Sensetime valued at 4.5b, page-17

To summarise the article,

*Until real event-based benchmarking datasets are available, a compromise is to transform frame-based datasets (CNN) into frame-free ones (SNN). >>> OEM customers can use CNN translator to evaluate the benchmark of SNN with their existing dataset based on CNN. This is extremely brilliant as we do not have to go to these customers and say "it's lower power, it's lower latency and you can do things that you otherwise couldn't do with a CNN, which includes autonomous, supervised and autonomous unsupervised learning." As LDN said in the webinar, all he needs to give them is CNN translator and have them tested. Easier to demonstrate the performance of the technology.

*As an example, frame based dataset (CNN) was converted to event based dataset (SNN) to allow the usage of the whole 60,000 samples for testing. The result were:
- the prediction error using a deep convolutional SNN was reduced to 0.9%
  - evaluates the result five times as fast

*The following photo shows the difference between the images/datasets required by each network.

CNN: frame based (a)
SNN: frame Free / Event baased (b)