Eyal: Jonathan, how about some background for our readers? Tell us, how did you start your deep learning (DL) journey?
Jonathan: Sure. I’m a machine learning (ML) guy at my core. I did my PhD at the Stanford AI lab (in the pre- Deep Learning era, huh) and after graduating moved back to Israel and joined Google Tel Aviv. I was in a product team and did some ML there, but not the most cutting-edge stuff (surprisingly, Google TLV was not a good place for that at that time). I was super excited about DL and desperately wanted to get into it, so eventually I left Google to join the Deep Learning revolution. I immediately started to learn everything there was to learn on this technology, read tons of papers, and implemented some of the cool algorithms. In the last two years I spent my time doing POCs and consulting for many startups (PointGrab being the most predominant one), but never committed to any single company. Eventually, the hunger to work in a real research environment started to creep up on me, and I began considering my options. Inbal Tal, a good friend whom I trust, introduced me to Eyal Gura.
Eyal: Nowadays with ML\DL being so hyped and also a part in other “cool” segments of the tech industry such as smart home or drones, I am sure you had many other offers, what “caught” you in zebra?
Jonathan: There are numerous reasons why I joined Zebra, and I can go ahead and list them for you, but let’s start with emotional reason. In a wise move, you (Toledano) put me in your research mailing group right after interviewing me. Once in awhile I would get an email from it, a researcher would describe a new paper or a new open-source library. Usually there would be a discussion and I would secretly listen-in like a facebook stalker. I don’t think the team knew I was there.
What I saw was a group of smart people trying to tackle something big together. They were not yet very experienced in Deep Learning, they were learning the ropes. They summarized papers for each other. They had discussions regarding the new features of Tensorflow. They were debating the right way to launch experiments on a cluster of GPUs, and on which version of the Inception network to train. These team interactions felt like the start of something big, and that’s what drew me in.
Eyal: Last year the team was very young and new. We knew we are sitting on a rare opportunity in terms of data and impact but indeed it was a challenge to build a great ML\DL team.
Jonathan: The realization that a research team was forming up, united not by the need to publish papers or make money, but by the desire to solve a challenging problem, and by the conviction that Deep Learning was the way to go – that’s what caught me. I felt excited knowing how super-relevant my knowledge and experience could be for this team and the impact I could make.
And also, this wasn’t the Zebra that I saw just one year before (back then you guys just offered your platform as a service). You shifted your strategy, and you convinced me that you were serious about this research team for the long run, and you were willing to invest what it takes for this team to grow, including letting the researchers take their time to study new technologies, go to conferences, and write papers.
Eyal: As the tech world becomes more connected and sometime cynical, every other developer these days does “something with deep learning” and it can indeed be applied in AdTech \ FinTech \ Cars \ etc. I know what drew me back to Israel to join Gura and start this but also wonder what did you feel when you learnt about what we are working on, and the potential impact on society?
Jonathan: For some people the potential impact alone on saving people’s life and working in the medical field is a thrilling-enough factor to get up in the morning. Not for me. I want to do good, and Zebra’s goal is definitely nobel, but more importantly I want to solve hard (and important) problems using the best available algorithms. I see this as the most authentic joy, there is nothing cynical about it, and I completely respect those who find this joy in other fields like the one you mentioned.
Here is what gets me excited. First, we have one of the largest medical data collections in the world, containing millions of X-Ray, CT, and mammography images, as well as their associated medical reports. And second, that you had the willingness to spend the necessary funds to process that data: not just the computational resources, but also clinical ones – a home team of skilled radiologists ready to manually tag those images when necessary. Deep Learning with my own private gold-mine of a dataset and more computational and clinical resources than I could have possibly imagined working on my own? Yes, please!
And, of course, there is the challenge itself. Understanding high-res visual data, in 2D and 3D. Marrying vision with language, coming to terms with the fact that many of the clinical findings are not well defined in the medical literature, and that even skilled radiologist disagree 30% of the time whether there is a tumor in the slide or not! Those are the hard research problems I’m talking about, and we employ the most cutting-edge machine learning tools to solve them (and yes, we use GANs!).
Eyal: While we were walking in the desert for 3-4 years, it seems that even within radiology, deep learning becomes a “thing” and it is no longer dismissed easily by doctors and developers. How close do you think we are to fulfill the bigger mission of bringing scalable diagnosis for hundreds of millions of patients?
Jonathan: I don’t know how much time it will take for the product to reach the masses, because ultimately that’s a business question, not a technological one. But even on the technological side, people have been talking for a while now about how Deep Learning is going to transform radiology, but it hasn’t happened yet. The DL technology is new, sure, and there were other barriers as well: logistic, regulatory, clinical, but it seems now that everything is aligned and 2017 is going to be the transforming year. All it takes now is for us to the job and make it happen.
Eyal: Any advice for fellow DL researchers?
Jonathan: I noticed that many researchers restrict themselves to pre trained networks, and they are afraid to build their own networks and train them from scratch. Don’t be. If you have enough data the performance difference is going to be negligible, but the benefit is that you now own your network’s architecture and can adapt it to utilize more available labels in your domain. Own your loss functions as well – you can do crazy things just by playing with them. Lastly, strive to operate in an environment in which you can casually launch a new training experiment, without it blocking your development. You don’t want to dismiss a useful experiment just because it takes a few days to train.