77±0.09% | 97.92±0.06% | 97.97±0.06% | 98.01±0.07% | 98.12±0.00% |
| 784-400-10 | 97.30±0.09% | 97.85±0.09% | 98.01±0.08% | 98.15±0.06% | 98.21±0.06% | 98.26±0.05% | 98.41±0.00% |
Our work owes a lot to earlier studies on binary stochastic neurons, the construction of this repository was inspired by other works, including but not limited to:
- A GitHub repository that constructs binary stochastic neurons using PyTorch.
- A comprehensive article introducing binary stochastic neurons and showcasing preliminary results.
If you find our data or code useful for your research, please consider citing the following paper:
S.-Y. Ma, T. Wang, J. Laydevant, L. G. Wright and P. L. McMahon. "Quantum-noise-limited optical neural networks operating at a few quanta per activation." arXiv:2307.15712 (2023)
The code in this repository is released under the following license:
Creative Commons Attribution 4.0 International
A copy of this license is given in this repository as license.txt.