General Bitwidth Assignment for Efficient Deep Convolutional Neural Network Quantization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (DCNNs) require significant computational resources, hindering deployment on devices with limited capacity.
• Model quantization, specifically reducing the bitwidth of weights and activations, is a key technique for efficient DCNN deployment.

Purpose of the Study:

• To develop a general algorithm for optimal layerwise bitwidth assignment in DCNNs.
• To minimize the trade-off between classification accuracy and model size through efficient quantization.

Main Methods:

• A prediction model using a geometrical approach to estimate classification accuracy loss from weight quantization.
• Dynamic programming to determine optimal bitwidth allocation for weights based on predicted accuracy loss.
• Optimization of activation bitwidth assignment considering the signal-to-quantization-noise ratio (SQNR) relative to weight quantization.

Main Results:

• The proposed algorithm effectively balances classification accuracy and model size across various DCNN architectures.
• Experimental validation confirms the efficacy of the bitwidth assignment algorithm and the accuracy loss prediction model.
• The algorithm demonstrates successful extension to object detection tasks.

Conclusions:

• The developed bitwidth assignment algorithm provides an efficient solution for quantizing DCNNs.
• The method offers a generalizable approach for optimizing deep learning models for resource-constrained environments.
• The technique shows promise for improving the efficiency of DCNNs in applications like object detection.