Supernetwork-based efficient mapping of deep learning applications to mixed-precision hardware using model adaptation

Area of Science:

• Computer Science
• Artificial Intelligence
• Hardware Acceleration

Background:

• Artificial Intelligence (AI) applications are rapidly expanding, requiring efficient and scalable solutions for real-world deployment.
• Heterogeneous accelerators, which combine analog and digital components, offer potential for localized and energy-efficient neural network computations.
• Optimizing performance on these accelerators involves a critical balance between energy efficiency and model accuracy through effective neural network layer mapping.

Purpose of the Study:

• To introduce Mixed-Precision Supernetwork, a unified framework for training mixed-precision supernetworks.
• To enable seamless integration of quantized digital layers with analog noise-sensitive layers.
• To develop a mapping-aware adaptation strategy for dynamic layer assignment and hardware-aware architecture search.

Main Methods:

• Developed a unified framework for training mixed-precision supernetworks.
• Incorporated a mapping-aware adaptation strategy for dynamic layer optimization.
• Utilized hardware-aware architecture search to refine neural network designs.

Main Results:

• Mixed-Precision Supernetwork achieves mappings approximately 2.2x faster than fully analog approaches.
• Demonstrates an average increase of 3.4% in model accuracy compared to fully analog methods.
• Improves energy efficiency by mapping up to 80% of model weights to analog hardware while preserving full-precision accuracy.

Conclusions:

• Mixed-Precision Supernetwork provides a groundbreaking approach for efficient deep learning model deployment on heterogeneous accelerators.
• The framework successfully balances energy efficiency and model accuracy through optimized layer mapping.
• This method enables significant performance gains and enhanced energy efficiency for AI on specialized hardware.