Monday, December, 23, 2024 06:22:07
Trending News:
- Hyundai to invest $85.41Bn amid plans to sell 2M EV units by 2030
- Lenovo to invest $1B to drive AI deployment for businesses worldwide
- UAE's Masdar consortium inks deal for $10Bn mega wind project in Egypt
- Xiaomi and Dixon join forces for manufacturing smart phones in India
- Wipro extends Google Cloud partnership to advance Generative AI adoption
- Apple and Google team up to foil risk of unwanted tracking devices
- First Republic reports plunge in deposits, 50% fall in bank stock
- Epic partners with Microsoft for generative AI deployment for improved EHRs
- Australia joins list of nations banning TikTok on govt devices
Date: 2019-06-12
Technology
A standard CNN description is taken by the conversion flow as input.
BrainChip Holdings Ltd, a neuromorphic computing solutions provider, has reportedly announced the availability of its powerful neutral network converter which allows users to easily convert current convolutional neural networks (CNNs) to an event oriented spiking neural network (SNN) compatible with Akida. The Akida Development Environment (ADE) is integrated in the converter to offer network simulation and conversion.
Reportedly, the integrated flow represents the first commercially offered development environment of the world. The network converter enables implementation of both SNN and CNN on the same hardware device with maintaining the power benefits and inherent performance of event-based neural networks. Higher performance can be attained by users with a native SNN and quicker time-to-market by utilizing the ADE and the CNN to SNN converter.
The latest CNN to SNN modification flow uses standard text files and is created for effortless usage. Many CNN architectures aimed at edge applications such as robotics, image processing, key word spotting, anomaly detection, and ADAS can be implemented by users. The conversion maintains near full precision with improved performance while reducing neural network computational operating cost.
Post-conversion, the whole network is implemented within the neural fabric of the Akida chip, which signifies elimination of the host computational necessities of the neural network. The host provides the data to the chip and retrieves the outcomes.
Apparently, a standard CNN description is taken by the conversion flow as input. The user changes the input through a logical method to Akida compatible layers. The modified network description is then managed via standard quantization and training.
To improve performance, the ADE aids programmable multiple bit-widths including 4-bit, ternary and binary for both activations and weights in every network layer. When the final network configuration is attained, the subsequent Akida compatible network description is outcome in industry standard .dat and .yml files. The performance information is generated from Akida device by running these files in the Akida emulation environment.
Source credit: https://www.brainchipinc.com/news-media/press-releases/detail/80/brainchip-introduces-a-powerful-neural-network-converter