The backpropagation algorithm implemented on spiking neuromorphic hardware.

Area of Science:

• Neuroscience and Artificial Intelligence
• Neuromorphic Engineering
• Machine Learning

Background:

• Natural neural systems inspire advanced machine learning and neuromorphic circuits.
• Modern deep learning, particularly backpropagation, faces challenges in neurophysiological plausibility and hardware implementation.
• Existing neuromorphic approaches often struggle to replicate the exact backpropagation algorithm.

Purpose of the Study:

• To present a neuromorphic, spiking backpropagation algorithm implemented on Intel's Loihi research processor.
• To demonstrate a proof-of-principle three-layer circuit capable of on-chip learning.
• To showcase the feasibility of exact backpropagation in a fully on-chip Spiking Neural Network (SNN).

Main Methods:

• Implementation of a synfire-gated dynamical information coordination and processing algorithm.
• Deployment on Intel's Loihi neuromorphic research processor.
• Training and testing on MNIST and Fashion MNIST datasets for digit and clothing item classification.

Main Results:

• Successful proof-of-principle demonstration of a three-layer spiking neural network learning task.
• Achieved classification accuracy competitive with off-chip trained SNNs.
• Demonstrated an energy-delay product suitable for edge computing applications.

Conclusions:

• This work represents the first fully on-chip, computer-in-the-loop-free implementation of the exact backpropagation algorithm in an SNN.
• The developed approach enables low-power, low-latency deep learning applications on neuromorphic processors.
• Highlights a viable path for integrating advanced machine learning with in-memory, massively parallel neuromorphic hardware.