Brain-inspired global-local learning incorporated with neuromorphic computing.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Neuromorphic Engineering

Background:

• Traditional AI learning methods include error-driven global learning and neuroscience-inspired local learning.
• Integrating these approaches offers complementary learning capabilities for diverse AI scenarios.
• Neuromorphic computing requires advanced algorithms and co-design for optimal performance.

Purpose of the Study:

• To present a neuromorphic global-local synergic learning model.
• To incorporate a brain-inspired meta-learning paradigm and a differentiable spiking model.
• To enable multiscale learning through meta-learned local plasticity and top-down supervision.

Main Methods:

• Developed a differentiable spiking neural network model with neuronal dynamics and synaptic plasticity.
• Integrated meta-learning for adaptive local plasticity control.
• Implemented the model on the Tianjic neuromorphic platform with algorithm-hardware co-designs.

Main Results:

• Demonstrated superior performance in few-shot learning, continual learning, and fault-tolerance learning tasks.
• Achieved significantly higher performance compared to single-learning methods.
• Showcased the model's ability to leverage neuromorphic many-core architectures for hybrid computation.

Conclusions:

• The proposed neuromorphic global-local synergic learning model effectively integrates diverse learning paradigms.
• The model shows significant advantages in various learning tasks and demonstrates efficient hardware implementation.
• This work advances neuromorphic computing by providing a versatile and high-performing learning algorithm.