Chapter 180: Ternary Gradient Compression for Trading

Overview

Sending thousands of model updates across a global trading network is a bandwidth nightmare. In the previous chapter, we explored basic quantization. Now, we implement Ternary Gradient Compression (TGC).

TGC reduces each weight update to just one of three values: ${-1, 0, +1}$. This not only compresses the data but also acts as a powerful regularizer, filtering out market noise.

Adaptive Thresholding

Markets are not static. During a high-volatility event, small gradient updates might be crucial. During a sideways market, most updates are just noise. Our implementation uses an Adaptive Threshold:

High Volatility: The threshold drops to capture more details.
Low Volatility: The threshold rises to maximize sparsity and save bandwidth.

Project Structure

180_fl_gradient_compression/
├── README.md           # English Overview
├── README.ru.md        # Russian Overview
├── docs/ru/theory.md   # Mathematical deep-dive
├── python/
│   ├── model.py            # Base Neural Network
│   ├── ternary_core.py     # TGC & Adaptive Logic
│   └── train.py            # Adaptive vs. Fixed compression simulation
└── rust/src/
    └── lib.rs              # Optimized 2-bit packing engine