Nadav Cohen is a professor of computer science at Tel Aviv University and the chief scientist and co-founder of Imubit, an AI developer that is being used to optimize refineries and chemical plants around the world using its advanced Artificial Intelligence models.
You’ve been working in the AI development space for many years – how did you first get involved, and what pushed you to create Imubit?
I started my technological career in one of the army’s cyber security units, implementing certain special algorithms relating to security on real world industrial systems. During that time, I felt that I wanted to take on the mission of using special, scientific algorithms and making a physical difference in the world; I felt that machine learning is an area where you can really combine science with real world impact.
And so Imubit was eventually born as a realization of that mission. How does the company use AI to make a real world impact?
There’s a lot of manufacturing in the world, and about half of it is continuous – things that come in the form of liquid or gas. Running the manufacturing plants is a very, very complex process, it’s super complicated and hard to really optimally control. These plants have regulations that require them to record sensor data: temperatures, pressures, everything, basically on a minute-by-minute basis. Imubit takes this data and uploads it to the cloud; from the cloud, a model goes down to the site, and just starts controlling the plant.
This is a type of AI that’s super challenging. It’s not like a machine learning model, classifying an image or translating text: it is making decisions in a dynamic environment, and it needs to take into account the impact of its decisions in its effort to implement long-term strategy. That’s called reinforcement learning.
Your company takes AI technology, which many are using for machine learning applications, and runs with it in a more complex direction. Are there others in the industry who are running in that direction with you?
Developing an understanding of deep learning is by no means an individual’s task. I’m not alone in this – there’s a community. It’s not millions of people, not even thousands of people, but there are other people around the world – maybe 100 or so – and everybody works on it from a different angle. So we improve our understanding of the technology and are able to use it more safely together.
There’s a concept in the AI space called “democratization,” which essentially equates to making AI development capabilities accessible to the most people possible. What does the effort toward realizing AI democratization look like? Has the AI community been successful in implementing it?
We weren’t the first to get into the AI field – there were several people who had been working on it for a really long time. Even when it didn’t really work that well, and nobody wanted to hear about it, they were kind of persistent. It used to be the case that only people from their groups were able to train models; that’s not the case anymore.
These days, there are hundreds of thousands of people, at least, that can train simple machine learning models for simple settings. You don’t have to be a professor: You can be fresh out of school, with one year of training and you’re good – but somebody like your next door neighbor, they’re not able to train models these days. That’s one of the main goals of democratization – to enable anyone to train their own models.
I would say that, among non-machine learning experts, there are few cases of successful democratization, Imubit being one of them.