EPOC: A 28-nm 5.3 pJ/SOP Event-Driven Parallel Neuromorphic Hardware With Neuromodulation-Based Online Learning

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Biologically Inspired Computing

Background:

• Neuromorphic hardware promises energy-efficient, adaptable AI but lacks unified learning frameworks.
• Existing systems underutilize spike-based parallelism, limiting computational efficiency and scale.

Purpose of the Study:

• To develop a unified, event-driven, and massively parallel multi-core neuromorphic online learning processor (EPOC).
• To introduce a neuromodulation-based framework for diverse SNN learning algorithms.
• To enhance computational efficiency and learning accuracy in neuromorphic systems.

Main Methods:

• Developed EPOC, a multi-core neuromorphic processor with a novel event-driven computation method.
• Implemented a neuromodulation-based online learning framework supporting supervised SNN learning.
• Utilized a low-memory streaming single-sample learning strategy.
• Leveraged parallel multi-channel and multi-core architecture for enhanced processing.

Main Results:

• EPOC achieved state-of-the-art accuracy on MNIST (99.2%), NMNIST (98.2%), and DVS-Gesture (94.3%).
• Demonstrated a 9.9x time efficiency improvement over baseline architectures.
• Local learning on EPOC showed a 2.9x time efficiency gain over global learning.
• Achieved high energy efficiency (5.3 pJ/SOP) and throughput (328 GOPS/51 GSOPS).

Conclusions:

• EPOC offers a unified, efficient, and bio-plausible learning framework for neuromorphic hardware.
• The event-driven design and parallel architecture significantly boost computational efficiency.
• EPOC advances the development of human-like intelligence in energy-efficient hardware.