Event-Driven Random Back-Propagation: Enabling Neuromorphic Deep Learning Machines.

Area of Science:

• Neuromorphic Computing
• Computational Neuroscience
• Artificial Intelligence

Background:

• Developing computationally efficient brain-compatible models for neuromorphic computing is a key challenge.
• Deep neural network algorithms inspire current neuromorphic approaches.
• Traditional Backpropagation (BP) requires high-precision memory and network-wide information, posing hardware implementation difficulties.

Purpose of the Study:

• To introduce a novel, efficient learning rule for neuromorphic systems.
• To demonstrate that exact backpropagated gradients are not essential for learning deep representations.
• To develop a rule suitable for implementation in neuromorphic hardware.

Main Methods:

• Developed an event-driven random backpropagation (eRBP) rule utilizing error-modulated synaptic plasticity.
• Employed a two-compartment Leaky Integrate & Fire (I&F) neuron model.
• The eRBP rule requires minimal operations per synaptic weight (one addition, two comparisons).

Main Results:

• The eRBP rule enables rapid learning of deep representations.
• Achieved classification accuracies comparable to GPU-based artificial neural network simulations on permutation-invariant datasets.
• Demonstrated robustness to quantization of neural and synaptic states during learning.

Conclusions:

• The event-driven random backpropagation (eRBP) rule is a computationally efficient and hardware-friendly alternative to traditional BP for neuromorphic learning.
• This method facilitates the development of advanced AI capabilities within the constraints of neuromorphic hardware.
• eRBP offers a promising pathway for realizing effective deep learning in brain-inspired computing systems.