Classification of Single Photons in Higher-Order Spatial Modes via Convolutional Neural Networks

Manon P. Bart, Sita Dawanse, Nicholas J. Savino, Viet Train, Tianhong Wang, Sanjaya Lohani, Farris Nefissi, Pascal Bassene, Moussa N’Gom, Ryan T. Glasser

​

​

Quantum Information and Nonlinear Optics Group

Dr. Ryan Glasser

Manon Bart

Physics PhD student

Sita Dawanse

Physics PhD student

Brad Mitchell

Physics PhD student

Optical Communication

• Benefits such as high bandwidth, narrow beam divergence, low power consumption over RF links
• Many degrees of freedom to encode information
• Amplitude/Phase
• Polarization
• Spatial
• Spatial degrees of freedom are particularly promising due to high dimensionality, high bit transfer rate

Introduction to higher order spatial modes

HG modes

IG modes

LG modes

• Communication protocols can encode information in order numbers
• Permits high capacity information transfer
• Compensation of distorted image at reciever necessary

Motivation

• Use machine learning to classify the mode and order numbers of Laguerre Gauss (LG), Ince Gauss (IG) and Hermite Gauss (HG) modes
• Analyze performance of spatial modes as information carriers

​

1. Convolutional Autoencoder to reduce effects of turbulence
2. Convolutional Neural Network to classify mode orders

Experimental Generation of Spatial Modes

• Experimentally generated images contain various turbulence levels and exposure times

​

​

​

​

​

• ~20,000 images generated for training, validation and testing

Background: Machine Learning Models

Background: Convolutional Networks

• Convolutional neural networks are useful in multidimensional data for analyzing and classifying visual data
• Main features include
• Convolutional layers
• Works to extract features
• Pooling layers
• Works to reduce dimensionality

​

https://www.ibm.com/think/topics/convolutional-neural-networks

Background: CNNs

• Convolutional neural networks are classification models
• Output consists of class predictions
• Relies on supervised learning to improve classification accuracy during training

​

Convolutional Neural Network

https://medium.com/@abhishekjainindore24/understanding-convolutional-neural-networks-cnns-with-an-example-on-the-mnist-dataset-a64815843685

Background: Denoising CAEs

• Denoising Convolutional Autoencoders are a generative machine learning model
• Consist of encoder, latent space, decoder
• Autoencoders proven useful in improving classification accuracy when image corruption present

​

Denoising Convolutional Autoencoder

Denoising autoencoders with Keras, TensorFlow, and Deep Learning, Adrian Rosebrock 

Latent Space

Model Architecture and Training

CAE Architecture and Training

• Training/validation performed on half of experimental images
• Optimized architecture parameters
• Latent space vector size 30
• ReLU activation function at each layer
• Grid search for optimal hyperparameters

​

​

CAE Architecture and Training

• Training loss metric was defined as mean squared error (MSE)
• Training run for 100 epochs over 5 trials to assure convergence
• Order of magnitude reduction in MSE

​

CAE Generated Images

• Output of CAE is a denoised representation of the original experimental images
• Output of CAE is used as input to the CNN classification model

​

Increasing Turbulence Strength

Sample Input and Output of the Denoising CAE

CNN Architecture and Training

• Training/validation on half of experimental images
• Output predicts mode and order number
• Optimized architecture parameters:
• ReLU activation function
• Dropout layers for overfitting

​

CNN Architecture and Training

• Training loss metric was defined as sparse categorical cross entropy
• Training run for 15 epochs, results presented over 10 trials

​

Model Testing and Results

Results: Classification Accuracy

• Final model of CAE and CNN achieves 99.2% accuracy across all modes
• Addition of the CAE improves classification accuracy, reduces variance
• HG modes are most robust to effects of turbulence

Test Set Accuracy of Individual Spatial Modes for varying Turbulence Strength

Results: Confusion Matrix

• Most predictions follow true labels
• Common errors between
• IG3,1and LG1,1
• IG4,4 and LG0,4
• Errors can be physically attributed to IG-LG transition

CNN only

CAE + CNN

Ongoing Work

• Generating simulated images for classification in the case of long distance propagation
• Further analysis on classification errors at transitions between IG-LG and IG-HG

Lohani, S., Knutson, E.M. & Glasser, R.T. Generative machine learning for robust free-space communication. Commun Phys 3, 177 (2020).

Conclusion

• Convolutional networks support information transmission in single photon limit, >99% accuracy across trials
• Addition of CAE is useful in case of high-loss environments such as propagation through the atmosphere
• Model can be pretrained and implemented in real time at a receiver
• Does not significantly affect size, weight and power requirements for free space optical communication schemes
• Machine learning models can be integrated with existing error mitigation techniques such as adaptive optics

Acknowledgements

• Many thanks to our many collaborators and colleagues
• RPI
• SMU
• NASA

Questions?

Thank you for your time!

Bart, Manon P., et al. "Classification of Single Photons in Higher-Order Spatial Modes via Convolutional Neural Networks." arXiv preprint arXiv:2412.07560 (2024).

​