TurboAE
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.
- Window 5, 100 Filter Conv w/ ELU w/ Zero Padding-> Batch x Block Length x 100
- Window 5, 100 Filter Conv w/ ELU w/ Zero Padding -> Batch x Block Length x 100
- TimeDistributed 1 Unit FC -> Batch x Block Length x 1
- Power Constraint
- We subtract normalize uniformly all outputs with
- Mean: -0.0215
- Standard Deviation: 0.5114
- For TurboAE-Binary, we quantize to $\pm 1$ using the Sign Function.
- We subtract normalize uniformly all outputs with
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
Total Parameters: 2,453,656
Counts were measured in cell 6 of this notebook.