In the lead-up to the presentation of the awAIre AI Startup of the Year Award at awAIre, we’re taking some time to get to know the leaders of each of the 5 Finalists. Today we are talking with Shivy Yohanandan, CTO, Xailient:
Hi Shivy – Tell us a bit about yourself – what led you to get involved in the AI market?
I’m a neuroscientist-bioengineer with a PhD in AI, and for 8 years I worked on bionic eyes, deep brain electrodes, and brain-machine interfaces that control robotic arms using thought. My biggest revelation from neuroscience was that brains only evolved because of senses and actuators. Without senses, actuators, and brains, life would just wander the Earth randomly towards food, or wait for food to randomly wander into its mouth, which is not very efficient. So therein lies the whole purpose of senses, actuators, and brains: efficiency!
Earth needs a brain too. Think of all the wastage that happens daily. All those skyscrapers and highways burning power when not a soul is around. All those appliances trickling power in your home for no reason. All those homes with more cars than people. All those cars driving around with only one driver. All that wasted food…water…time at shopping centers searching for products…businesses open when there are typically no customers at that particular time…the list goes on! As a collective, we’ve become like an animal with no senses…wandering aimlessly while burning through tons of resources that will inevitably end one day.
It then occurred to me that, to be more efficient at managing our limited resources, this planet
needs senses and a brain . In practice, that means sensory receptors in the form of IoT devices
(e.g. cameras) and the brains in the form of AI.
The brains need to act, so AI needs to send information to actuators: turn the lights/heating off;
stock shelves at this time when it’s less busy; pick up these four people since they’re on the way to your destination; only make this much food, and so on!
So, the ultimate goal of AI IoT is to be aware of our surroundings through smart sensors which
can then respond in real-time to be more selective with how we use and manage our limited
resources. This significantly reduces business and environmental costs.
But we already have IoT sensors and AI right? What’s the problem?
The big problem with AI IoT is that state-of-the-art AI uses more resources to process IoT data
than the resources it’s trying to save, which is paradoxical.
The problem became very apparent to me in 2016 when AlphaGO beat Lee Sedol. After the
tournament, someone posted a meme showing how brutally power hungry the AI was:
1,000,000 vs 20 W, and that’s just to play Go, a very narrow AI task compared to what the
human brain is capable of. You’ve gotta look at that number again — embarrassing!
It turns out that our current approach to IoT and AI is not that dissimilar to AlphaGO. The key
cause? Expensive brute-force algorithm families like YOLO, SSD, R-CNN, and their derivatives, which account for most of the computer vision algorithms used by everyone! So there’s clearly a dangerous trend here, where we just sweep bruteforce compute under the rug and ignore cost and environmental impact!
That’s when I made it my mission to break this paradox and is what got me passionate and
determined to pursue efficient AI.
What is the company uniquely bringing to the AI market?
Xailient brings the world’s smallest computer vision AI models.
Staple computer vision algorithms throw away 95% of your data, and then struggle with
accuracy. Duh! YOLO, SSD and R-CNN do object detection by shrinking the full resolution
image to 416×416 or 300×300 and then doing both localization and classification on this
shrunken image. Higher resolutions are impractical because the models are so huge and
brute-force. But you’ve lost >95% of information from the original image, which is why accuracy is poor and not robust and doesn’t generalize across many cameras.
So, these models are already huge, and everyone is trying to fit them on the edge by shrinking
them, but only makes accuracy and generalization worse on a model that was, charitably, barely accurate to begin with.
Xailient solved this problem by cracking a 600-million-year-old secret in biology which led us to the world’s smallest computer vision AI model. At 5000x smaller than YOLO, nothing even comes close. You can then use your own flavour of classifier to process each detected ROI one-by-one but now using a crop from the original image thus preserving more information for better accuracy. So we’ve solved both model size and accuracy in one hit! This allows Xailient to fit object detection on ultra-low power devices (e.g. smart camera sensors), which is exactly what we need to break the paradox above. And now we built a platform giving everyone easy
access to this new kind of computer vision that’s much more efficient and accurate, and you don’t even need model compression!
In a nutshell, Xailient, together with our hardware partners, is bringing affordable and accessible smart sensors to the AI market thanks to a significant step-change in efficiency that’s years ahead of everyone else.
What other factors make the company compelling?
We’re hardware agnostic and can already run on commodity low power devices and micros,
making smart sensing even more accessible and affordable. We can run directly on your
camera or cloud if that’s what you need. We are a privacy-first company: no sensitive visible
spectrum data (e.g. personally identifiable information) leaves the edge!
Our philosophy is to build complex things from simple things. This is easier for enterprises to
adopt and more practical to maintain. Agile process & “ensemble AI models”, make things
“increasingly valuable and increasingly aware”.
“Big think problems” like AI require people from many different backgrounds. People who all
work and come from the same background and field are often conditioned to think a certain way. We’re a diverse and multi-disciplinary team with different backgrounds, together we can see these problems in a new light. That’s why Xailient has experts in the evolution of vision, biology, neuroscience, physics, etc. We mimic Natural Intelligence to build AI; we dont believe in brute-force & blind NAS. Instead, we use clever humans to narrow the search space for AI, which significantly speeds-up the development cycle.
Finally, the key to being a revolution catalyst is not just affordability and unit economics, but also accessibility. That’s why we have veterans in the fields of software engineering and security that make it easy to deploy our models, maintain and do it with minimal effort and maximum safety. Our CEO is ex Symantec, and Chief Architect is ex Symantec & FireEye, so data security is paramount in our operation.
How has the company been involved in the development of context & AI?
Xailient has already helped build the next generation of services that reference their
surroundings and contextual data to take action. We supply the AI, and partner with Edge
Compute leaders to bring Sensing IoT to customers so they can extract analytics to make better business decisions.
We work with a large Australian food company, Bega, to protect Australian bees from Varroa destructor mites that are killing bees everywhere else in the world. Smart Beehives monitor bees using cameras and low-power ARM chips with our edge AI. They run on solar power in the middle of nowhere with no network connectivity.
Xailient believes that today’s state-of-the-art AI is still very narrow, and hence very context sensitive, so we should pay close attention to context to drive immediate incremental business value rather than wasting time and resources chasing a wild goose like artificial general intelligence (AGI). To that end we’ve mimicked nature by adopting her ingenious ability to speciate animals to different environments and contexts. This is called Adaptive Radiation and with our proprietary AutoML training techniques, we’re able to adapt each AI model to its specific use-case and sensors.
How do you see the Context & AI market developing over the next few years?
Around 600 million years ago we had the Cambrian Explosion. It’s called an explosion because
as soon as there was a step-change in efficiency (hypothetically due to the arrival of eyes and
vision in those first animals), many different animals evolved, which led to the massive diversity
of life as we know it today. We believe the same thing will happen with a digital cambrian
explosion in smart AI sensors, and the step-change in efficiency will be the true catalyst here,
which Xailient believes we have brought to the field.
Innovators and builders will use this technology to build and innovate things that nobody even dreams of today, just like nobody could predict what kinds of life would arise from the Cambrian period. This will lead to significant reduction in business and environmental costs, and generate new business value. Context aware AI, along with affordable and accessible low-powered sensors, will usher in the 4th industrial revolution that everyone’s been waiting for: true automation!
Thanks Shivy, and best of luck with the Award !
Register FREE to see see Shivy pitch-off with the other finalised for the Award at AwAIre on June 22nd.