Event management for large scale event-driven digital hardware spiking neural networks

Area of Science:

• Neuromorphic Engineering
• Computer Science
• Computational Neuroscience

Background:

• The growing interest in brain-like computation has spurred the development of numerous neuromorphic systems.
• Spiking Neural Networks (SNNs), event-driven simulations, and digital hardware architectures are key areas of focus.
• A significant gap exists between the software popularity of event-driven SNNs and their digital hardware implementations due to poor scalability of existing event management solutions.

Purpose of the Study:

• To introduce a novel pipelined digital hardware data structure, the structured heap queue, for efficient event management.
• To demonstrate the suitability of the structured heap queue for large-scale digital hardware event-driven SNNs.
• To address the scalability limitations of current hardware solutions for event management in neuromorphic systems.

Main Methods:

• Introduction of the structured heap queue, a pipelined digital hardware data structure.
• Demonstration of the structured heap queue's performance on a field-programmable gate array (FPGA).
• Implementation of a large-scale SNN comprising 65,536 neurons and 513,184 synapses for an image segmentation task.

Main Results:

• The structured heap queue exhibits linear scaling for memory, logarithmic scaling for logic resources, and constant scaling for processing time with an increasing number of events.
• Events are processed at a rate of 1 every 7 clock cycles.
• A 406×158 pixel image was successfully segmented in 200 ms using the implemented SNN.

Conclusions:

• The structured heap queue provides a scalable solution for event management in digital hardware event-driven SNNs.
• This data structure facilitates the efficient implementation of large-scale neuromorphic systems.
• The demonstrated FPGA implementation highlights the practical viability and performance benefits of the structured heap queue for real-world applications like image segmentation.