A saying practice makes it perfect, can be perfectly applied not only for usual mundane situations but for every “living” thing.
Even the greatest inventions like AI are quite dumb, says today’s guest, therefore we have to train them to do a very specific task. For this reason, mathematicians created optimizers – algorithms or methods used to change the attributes of your neural network.
Today’s interviewee, Nick Romano, CEO of Deeplite – a provider of intelligent AI optimization software – was invited by the Cybernews team, in hopes to find out what ideas are hiding behind this technology and what benefits it can bring.
How did Deeplite originate? What has your journey been like?
Deeplite began in a start-up accelerator in Montreal called Tandem Launch. Two founders, Ehsan Saboori and Davis Sawyer, worked on the initial intellectual property for the Deeplite project.
About eighteen months later, I exited my last company through a private equity transaction and was introduced to Tandem Launch, which then introduced me to Ehsan and Davis. I invested some of my own money and joined Deeplite in August 2019 as the third founder and soon after became CEO.
The journey has focused on three areas: 1) fundraising 2) product development and 3) commercial traction. As you scale up the business, you need to build each of these three components.
Can you introduce us to your software accelerator? What are its key features?
The Deeplite platform has two primary components: Deeplite Neutrino and Deeplite Runtime (DeepliteRT). Neutrino is the engine where we do all the automation to make Deep Neural Networks more optimized and efficient.
It takes AI models that are trained for accuracy but are too big, too slow, or too power consumptive for a given use case, and then we run them through our algorithms to transform the model architecture into a smaller, faster, less power-consumption version of its former self. That type of model change can be applied to any type of hardware.
DeepliteRT allows us to do hardware-aware optimization with our low-precision quantization, taking the deep neural networks and quantizing them down to two-bit and one-bit precision while preserving accuracy and being able to deploy on low-power, commodity hardware like Arm Cortex A-CPUs.
In your opinion, which industries would greatly benefit from adopting AI-powered solutions?
Many industries would benefit from AI-enabled solutions, but our primary focus right now is on computer vision and perception tasks, especially those taking place on commodity hardware.
So, industries like security and surveillance, smart homes, industrial IoT, and automotive are all a good fit, because they have situations with extreme edge inference and embedded edge inference. They are a great opportunity for our tools and solutions.
How did the recent global events affect your field of work? Have you noticed any new security issues arise as a result?
The move to being all virtual transformed how we operate as an organization, but being a software company, it was easier to adapt to that change and continue to grow and do the things that we have done.
We’ve raised two rounds of funding since the beginning of the pandemic, and though we have not even met some of the investors in person yet, we’ve still been able to grow and scale the business.
For security, since we are all virtual, info security is a big focus for us. It was important before the pandemic too, but it was even more heightened when we went all virtual. Deeplite strives to make sure that our code and customer information are secure.
Since AI is a relatively new technology, people still tend to have some misconceptions and myths regarding it. Which ones do you notice most often?
People misunderstand the power of AI and what it can actually do today. For example, the idea that AI is going to take over the world and that we’re going to lose control of it is a very common misconception.
In reality, most AI is actually quite “dumb” – that is why you have to train your AI models thoroughly, and train them to do a very specific task. Then your AI becomes good at that specific task but that is about it. It is not sentient at all.
What predictions do you have for the future of deep neural networks?
Deep Neural Networks are evolving – the models are getting better and better, and they are becoming more efficient as a result of our tools.
There are also different types of DNNs coming to market that address different tasks, so it is really evolving as a school of thought and having new and different uses.
What other aspects of our daily lives do you hope to see automized or enhanced by technology in the next few years?
AI is designed to make human decisions easier – it can take some of the decision-making out of the hands of individuals so that processes can be more reliable and predictable while allowing people to focus more on core activities.
The “human in the loop” can make more of an informed decision with the assistance of AI, rather than the human having to determine everything in regard to what happens next.
With so many connected devices nowadays, what cybersecurity measures do you think are essential for everyone to keep their devices safe?
This is an important challenge for embedded AI in particular – if the AI happens inside the device itself then the device needs to be secure. A great example is an autonomous car – it can’t be hacked and allow someone else to take over the control of it.
The security on the end device has to be protected. However, one security benefit to having AI embedded on a device is that you are not transmitting potentially sensitive data out to other locations – it stays on the device itself which can help keep it safe.
Would you like to share what’s next for Deeplite?
We continue to innovate on the platform and expand the scope and reach of what we do to more hardware types and different use cases.
We continue to grow. Our next financial goal is our next round of funding, while we continue to deliver our solutions to more customers.