The effects of quantization on multilayer neural networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Multilayer neural networks (NNs) are computationally intensive.
• Weight quantization is a technique to reduce NN model size and inference time.
• Quantization can lead to performance degradation in NNs.

Purpose of the Study:

• To analyze the impact of weight quantization on multilayer neural network performance.
• To develop a predictive method for performance degradation due to quantization.
• To introduce an algorithm for mitigating quantization-induced noise.

Main Methods:

• Derivation of a method to predict performance degradation based on network properties and quantization bits.
• Evaluation of the predictive method against simulation results.
• Development and application of an algorithm to decrease output noise.

Main Results:

• The proposed method accurately predicts performance degradation caused by weight quantization.
• Simulations validate the predictive accuracy of the derived method.
• The noise reduction algorithm demonstrates effectiveness in improving output quality.

Conclusions:

• Weight quantization in NNs necessitates methods for predicting and mitigating performance loss.
• The developed predictive model and noise reduction algorithm offer practical solutions for quantized NNs.
• This research contributes to the efficient deployment of NNs through effective quantization strategies.