Artificial intelligence has arrived at the point where most people know what it can do… but have no idea how it does it. The mere thought of building and training deep neural networks and algorithms is so daunting that many enterprises which could benefit from AI are reluctant to engage with the transformative tech.
Google’s AutoML is creating a buzz among AI watchers and the “AI-building-AI” concept is now a hot industry topic — what could be more appealing than having this mysterious new tech actually build your models for you?
Fledgling auto machine learning technologies can already automate the design of machine learning models without human supervision. The tradeoff, however, is the very high computational cost. Google AI Chief Jeff Dean estimates that eliminating human expertise when building and training ML models would require boosting compute power by a factor of one hundred.
Enter DarwinAI, a Waterloo, Ontario based AI startup which recently released a beta version of an automated machine learning solution it says can generate models ten times more efficiently than comparable state-of-the-art solutions.
The company uses a machine learning technique called Generative Synthesis, which draws on the interplay between a generator and an inquisitor pair working in concert to learn the intricate inner workings of deep neural networks and generate better, highly-efficient models.
DarwinAI Co-Founder Alexander Wong — who contributed to the creation of Generative Synthesis — told Synced his startup can “rapidly build and deploy deep learning solutions in scenarios where computational and power resources are limited, typically at the edge.” The company currently provides optimized network solutions for perception tasks such as crowd pose estimation, real-time identification of objects, and video resolution enhancement.
An Associate Professor at the University of Waterloo (UW) and Canada Research Chair in the area of Artificial Intelligence, Wong has spent more than a decade exploring the design of automated learning systems. Wong’s aim is to remove what he sees as the most significant barrier to AI’s widespread adoption: The difficulty in designing and building understandable AI systems for specific operational requirements.
“I decided the only way to truly address this challenge is to reinvent the way deep learning was developed, wherein AI itself is used as a collaborative technology that works with humans to design and build much better deep neural networks,” says Wong.
Wong founded DarwinAI in 2017 with fellow UW professor Mohammad Javad Shafiee and two of his graduate students, Francis Li and Brendan Chwyl. Also joining the team was CEO Sheldon Fernandez, whose software company Infusion was acquired by Avanade last year.
The term “AutoML” was coined in 2013 when a team of University of British Columbia researchers presented an automotive approach to solving the problem of simultaneously selecting a learning algorithm and setting its hyperparameters. The idea did not garner much attention until the end of 2016 when Google Researcher Quoc Le and his fellows put AutoML in the spotlight with their Neural Architecture Search (NAS) — the technique behind Google Cloud AutoML. Google researchers applied a reinforcement learning agent to search the architectural building blocks. Obtaining a state-of-the-art architecture however proved compute-heavy, requiring 1800 GPU days or five years on one GPU.
Unlike Google AutoML, the DarwinAI generator employs a human-crafted rough network prototype along with the usual design needs and requirements. This allows it to learn to automatically determine the best parameters at a significantly finer level of detail than human researchers are capable of.
“Designing network architectures that are most suitable for a particular task and set of design needs and requirements involves considerable guesswork. Technology like Generative Synthesis eliminates such guesswork and allows both experts and non-experts to create powerful deep neural networks in a fraction of the time and in a reliable manner,” says Wong.
DarwinAI’s generator is able to build a network that is 4.5x more computationally efficient than a state-of-the-art NASNet-L2C network produced by Google, while achieving the same accuracy. The generator also outperformed MobileNet, building a network with the same accuracy but 26x smaller. For object detection tasks, Generative Synthesis can produce a network that is more efficient than DetectNet yet 10x smaller and 4x more energy efficient on a Nvidia chipset.
DarwinAI’s beta solution targets enterprise-level users, who provide initial network designs, data, and requirements. Wong however envisions that in a not-too-distant future, DarwinAI will be able to generate its own new, optimized deep neural networks without any human assistance whatsoever.
Last month DarwinAI announced CDN$3 million in funding led by by Obvious Ventures, iNovia Capital, and angels from the Creative Destruction Lab in Toronto.
The growing and lucrative AutoML market is attracting tech giants and new players alike. And all that competition means a plucky startup like DarwinAI has to continuously maximize its resources and push its technology in order to stay in the game. It’s a challenge Wong believes DarwinAI is prepared for: “Our plan for the three to five years is to usher in the next wave of collaborative AI technologies to truly democratize AI and open up the field to non-experts and facilitate impactful AI solution across numerous verticals.”
Journalist: Tony Peng | Editor: Michael Sarazen