Design of a 2-Bit Neural Network Quantizer for Laplacian Source

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Real-time inference is crucial for neural network applications on resource-constrained mobile devices.
• Model compression techniques like quantization are vital to address storage and computational limitations.

Purpose of the Study:

• To design and analyze a 2-bit uniform quantization model for Laplacian sources.
• To improve the efficiency and speed of neural network inference for mobile deployment.

Main Methods:

• Developed a 2-bit uniform quantization strategy tailored for Laplacian source data.
• Implemented and evaluated the quantized model using Multi-Layer Perceptrons (MLP) and Convolutional Neural Networks (CNN).

Main Results:

• Achieved high classification accuracy: over 96% for MLP and over 98% for CNN.
• Demonstrated competitive performance compared to other quantization solutions with near-optimal precision.
• The 2-bit uniform quantization offers implementation simplicity, leading to reduced processing time.

Conclusions:

• The proposed 2-bit uniform quantization model effectively balances accuracy and efficiency for neural networks on mobile devices.
• This approach facilitates faster inference and simpler implementation without significant performance degradation.