Eyal Amir, Co-founder and CEO of Parknav, is a well respected Artificial Intelligence scientist. He received the Arthur L. Samuel Award for the best PhD thesis at Stanford’s Computer Science Department, and the world’s largest association of technical professionals, IEEE, included him on the list of “10 to watch in AI.”
For years, he was a tenured professor at the University of Illinois at Urbana-Champaign Computer Science Department where he led research on AI specializing in machine learning.
Rarely do successful professors leave their faculty positions at one of the top 5 universities. Why did you decide to pursue the unpredictable and challenging entrepreneurial world to start Parknav?
At Stanford, we were told that we can change the world for the better, meaning that the best of the world’s talent is here, and if we don’t improve the world, nobody will. This drove my desire to make an impact. After spending years working at UIUC, making an impact in science, I realized that I could do much more by creating a great product to benefit society. I saw an opportunity in the AI techniques that I helped develop, and I made that risky but exciting jump into entrepreneurship.
Why did you decide to build a product in the parking field?
From personal experience, I knew how frustrating finding parking is in Tel Aviv, Chicago, and elsewhere. From a technology perspective, I knew how to solve the problem of finding parking and believed that the AI techniques that I would develop would have a much broader impact beyond parking. I felt parking was the first step, and I thought, if we can do that, we can do much more with the same data and technology.
Was it hard to transition from academia to entrepreneurship as you didn’t have experience back in 2009 when you started the company?
It was challenging. So I looked for a co-founder and advisors at that time. There was a guy I tried to recruit as my co-founder, and he said, “Stop trying to recruit me! You’d better start taking steps to build your product.”
So, I started doing what I could then — collecting data for the product.
I walked the streets with a clipboard around my neighborhood in Chicago where I lived at that time and wrote down how many available spots there were, if they required payment, or if they were free, and any other information that affected parking. I literally hit the pavement and collected the data myself from scratch — true entrepreneurship.
It took about a month to collect at least some useful data that I could work with to build predictable models around my neighborhood.
I once had a party at my place, and I thought it was a good occasion to test my models. I went down to meet my guests with printed sheets of driving instructions for where they might find a parking spot. That helped me verify my models at that time.
Around that time, I met my co-founder Sergei Kozyrenko, who loved the idea of the product. We became more efficient at collecting data and began buying it from partners (e.g., car manufacturing and navigation companies, fleet operators, and telecom operators).
We built models of similar streets and neighborhoods and scaled it to the entire city of Chicago. We began taking into account a variety of new factors, such as time of day, nearby schools, baseball fields, parks, and many others.
One of the most challenging things related to AI products is finding the right business model. How has that worked for you?
We pivoted several business models, such as sales to real estate, sales to cities, and mobile payments. However, it turned out that the most effective business model would be licensing the technology to automotive manufacturers. Drivers would pay for the service through the cars they bought.
We were fortunate that the automotive industry woke up and started putting money on the table. I found our first automotive client in 2011 and started traveling to Munich; we had more conversations with automakers and tier-one suppliers. Today, we have several major automakers as clients
What parts of the product do you apply AI?
We combine data from many sensors (IoT devices) to predict or estimate where a parking spot will be. Another area of implementation is measuring the quality of our data and estimates. This requires non-standing techniques to determine how well the system is working.
We use a lot of Bayesian techniques for estimations and corrections as well as a lot of statistics — combinatorial algorithms — which are underestimated, in my opinion.
Combinatorial algorithms allow you to find ‘your way out of a labyrinth,’ meaning it’s a more efficient approach in terms of the output one may receive after training a model with a limited amount of data available.
We also apply neural network technologies for our video-understanding datapath.
Other approaches are used for other sensors, and we then put all those sensor models together into our multi-modal inference and prediction engine.
What are the main challenges of building an AI product?
Navigation services and apps often don’t give you the data you’d expect; specifically, such services or apps behave unpredictably, or the data isn’t updated frequently enough. We’d see a dot for the movement of a car (a signal we tried to detect where people park), but then the dot would disappear.
What did that tell you?
Various things — that a user stopped to park, turned his or her app off, lost the signal, went into a garage, or something else. There are many ways to interpret the same data.
So, you have to create statistical models for each use case and fit those cases with given information, which is more complicated.
It’s important to understand the compound errors that you get and make decisions from. You might need more data, different models for the data, or you may need to combine existing data with new data. This involves a lot of experimentation.
Our goal is to increase the precision, acquire more and better data, and make better use of such data. The precision is now around 80%.
What do you think about the hype around AI, specifically neural networks, as a top data scientist?
In 5 to 7 years, the hype around neural networks will die down. Neural networks are a set of approaches among hundreds of different AI models which are applied in the industry. Each of them typically is useful for only some tasks. To build a great product, there is no need to specifically use neural network technology; many other AI approaches can be used.