Layer-specific approximate multipliers for energy-precision trade-offs in convolutional neural networks

Area of Science:

• Computer Engineering
• Artificial Intelligence
• Hardware Acceleration

Background:

• Approximate computing offers hardware efficiency for error-resilient applications like neural networks.
• Traditional computing prioritizes precision, often at the cost of efficiency.
• CNNs present unique characteristics influencing the suitability of approximation techniques.

Purpose of the Study:

• To develop and evaluate a novel CNN-specific approximation methodology.
• To enhance hardware efficiency and energy savings in CNNs.
• To balance computational complexity and accuracy in deep learning hardware.

Main Methods:

• Identified characteristics of approximate multipliers suitable for CNNs based on weight distribution.
• Designed novel approximate multipliers (AM_5×5, AM_4×4, AM_3×3) using operand truncation with adjustable accuracy.
• Proposed two training algorithms: one for direct optimization and another using a gradual strategy.
• Implemented and evaluated the methodology on ASIC in 28 nm CMOS technology using VGG16, VGG10, and AlexNet.

Main Results:

• Achieved significant energy efficiency gains per operation: up to 86% (VGG10), 95% (VGG16), and 88% (AlexNet) with the first strategy.
• Second strategy yielded energy efficiency improvements of 81% (VGG10), 92% (VGG16), and 84% (AlexNet).
• Demonstrated scalability of approximate multipliers and adjustable accuracy through operand bit truncation.

Conclusions:

• The proposed methodology effectively leverages CNN features to enhance hardware efficiency.
• Experimental results validate the potential for optimizing both energy and hardware resources in CNN designs.
• This approach advances practical applications of CNNs by improving their energy and computational efficiency.