Convolutional networks for fast, energy-efficient neuromorphic computing.

Area of Science:

• Artificial Intelligence
• Computer Engineering
• Neuroscience

Background:

• Deep neural networks achieve human-level recognition performance.
• Neuromorphic computing offers energy-efficient hardware using spiking neurons and low-precision synapses.

Purpose of the Study:

• To demonstrate the implementation of deep convolution networks on neuromorphic hardware.
• To evaluate the performance, energy efficiency, and trainability of these networks.

Main Methods:

• Implementing deep convolution networks on a novel neuromorphic chip architecture.
• Training networks using backpropagation.
• Evaluating classification accuracy on eight standard vision and speech datasets.
• Measuring inference speed and power consumption.

Main Results:

• Deep convolution networks on neuromorphic hardware approach state-of-the-art accuracy.
• Inference achieved high throughput (1,200-2,600 frames/s) with low power consumption (25-275 mW).
• Networks were specified and trained using backpropagation with ease.

Conclusions:

• Neuromorphic computing can effectively implement deep learning algorithms.
• This integration merges deep learning's algorithmic power with neuromorphic efficiency.
• Enables progress towards embedded, intelligent, brain-inspired computing.