Comparing Neuromorphic Solutions in Action: Implementing a Bio-Inspired Solution to a Benchmark Classification Task...

Area of Science:

• Neuromorphic computing
• Computational neuroscience
• Artificial intelligence hardware

Background:

• Neuromorphic computing utilizes neuronal circuit models for problem-solving.
• Availability of neuromorphic hardware and development of neuromorphic algorithms are increasing.
• Assessing and comparing neuromorphic systems is crucial for practical applications.

Purpose of the Study:

• To practically implement and compare a bio-inspired spiking network for multivariate classification on three distinct neuromorphic platforms.
• To evaluate performance, ease of implementation, speed, scalability, and power efficiency across different hardware.
• To identify bottlenecks in neuromorphic system deployment.

Main Methods:

• Implementation of a spiking neural network for handwritten digit classification.
• Deployment on three platforms: Spikey (hybrid digital/analog), SpiNNaker (digital spike-based), and GeNN (GPU meta-compiler).
• Assessment of classification performance, execution speed, and power consumption.

Main Results:

• Comparable classification performance across all three platforms, indicating model capability over platform limitations.
• Significant portions of computation time and power consumption occurred on the host machine, not the neuromorphic device.
• Host-device communication, data preparation, and result decoding were major contributors to overhead.

Conclusions:

• Specialized neuromorphic hardware benefits can be negated by inefficient host-device communication and non-neuronal computations.
• Optimization of host-device communication architecture is critical for scalability, throughput, and latency.
• Minimizing host-device interaction is essential for efficient neuromorphic computing system design.