Abhijeet Mulgund's Personal Webpage

Search

Search IconIcon to open search

TurboAE

Last updated Nov 7, 2022

TurboAE is a Deep Learning Machine Learning Model implemented and published by Yihan Jiang. The model attempts to solve the Channel Coding problem by searching for a Neural Network Coding Encoder and Coding Decoder that mimic the architecture of a Turbo Code.

# Architecture Details

# Encoder

The TurboAE encoder is made up of 3 encoders with the following architecture

Input: Batch x Block Length x 1; Block Length is usually 100.

  1. Window 5, 100 Filter Conv w/ ELU w/ Zero Padding-> Batch x Block Length x 100
  2. Window 5, 100 Filter Conv w/ ELU w/ Zero Padding -> Batch x Block Length x 100
  3. TimeDistributed 1 Unit FC -> Batch x Block Length x 1
  4. Power Constraint
    1. We subtract normalize uniformly all outputs with
      1. Mean: -0.0215
      2. Standard Deviation: 0.5114
    2. For TurboAE-Binary, we quantize to $\pm 1$ using the Sign Function.

Parameters per encoder: 50,801 Total Parameters: 152,403

Counts were measured in cell 7 of this notebook. These counts match manual calculations: $(51100 + 100) + (5100100 + 100) + (100 + 1) = 50,801$

# Decoder

TODO

Total Parameters: 2,453,656

Counts were measured in cell 6 of this notebook.