Machine learning for optical communication over a dispersive fiber

The goal of this project is to use machine learning techniques to decode a distorted signal received through a simulated optical fiber.
We examine two encodings: binary (NRZ) and 4-bit (4-PAM)

Maximum Likelihood Estimator (Matlab)
We assume each time sample within a clock cycle has a Gaussian distribution.
An input sample is scaled by the variance to find the deviation from the mean of each symbol, and the prediction is given by the nearest symbol.

Convolutional Neural Network (Python)
A CNN typically used for 2D image classification is modified to classify 1D waveforms.
Instead of 2D convolutions, we use 1D dilated causal convolutions.

Support Vector Machine (Matlab)
The time-domain signal is partitioned into sets of samples based on the bit rate.
The SVM is trained on the samples after they have been cast as 16-dimensional feature vectors, where each dimension is a sample within the clock cycle.

Dependencies

Matlab: None
Python: Python 3.6, Numpy 1.14, TensorFlow 1.6

Feature Extraction files
auto_labeler.m - run program to generate label file from TX file
data_parser.m - function to parse RX and label files into training and test arrays

Maximum Likelihood Estimator files
MLbin.m - runs max likelihood on binary data
MLpam.m - runs max likelihood on 4-PAM data

Convolutional Neural Net files
dilated_neural_net.py - defines helper functions for training the CNN
test_nn_binary.py - run to train CNN to classify binary data
test_nn_pam4.py - run to train CNN to classify 4-PAM data
To run visualizations, use the following command (alternatively python -m tensorboard.main)

tensorboard --logdir=/tmp/dilated_cnn_model_binary

or

tensorboard --logdir=/tmp/dilated_cnn_model_pam4

Once TensorBoard is running, navigate your web browser to localhost:6006 to view the TensorBoard.

Support Vector Machine files
SVM_train.m - function to train an SVM hyperplane using training array and hyperparameters
SVM_test.m - function to test an SVM hyperplane using test array and hyperparameters
binary_SVM.m - run SVM for Binary RX data
svmp4.m - run SVM for 4-PAM RX data

svmp4_helper.m - Calls svmp4.m N times and returns averages
binary_SVM_helper.m - Calls binary_SVM.m N times and returns averages
hard_margin_main.m - Collects and plots data for binary and 4-PAM SVM
test_regularizer.m - run program to test soft-margin SVM
test_desynch.m - run program to test the effects of desynchronization

Gaussian Radial Basis Kernel files
kernelbin_SVM.m -this function contains the RBF kernel code for binary data training and testing
kernelpam_SVM.m -this function contains the RBF kernel code for 4PAM data training and testing
kernelfindparam.m -m file that deploys the binary kernel. Oneshot variable determines if a tuning sweep is to be performed using variable vectors, otherwise the function will be run once with our default tuned values.
kernelfindparampam.m -m file that deploys the 4PAM kernel.