Each week one of our AI team members dedicates a few days to learn novel concepts in deep learning or software engineering and present them to the team – we call these talks “Deep Snips”.
We do this mainly by reading academic papers from leading groups, books and blog posts. We are not intimidated by anything and two of the most important values in our team are independent learning and curiosity.
We’ve covered many topics in-depth: from object detection and unsupervised learning in images, through the internal workings of ‘git’ and to an academic literature survey of code reviews and how to make them effective.
From many of these talks, we absorb immediate ideas for improving our algorithms, code or processes. Other insights just add up to our really unbelievable knowledge base, and somehow always become applicable at the right times – sometimes even a year or more after the talk has been given.
When I was a special forces team leader, my company commander used to tell us: postpone ANYTHING you want, except for operational duty and physical exercise. If you’ve run only twice this week instead of five times, you will not be able to complete the missed runs the week after. There are some things that must be done on a consistent and gradual basis to be effective.
What I’m doing today as the head of AI is so far from the bomb squad, but also always surprisingly close. Postpone ANYTHING you want – except for production-duty and cognitive-exercise.
If you’re an Olympic level learner – knowledge that you have not gained this month, you will not be able to compensate for next month by performing a blitz. There is only so much you can squeeze-in effectively in a day.
No matter how many urgent things there are today – if you are obligated to long-term success you cannot allow yourself to wake up two years from now and decide you must become a deep learning expert or a software engineering craftsman by the next month. You must build your knowledge and know-how consistently and gradually.
Cognitive Exercise in Practice
In the last two months, I’ve decided to cover an entire computer vision topic called semantic segmentation towards my next “Deep Snip”. This is a topic I already knew pretty well in-practice and from basic principles but wanted to dive in and understand the most-novel advancements thoroughly. I chose this topic because it’s clear that this is a topic that will have strategic importance for our company, even though it’s not the core part of our current algorithms.
I’ll elaborate a little bit about the learning style with which I chose to “tackle” this topic. I believe in being thorough. Very thorough. I strive to understand everything from the basic principles to the nuts and bolts of the most advanced algorithm. In my opinion, this approach, while slow in the beginning, pays off fairly quickly – and immensely.
I started by reviewing the leaderboards of the three most significant challenges (competitions between the algorithms of the leading semantic segmentation groups in the world). In some of them, the results are fragmented between several websites and I aggregated them into my own list. I started reading the papers of the 10 most highly achieving algorithms (they are rather comparable in performance, each taking a different approach, so it’s worth it to learn all of them). I read many additional papers that I reached after they were mentioned in one or more of these papers. I decided to read the additional papers that were mentioned by many of these leaders, or papers that seemed to explain basic-principles knowledge that the current paper is building on, or papers that just seemed to do something very novel and relevant even if they don’t answer the above two conditions.
Since I’m already fairly competent in reading computer vision papers after doing this many times already – it took me about 1 hour to read each of these papers very thoroughly. I did this every other morning in a coffee shop next to our office, before starting the rush of the workday.
Just to demonstrate the power of our “deep snip” habit – you can see the number of papers I ended up reading towards “deep snip” in the image, I call this the “semantic segmentation book”. I gave a 120 slide presentation for my team, which took about 4.5 hours in total. Now they also know almost everything I know after reading “the book”. We’re already working towards implementing some of these concepts into our algorithms.
I believe we’re truly lucky to be working in a company where the executive management understands and encourages such behavior. Actually, they don’t only encourage it. Our founders started doing this even before they raised the first round. This is an important aspect of this culture, but only one aspect out of many, that are common in a company where the management understands that you can postpone ANYTHING you want – except for production-duty and cognitive-exercise.
Idan Bassuk is the Head of AI at Aidoc. This post was originally posted on Medium.com.