Nvidia provides third party validation of DarwinAI’s Generative Synthesis technology
Waterloo, ON, March 12, 2018 - DarwinAI is pleased to announce that Nvidia corp. has performed a laudatory and independent validation of its Generative Synthesis technology.
The validation report, which can be accessed here, confirms several impressive aspects of the company’s Generative Synthesis technology. Specifically, Nvidia engineers subjected an optimized deep neural network generated by Darwin's technology based on the company’s popular object detection network, DetectNet, to detailed analysis and reported that the newly generated network:
1. Contained 13 times fewer parameters than the original network
2. Required 11 times fewer multiply-accumulate operations (‘MAC’)
3. Improved inference performance 4 times on a Jetson TX2 chip
4. Maintained mean average precision within 1 point of the original network
Unlike existing optimization techniques that perform model compression on the original network – pruning neurons/synapses or reducing weight precision – Generative Synthesis leverages AI itself to observe the inner workings of a neural network during training and then automatically generate an entirely new and highly compact deep neural network.
In addition to world-class optimization, this ‘AI building AI’ approach facilitates explainable deep learning - the ability to understand how a network reaches the decisions it does.
Moreover, it allow developers to focus on the important task of getting a network to properly work under different functional scenarios, while offloading the responsibility of meeting certain performance requirements to the DarwinAI platform. This capability will be especially important for enterprises that don’t have substantial deep learning resources in-house.
DarwinAI’s optimization platform is expected to be ready in Q2 of 2018, followed by explainability features thereafter.