Efficient Deep Learning Model Compression for Sensor-Based Vision Systems via Outlier-Aware Quantization

Area of Science:

• Computer Vision
• Deep Learning
• Sensor Technology

Background:

• Deep neural networks (DNNs) require efficient models for real-time image feature extraction, especially in resource-limited settings.
• Existing quantized DNNs struggle with outliers, causing performance degradation in low-precision scenarios.
• Outliers increase dynamic range, reducing quantization resolution crucial for sensor-based image analysis.

Purpose of the Study:

• To propose an outlier-aware quantization (OAQ) method to improve quantization accuracy in deep learning models.
• To address the sensitivity of quantized DNNs to outliers in weight distributions.
• To enhance the performance of deep learning models for sensor-based vision applications.

Main Methods:

• Developed an outlier-aware quantization (OAQ) method to reshape weight distributions.
• Analyzed outlier-handling techniques using structural similarity (SSIM).
• Validated OAQ's compatibility with existing quantization schemes (post-training quantization and quantization-aware training).

Main Results:

• OAQ significantly reduced the negative impact of outliers while maintaining computational efficiency.
• OAQ demonstrated orthogonality to existing quantization methods, offering compatibility without extra overhead.
• 4-bit OAQ ResNet20 in PTQ showed improved accuracy over full-precision models.
• OAQ enhanced 2-bit quantization performance by 43.55% in QAT compared to baselines.

Conclusions:

• OAQ effectively mitigates quantization errors in deep learning models.
• The proposed method optimizes deep learning for sensor-based vision applications.
• OAQ shows significant potential for improving low-precision deep learning models.