This blog post was originally published by Opteran Technologies. It is reprinted here with the permission of Opteran Technologies.
It’s been another tough year for anyone organising virtual high-tech conferences. Rarely do these events capture the atmosphere and buzz of a real-world face-to-face gathering. So for all of us involved in the computer vision industry we can be truly thankful to the Edge AI Vision Alliance, under the direction of the brilliant @Jeff Bier, for firmly placing the 2021 Embedded Vision summit into the “very well organised conference” category. And of course we’re not just saying that because Opteran was lucky enough to have been crowned a winner in this year’s Vision Tank start-up competition. Indeed not! Rather we are all a bit geeky and just love to see the latest in computer vision technology and to hear robust debates from industry leaders and competitors.
So with our CEO David Rajan navigating his way through multiple rounds of interviews and start-up pitches to the tech dragons, the rest of us took the opportunity to jump into the many interesting sessions and virtual demonstrations to see what was cookin’. As well as wanting to see the latest developments in Edge-based AI devices, I was also deeply intrigued to hear the keynote from Covariant’s Co-Founder @Pieter Abbeel (if you’re not hooked on his “The Robot Brains” podcast then you should be). Pieter did not disappoint, offering more thoughts and insight into Convariant’s approach to Reinforcement Learning and its application to some difficult robotics problems.
The 3 days were also packed with great sessions and virtual demonstrations, and of course it was no surprise after a decade of technical innovation in the application of neural networks for computer vision, that the conference was heavily dominated again by deep learning problems and solutions.
So in our post-conference celebrations and ‘what we learned’ mop-up a few of us had taken the same note from Jeff during one of the fireside chats where he said:
“I’ve also seen a bias towards using deep learning when a problem could be solved less expensively using conventional ML/CV methods. It is like peeling an apple with a sword. Unfortunately, there seems to be a wave of AI engineers who only know deep learning!”
We agree Jeff, and so do the bees!
Whilst deep learning has been hugely successful in addressing many commercial and consumer challenges it is clear there is a growing problem. The problem of “data”; And if you thought data wasn’t already too big for its own boots then think again….on show were solutions for dataset capture, dataset management, data curation, data synthesis….data…data… lots of data …because data is really big.
OK so we’re getting a bit snarky about the “data” thing but for good reason. Simply put we as humans do not learn by training on millions of images. We do not need to synthesise edge case situations to learn about those dangerous ‘black swan’ events. We humans with our 86 billion neurons don’t and neither do bees and insects with their much simpler sub-million neuron brain structures.
So at the end of a decade of deep-learning innovation we can reflect on what seems to be the only horse on the track and ask:
Q. Will deep learning keep running through the next decade?
A. Of course it will. But it could be running out of steam….or hay.
Q. Is deep learning really the only horse on the track for the next decade?
A. Of course not! That’s why we’re here.
Our response to the challenges faced by deep learning and an alternative to needing more and more data is driven by our extensive research into how nature really solves many problems from simple decision-making, to complex navigation, sensing and avoiding obstacles.
Take the navigation problem. As humans, if we are navigating from one room to another we do not need to build a detailed metric SLAM map to safely navigate any obstacles along the way. Rather our brains have a map of connected spaces that means we can do it without consciously thinking about it.
While a bee is much simpler than the human brain, examining how an insect brain works shows how nature has found elegant, ultra ultra low power solutions to these problems…..ones without the need for a data centre. At Opteran we have been able to reverse-engineer this capability into algorithms and onto silicon, which require orders of magnitude less power than equivalent deep learning solutions and don’t require extensive training. The potential for these light-weight, effective solutions is bigger than big….huge. We believe our natural-intelligence approach will create a vast array of new opportunities to help machines see, sense, act and make decisions.
So with a lot of smart research, researchers and of course 600 million years of evolution behind us we’re confident at Opteran we have the breakthrough that will address a myriad of problems deep learning solutions are facing today and have been clearly visible at the Embedded Vision Summit.
So what does the next decade of innovation hold for us all? Well nature has always found much more efficient solutions that when translated effectively into technology will create a very different future for machine autonomy. Watch this space for more thoughts on the future of computer vision and autonomy at the edge…
Chief Product Officer, Opteran Technologies