Unlike existing optimization techniques that reduce the size of a network through compression strategies such as weight pruning or weight precision reduction, DarwinAI produces an entirely new and highly compact network based on an intricate understanding of the original network obtained through AI.
DarwinAI's Generative Synthesis platform uses AI itself to understand a network at a foundational level and then employ this understanding to generate new and unique networks based on user-defined requirements.
The new networks are considerably smaller, infer faster and maintain the functionality accuracy of the original network. The potential of this approach is evidenced by one of the team’s early demos, which illustrates 110 ‘perception networks’ running on a single chip.
DarwinAI gives rise to ‘explainable’ deep learning whereby developers can understand, interpret, and defend the inner workings of a network and how it reaches its decisions –imperative for network debugging, design improvement, and addressing regulatory compliance.
This approach diverges from existing techniques, as it continuously improves its understanding of the inner workings of a network in way that not only demystifies how it reaches its decisions but reveals critical inefficiencies and imbalances in the network.
The platform runs natively in TensorFlow, the popular open source machine learning framework, and can thus be run on-premise in a customer’s private data center eliminating the need to share or upload sensitive data or proprietary models to a public cloud environment.
DarwinAI provides a vendor-neutral approach, a key differentiator over alternative Automatic Machine Learning techniques, which often must run in specific cloud environments and not on-premise.