Bitdefender at TMLSS2018
Bitdefender's Machine Learning Unit participated with eight of its members at the first edition of the Transylvania Machine Learning Summer School that took place at the end of July 2018 in Cluj-Napoca.
The organizers are well known DeepMind figures (and also Romanians :D), Doina Precup, Razvan Pascanu and Viorica Patraucean and RIST institute from Cluj - Luigi Malago and Razvan Florian.
The motivation for the Summer School was to facilitate new relations between the Machine Learning research community of Eastern Europe with the rest of the world by encouraging collaboration between students and researchers from Eastern Europe with other more developed research centers and also by offering more accessible fees to the school attendees.
Since the school had excellent lecturers, the competition was tough. From over 800 participants, only 100 were selected to participate and to present a poster with his/hers research results or work in progress. The diversity of the participants was a keypoint in the selection, roughly half of the participants were Romanians.
Some words about the schedule. Each day we attended to several lectures, some of them introductory, other more advanced. At the end of the day we had lab sessions on Math, Computer Vision, Language, Generative Models, Unsupervised Learning and Reinforcement Learning. At the poster sessions, each of us had up to 4h30 to present the poster to anyone interested walking by. A detailed program can be found here tmlss.ro/programme.php.
Even though there were six full and exhausting days, our team really enjoyed it and tried to get the most out of the lecturers and from the interactions with the participants and the lecturers. Bitdefender was very well represented and we came back home inspired and eager to continue our research. More, we also got a Best Poster award for Computer Vision (yey!).
Actually our team enjoyed this experience so much that some of them decided to give an individual account of some of the highlights of TMLSS2018 and the lectures that gave them a long lasting impression.
Tudor Berariu:
The Transylvanian Machine Learning Summer School represents the perfect venue for all who are interested in pursuing a research career in machine learning. For those that are yet unfamiliar with the field TMLLSS gives the chance to grasp the fundamental theory of a broad set of core topics. At the same time, the more experienced participants are being offered a glimpse of the top researchers' reflections on the open problems in the field. All in all, Cluj was the place to be this July if interested in machine learning.
Amidst all those captivating lectures spanned over six days one caught my eye in particular: Ulrich Paquet's introduction to Variational Autoencoders. Although I held some lectures on the topic myself in the past years, I found Ulrich's presentation remarkable for its concise, visual, extremely clear, while at the same time strongly theoretical elucidation of what and how VAEs do achieve. I wish my lectures were half as good as his.
Later that day, while visiting the boring salt mine near Turda, I also had the chance to chat with Ulrich and ask him a few questions regarding generative models, especially Boltzmann Machines. His exact words Do not give up research on Boltzmann Machines. convinced me to recommence my efforts on training different flavours of energy-based models to structure knowledge in reinforcement learning setups.
Iulia Duță:
The main focus of the summer school was to popularize the machine learning fields in Eastern Europe, so the lectures' content was a mixture of introductory and more advanced topics. Knowing that beforehand, I arrived in Cluj both with the goal of capturing speaker's intuitions on fundamental concepts in machine learning and insights on open, more recent challenges in the field.
One of the lectures that I remarked as being well structured and created exactly for this purpose was Dumitru Erhan's presentation: "Computer Vision: Way Beyond Classification". He build the lecture incrementally, from a classical computer vision problem (Object Detection) to a much hotter topic Domain Adaptation. Both of them were presented in terms of how the field evolved over time, highlighting advantages and disadvantages of each approaches. While the first part addresses a common problem for computer vision researchers, Domain adaptation is a topic extremely important for all machine learning subfields, being a worth studying open problem. So besides useful knowledge and interactions, TMLSS showed us a glimpse of novel research and a long bibliography to follow.
Andrei Nicolicioiu:
The lecture that impressed me the most was the one about Graph Networks presented by Răzvan Pașcanu. He showed intuitions about the priors existing in deep networks and the need for models better suited for domains having strong relational priors. We were presented in a unified way different views of the domain from general graph networks to simpler graph convolutional architectures, showing the usefulness of relations both in problems involving data with a more obvious graph structure as molecules and in unexpected places like the interactions of memories in a recurrent model. The presentation certainly inspired me to study more about these approaches and how can they be applied in my research.
Dana Axinte:
Neural networks, gradients, posters, enthusiasm, magic. These are some of the settings of the first Transylvanian Machine Learning Summer School held in Cluj, Romania. Looking through the schedule and speakers, I expected the school to be strong and that I would have the chance to settle new concepts while I would engage with other participants with different backgrounds, but same interests.
I liked the balance between theoretically focused lectures and the ones that presented the industry usage of those concepts. Such pair of lectures that I particularly appreciated were the ones by Doina Precup who took a progressive approach in presenting value-based methods with function approximation for reinforcement learning and the one of Raia Hadsell who showed how deep reinforcement learning can be applied into robotics, with the challenges that exists and the possible solutions.
One of the parts of the summer school that I especially enjoyed was the panel session where some lecturers shared their opinion on different interesting topics. The reproducibility of experiments, the future of AI research and advices to increase the local AI ecosystems, as what could be the involvement of governments, companies and teachers were few of the issues discussed. There were no hype or buzzwords, but the lecturers were people who know today’s limitations, needs and questions and get things done step by step, with an engineering mindset. A fundamental feeling with which I left is that it is not only about learning, but sharing the experience and knowledge gained to empower others to create and to strengthen the community and motivation.
Conclusions
It was only the first edition, but every detail was at a very high standard. In the last days, we presented with Traian Rebedea the advantages of having the next Summer School edition here in Bucharest. Looking forward to TMLSS2019 :)