FAT: Frequency-Aware Transformation for Bridging Full-Precision and Low-Precision Deep Representations.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Quantizing convolutional neural networks (CNNs) to low bitwidths is challenging due to significant performance degradation.
• Existing methods often require complex hyperparameter tuning (e.g., nonuniform stepsize, layerwise bitwidths) due to discrepancies between full- and low-precision representations.

Purpose of the Study:

• To present a novel quantization pipeline, Frequency-Aware Transformation (FAT), that simplifies the process of learning low-bitwidth CNNs.
• To demonstrate that FAT enables effective quantization with simple quantizers and minimal hyperparameter tuning.
• To provide a new frequency-based perspective for model quantization.

Main Methods:

• FAT transforms network weights in the frequency domain to remove redundant information prior to quantization.
• The transformed weights become amenable to training in low bitwidth using simple, standard quantizers.
• Theoretical analyses confirm that FAT minimizes quantization errors for both uniform and nonuniform quantization schemes.

Main Results:

• FAT facilitates the embedding of CNNs in low bitwidths without complex hyperparameter adjustments.
• When combined with simple uniform or logarithmic quantizers, FAT achieves state-of-the-art performance across various bitwidths and model architectures.
• The method is easily integrated into diverse CNN architectures.

Conclusions:

• FAT offers a novel and effective frequency-based approach to model quantization.
• The pipeline simplifies the training of low-bitwidth CNNs, overcoming performance drops associated with quantization.
• FAT achieves competitive or superior performance compared to prior methods, highlighting its practical utility.