HybMED: A Hybrid Neural Network Training Processor With Multi-Sparsity Exploitation for Internet of Medical Things

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Hardware Acceleration

Background:

• AIoMT applications face accuracy issues due to patient variability in cloud/edge modes.
• On-chip learning offers user-specific adaptation but current processors lack versatility and efficiency.
• Existing solutions struggle with resource utilization and energy consumption for AIoMT.

Purpose of the Study:

• To develop a novel neural signal processor, HybMED, for versatile on-chip AIoMT training.
• To enhance resource utilization, area efficiency, and energy efficiency in AIoMT devices.
• To address accuracy degradation in AIoMT applications caused by physiological signal variations.

Main Methods:

• Proposed HybMED, a neural signal processor supporting on-chip hybrid neural network training.
• Utilized a composite direct feedback alignment-based paradigm for neural network adaptation.
• Implemented a reconfigurable homogeneous core with heterogeneous data flow and exploited sparsity for efficiency.

Main Results:

• Achieved an average accuracy improvement of 41.16% in physiological signal processing tasks via online few-shot learning.
• Demonstrated area efficiency of 1.17 GOPS/mm² and energy efficiency of 1.58 TOPS/W.
• Showcased a 65x improvement in area efficiency and 1.48x in energy efficiency compared to prior art.

Conclusions:

• HybMED is a suitable processor for general-purpose health monitoring AIoMT devices.
• The chip design significantly improves resource and energy efficiency for on-chip learning.
• HybMED offers a substantial advancement in AIoMT hardware for personalized healthcare.