Unsupervised Network Quantization via Fixed-Point Factorization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) offer high performance but demand substantial memory and computational resources.
• Network quantization, particularly fixed-point quantization, aims to accelerate and compress DNNs.
• Existing methods face performance degradation at extremely low bitwidths and rely heavily on labeled data for retraining.

Purpose of the Study:

• To propose an efficient framework, Fixed-Point Factorized Network (FFN), for extreme low-bit quantization of DNNs.
• To enable significant model compression and computational efficiency without substantial accuracy loss.
• To reduce reliance on large labeled datasets for retraining quantized networks.

Main Methods:

• Developed the Fixed-Point Factorized Network (FFN) framework to convert all network weights to ternary values (-1, 0, 1).
• Quantized network activations to an 8-bit format.
• Leveraged low-bit fixed-point additions, replacing computationally expensive 32-bit floating-point multiply-accumulate operations.

Main Results:

• Achieved negligible performance degradation with ternary weight quantization, even without supervised retraining.
• Demonstrated over 20x compression on large-scale datasets like ImageNet and MS COCO.
• Significantly reduced multiply operations while maintaining comparable accuracy.

Conclusions:

• The FFN framework offers an efficient solution for compressing and accelerating deep neural networks.
• FFN enables significant memory and computational savings with minimal accuracy compromise.
• This approach is particularly valuable for real-world applications where labeled data is scarce.