From Y-DATA to ACDL 2019
Hi, my name is Ilia Kravets and I’m going to tell you what I did last summer.
I’ll start right where my almost-a-year-long studies at Y-DATA program ended: At Y-DATA Industry Projects Demo Day, where Tal Heletz, Roman Gurevich and myself won the best project award.
Our project dealt with Pulmonary Nodules detection in CT scans. The goal of the work was to improve both speed and quality of lung cancer diagnosis by radiologists by creating an end-to-end ML pipeline to assist their work. Our pipeline first detects potential tumors (pulmonary nodules) and then retrieves similar cases from a historical database, containing both diagnosis and further disease development history.
The prize was the best thing a machine learning freshman could dream of — more ML studies! Yandex sponsored our participation in 2nd Advanced Course on Data Science & Machine Learning, a.k.a. ACDL 2019, and a few weeks later we found ourselves on a plane to Italy.
The course venue is nothing like your usual campus or conference venue. Certosa di Pontignano is a real medieval monastery situated in a rural Tuscany near Siena, Italy. We joked that for this week we became monks worshiping ML gods. And so the prophets came down to us (down the path of steepest descent, of course) in the form of top-notch ML professors and industry experts to teach us the Word of Machine Learning. But that’s enough theology for now — let’s move to what really went on in ACDL. So as not to bore you, I’m going to choose and briefly describe only seven of the topics covered in the course.
- Josh Tanenbaum from MIT introduced us to computational cognitive science and described how he uses probabilistic programming to “reverse-engineer the cognition” and allow the computers to learn a little bit more like humans do.
- Leslie Kaelbling from MIT took the opposite approach and described how she makes robots learn and perform complex tasks without striving to emulate human cognition.
- Phillip Isola from MIT talked about generative models, and generative adversarial networks in particular. Fellow Y-Daters may know him as a co-author of the CycleGAN paper we studied back in class.
- Ruslan Salakhutdinov from CMU and Apple talked about natural language processing and incorporation of domain-knowledge in deep learning.
- Joaquin Vanschoren from TU/e took us all the way from introduction to AutoML and hyperparameter optimization to network architecture search and meta-learning.
- Naftali Tishby from HUJI blew everyone’s mind with the information-theoretical approach he invented to analyze and explain deep neural networks’ performance.
- Oriol Vinyals & Ioannis Antonoglou from DeepMind talked about the reinforcement learning advancements leading to the famous AlphaZero and AlphaStar achievements.
When we weren’t learning we had a chance to mingle with young researchers and seasoned scientists alike, discussing everything from a recent lecture topic to our life stories. One afternoon was dedicated to a guided tour around old Siena, where we saw some churches, found our way through the tiny old streets and heard stories about the Siena horse-racing tradition and the ancient rivalry between Siena and Florence.
Then there was the poster session: an evening dedicated to young researchers looking to share their recent work. Here attendees had an opportunity to show their ongoing research as well as appreciate the work of others. As many of the attendees were either academic or industry researchers, both the quality and quantity of the works presented was impressive. Unfortunately, I missed many of the presentations by the authors due to being busy presenting our own project — the one which earned us our ticket to ACDL in the first place. However, it turned out to have paid off big time: we were recently notified by the ACDL committee that our project has won ACDL 2019 Best Poster Award! Such a huge achievement for somebody who was only studying introduction to machine learning less than a year ago. Needless to say that without Y-DATA program we could only dream about this.
