2021 BRN Discussion, page-3972

This is awesome, thanks for the share Benny.

Amazing how simple concepts can demystify complex subjects like Machine Learning, quickly!

So, thinking through the cnn2snn converter, I did a quick search and came across this old Linley overview of Akida. Essentially it appears we have a few secret sauce recipes (a proprietary algorithm that converts convolution data to spikes and mesh connectivity between NPU’s) coming together to make this happen.

The new chip is built around a mesh-connected array of 80 neural processor units (NPUs), as Figure 1 shows. The interconnect is a proprietary design that supports core-to- core packet transfers as well as intrachip communications among the various function blocks. The cores employ digital logic that mimics the spiking behavior of biological neurons, but the chip also integrates spike converters, which enable it to run popular convolutional neural networks (CNNs) such as MobileNet. The converters can generate spikes from au- dio, image, lidar, pressure, temperature, and other sensor data, as well as from Internet packets and multivariate time- series data. Although Akida is primarily an inference engine, its native SNN mode allows it to learn new spiking patterns, too.

Designers can use Akida as a native SNN processor, or they can use the BrainChip CNN2SNN software to retrain their CNNs, reducing power by changing convolutions to event- based computations. Customers can directly retrain net- works built with TensorFlow and Keras. To run object- classification networks such as MobileNet, Akida’s pixel- to-spike converter implements a proprietary algorithm that the company calls high-precision convolution (HPC). This algorithm’s INT8 convolutions generate events that serve as the inputs to the NPUs. Although the company withheld details, the technique involves detecting intensity varia- tions in small regions of the image, converting the result to a spike pattern.

source: https://d1io3yog0oux5.cloudfront.net/brainchipinc/files/BrainChip+Akida+Is+a+Fast+Learner.pdf