Ternary Quantization
An extreme neural network model compression technique that restricts model weights to three discrete values: -1, 0, and +1. Each weight theoretically requires approximately 1.58 bits of storage, achieving compression ratios exceeding 10× compared to FP16. Maintaining accuracy is challenging, and compensation techniques such as quantization-aware training are typically required.
Timeline (last 90 days)
Ternary quantization restricts neural network weights to three discrete values (typically -1, 0, and +1), requiring theoretically only about 1.58 bits of storage per weight.
Extreme quantization trades size advantages potentially at the cost of longer inference times.
In ternary quantization, matrix multiplication can degenerate into addition and subtraction, theoretically allowing multipliers to be bypassed entirely.
All Facts (3)
Ternary quantization restricts neural network weights to three discrete values (typically -1, 0, and +1), requiring theoretically only about 1.58 bits of storage per weight.
50%UnverifiedExtreme quantization trades size advantages potentially at the cost of longer inference times.
50%UnverifiedIn ternary quantization, matrix multiplication can degenerate into addition and subtraction, theoretically allowing multipliers to be bypassed entirely.
50%