Our Investment in DarwinAI
Driving towards a safe, high performance future for AI
A University of Toronto paper published in 2012 called “ImageNET Classification with Deep Convolutional Neural Networks” outlined a fundamentally different approach to the ImageNET computer vision competition and went on to outperform previous approaches by over 30%. Since then, the deep learning approach to clustering and classification using neural networks has been applied towards many real-world applications including Go, autonomous vehicles, NLP, physics engines, and more.
However, neural networks come with significant challenges. For one, they are computationally burdensome to train and to run due to their complex and large architectures. They require high performance computing systems (such as supercomputer clusters and GPU arrays, if not emerging custom ASICs such as Intel Nervana) and are difficult to deploy on the edge because of this inefficiency. Additionally, deep neural networks require machine learning experts to delicately design and fine-tune the large, complex architectures without a full awareness of the inner workings. This is often described as the black box problem.
Coincidentally, miles away from the team behind the 2012 neural network paper, Waterloo/Toronto based DarwinAI is working to solve some of the outstanding issues of modern deep learning. The company is building an analytics package for use in model training & triage (explainability) as well as model deployment (optimization).
By using a novel approach called generative synthesis, DarwinAI is able to unpack deep neural networks, map out nodes and layers within the network, and rebuild a leaner network with similar performance.
DarwinAI is built around the “Generative Synthesis” approach to probing, studying, and generating better neural networks. This approach involves a closed loop, continuously-iterating “dialog” between two core components: the Generator and the Inquisitor. The Inquisitor probes segments of a network to study the inner workings of the network. These learnings are then translated to the Generator to learn how to create a new, better network. This cyclical learning process ultimately drives an evolution in network efficiency (Darwinism!). The process also sheds light on regions of the network — helping to inform developers on how to drive improvements to the network with additional training data.
DarwinAI has already been engaged with many companies working with deep learning throughout a number of industries. Darwin’s neural network explainability solutions have the potential to speed up model development and augment the productivity of data science teams. In addition, the optimized networks that Darwin is able to generate have consistently outperformed industry standards such as AutoML and SqueezeNet.
We are in the early stages of an AI revolution where programmatic intelligence will be introduced across many industries. Researchers confirm that neural network-driven approaches will play a large role in this technological shift. Solving two of the fundamental challenges in Deep Learning — feature obfuscation and computational intensity — are increasingly urgent needs from this rapidly developing industry. We are looking forward to supporting DarwinAI in working to drive this progress forward.
We are thrilled to announce that Obvious Ventures has partnered with iNovia Capital to support DarwinAI by co-leading a $3M Seed.
