Whether the Support Region of Three-Bit Uniform Quantizer Has a Strong Impact on Post-Training Quantization for MNIST...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Neural network (NN) weight compression is crucial for deploying models on resource-constrained edge devices.
• Post-training quantization (PTQ) offers a method for NN compression without retraining.
• Uniform quantization, especially at low bit-widths like three bits, presents an opportunity for significant model compression.

Purpose of the Study:

• To investigate the performance of three-bit post-training uniform quantization.
• To analyze the impact of the support region threshold parameter on quantization accuracy for MLP and CNN models on the MNIST dataset.
• To determine the criticality of the support region threshold in three-bit uniform quantization compared to lower bit-widths.

Main Methods:

• Studied three-bit post-training uniform quantization.
• Evaluated the impact of various support region threshold values on NN accuracy.
• Compared the performance of three-bit quantization with two-bit quantization on MLP and CNN models for MNIST classification.

Main Results:

• The choice of support region threshold has a limited impact on the accuracy of three-bit uniform quantization for both MLP and CNN models.
• Accuracy preservation is largely maintained with three-bit uniform quantization, irrespective of the threshold parameter.
• This insensitivity contrasts with two-bit uniform post-training quantization, where threshold selection is more critical.

Conclusions:

• Three-bit uniform post-training quantization is a resilient technique for NN compression.
• The minimal impact of the support region threshold simplifies the quantization process and enhances its applicability.
• This quantization method holds significant potential for practical deployment on edge devices due to its robustness and simplicity.