Implementation of Model Free Training for End-to-End Communication Systems

The authors in the paper “Model-Free Training of End-to-End Communication Systems” consider auto-encoder based architecture for the entire transmitter and receiver model with the channel between the encoder (transmitter) and decoder (receiver). The following is the brief explaination of the paper.

• Here, model-free refers to the communication model where the channel is unknown or has non-differentiable components. Hence the encoder (transmitter) cann’t be trained through backpropagation.

• The objective was to train the autoencoder such that the decoder is able to detect the transmitted messages that are transmitted through the channel.

• To train the autoencoder, the authors propose an alternating algorithm where the decoder (receiver) and the encoder (transmitter) are trained separately.

• Since the encoder cann’t be trained as the channel can be non-differentiable, approximate gradient of loss function is used to train the encoder (transmitter).

Neural Network Architecture:

• Input: One-hot encoded messages that are transmitted through AWGN/ Rayleigh Block Fading (RBF) channel after encoding.

• Encoder: 2 layers of fully connected network with elu as the activation function with normalization to satisfy the power constraints. The output of the encoder is the symbol that is transmitted.

• Decoder: 2 layers of fully connected network with elu as the activation function for the $1^{st}$ layer and softmax as the activation function for the $2^{nd}$ layer. Input to the decoder is the noisy version of the transmitted symbol.

• Output: Predict the transmitted symbol using the received symbol.

• Loss Function: Categorical cross-entropy

• Optimizer: Adam

• For each training iteration, the receiver and the transmitter are trained alternatively.

My replication of certain results of the mentioned research paper can be found here.