Neural network converter converts existing CNNs to SNNs

June 17, 2019 //By Rich Pell
Neural network converter converts existing CNNs to SNNs
BrainChip Holdings (San Francisco, CA) has announced the availability of a neural network converter that enables users to easily convert existing convolutional neural networks (CNNs) to an event-based spiking neural network (SNN) compatible with the company's Akida neuromorphic computing architecture.

The converter is integrated with the company's Akida Development Environment (ADE) to provide network conversion and simulation. This unified flow, says the company, represents the world's first commercially available development environment enabling both CNN and SNN implementations on the same hardware device while maintaining the inherent performance and power benefits of event-based neural networks.

"The low power inherent in the Akida device will set a new standard in neural network design, implementation and performance," says Roger Levinson, BrainChip COO. "With no barriers to utilizing the Akida platform, Edge solution developers will have their cake and eat it too. They can leverage existing CNN solutions as well as incorporate next generation SNN solutions all in a single development environment and on a single device and achieve a low power solution without sacrificing performance."

The CNN to SNN conversion flow is designed for ease of use and uses standard text files. Users can implement many CNN architectures targeted at edge applications such as image processing, anomaly detection, ADAS, robotics, and key word spotting.

The conversion, says the company, maintains near full accuracy with increased performance while eliminating neural network computational overhead. Post-conversion, the entire network is executed within the neural fabric of the Akida chip, which means that the host computational requirements of the neural network are eliminated. The host delivers the data to the chip and retrieves the results.

The conversion flow takes an input as a standard CNN description, which the user modifies through a logical process to Akida compatible layers. The modified network description is then processed through standard quantization and training.

To optimize performance, the ADE supports programmable multiple bit-widths including binary, ternary, and 4-bit for both weights and activations in each network layer. Once the final network configuration is achieved, the resulting Akida compatible network description is output in industry standard .yml and .dat files. These files are run in the Akida emulation environment to generate performance


Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.