HSE University Opens Joint Laboratory with Samsung Research
© Signature/ iStock
Samsung-HSE Laboratory will develop mechanisms of Bayesian inference in modern neural networks, which will solve a number of problems in deep learning. The laboratory team will be made up of the members of the Bayesian Methods Research Group — one of the strongest scientific groups in Russia in the field of machine learning and Bayesian inference. It will be headed by a professor of the Higher School of Economics Dmitry Vetrov.
Neural networks and Bayesian models are two popular paradigms in the field of machine learning. The first made a real revolution in the field of processing of big data, giving rise to a new direction, dubbed deep learning. The latter have traditionally been used to process small data. A new mathematical tool, developed in 2010, allows you to design scalable Bayesian models. This makes it possible to apply the mechanisms of Bayesian inference in modern neural networks. Even the first attempts to construct hybrid neuro-Bayesian models lead to unexpected and interesting results. For example, by using Bayesian inference in neural networks, it is possible to compress the network by approximately 100 times without losing the accuracy of its operation. On the other hand, in the very procedure of the approximate Bayesian inference one can also use a neural network to approach the exact a posteriori distribution. Thus, mutual penetration of the two technologies is obtained.
Neuro-Bayesian approach can potentially solve a number of open problems in deep learning: the possibility of a catastrophic over fitting for noised data, the self-confidence of a neural network even in erroneous predictions, uninterpretable decision-making, and vulnerability to adversarial attacks. All these problems are recognized by the scientific community, many teams around the world work on their solution, but there are no ready answers yet.
‘Samsung Electronics is one of the world's technological leaders. In our development we use many models of deep learning. But in order to keep up with competitors, it is not enough just to use ready-made models. We need to create new technologies of machine learning. This is all the more important because the field of deep learning has not yet "settled" and every year there are new models, and existing ones quickly become obsolete,’ explains (Geunbae Lee, the Head of the AI Center, Samsung Research). ‘All this means that humanity has not yet found the optimal solution for processing big data. Therefore, cooperation with leading scientific groups in the field of machine learning and artificial intelligence in universities around the world allows us to "keep our finger on the pulse" and keep track of the latest achievements in the field, as well as get exclusive access to technologies created in partner laboratories.’
‘Samsung's decision to choose our group as a key partner in Russia, giving us the opportunity to focus exclusively on basic research, is a sign of recognition of our scientific achievements and at the same time a credit of confidence that we will try to fully justify,’ says the head of the joint laboratory and the head of the Bayesian methods research group, Dmitry Vetrov. ‘Usually large companies try to use scientists to solve specific applied problems. I am glad that our Korean partner understands the importance of research on the development of new technologies, rather than solving specific problems. Our laboratory will deal with the creation of new technologies, that is, the most interesting from the point of view of the scientist. Our goals completely coincide with the wishes of our partners, which serves as a guarantee of successful and long-term cooperation.’
In addition to scientific projects, the HSE-Samsung joint laboratory will actively participate in educational activities. Students and post-graduate students of the Faculty of Computer Science will be attracted to work in it. In August 2018, with the support of Samsung, the second summer school on neuro-Bayesian methods will be held. This time it will be conducted in English and several leading scientists will take part in it. The registration is still open for the summer school.
Researchers have used machine learning to discover that the two most widespread DNA structures — stem-loops and quadruplexes — cause genome mutations that lead to cancer. The results of the study were published in BMC Cancer.
Part of the Centre of Deep Learning and Bayesian Methods and another partner project between Sberbank and HSE University’s Faculty of Computer Science, the laboratory will focus on applying machine learning methods to financial services.
The HSE Faculty of Computer Science and Yandex with the support of Sberbank are to organize the 2nd International Data Analysis Olympiad (IDAO). The Olympiad is held by leading experts in data analysis for their future colleagues and aims to bring together analysts, scientists, professionals, and junior researchers.
The second Summer School on Deep Learning and Bayesian Methods (DeepBayes 2018) was held in Moscow from August 27 to September 1, 2018. The summer school was organized by HSE Centre of Deep Learning and Bayesian Methods and Samsung AI Center in Moscow. The lectures were taught by researchers from two centers-organizers, Skoltech, and Lomonosov Moscow State University.
In July, HSE is launching two English-taught courses on Coursera. Enrollement for the courses is still open. The full list of online courses offered by HSE is available on HSE MOOCs page.
On May 29, Samsung opened its new Artificial Intelligence Centre in Moscow. Dmitry Vetrov, Professor of the HSE Faculty of Computer Science, will become one of its leaders and oversee research in machine learning.
HSE and Samsung Electronics have signed an agreement to jointly implement the ‘IoT Samsung Academy’ programme. As part of this agreement, MIEM will start teaching specialists in the Internet of Things (IoT) from September 2018.
Researchers at HSE’s Laboratory of Methods for Big Data Analysis (LAMBDA) and the Yandex School of Data Analysis have significantly reduced the cost of CERN’s future SHiP detector. The detector will search for particles responsible for still unexplained phenomena in the Universe. With use of modern machine learning methods, LAMBDA and Yandex scientists came up with very effective configuration of magnets which protect the detector from background particles. As a result, the cost of the entire structure was reduced by 25%.
On February 20, the first online stage of the International Data Analysis Olympiad (IDAO) was completed. IDAO was organised by the Faculty of Computer Science of the Higher School of Economics in partnership with Harbour.Space University (Barcelona), Yandex and with the Gold sponsor, Sberbank.
The IDAO (International Data Analysis Olympiad), created by leading experts in data analysis for their future colleagues, aims to bring together analysts, scientists, professionals, and junior researchers from all over the world on a single platform. This is the first time an event of this scale will be held in Russia. The HSE Faculty of Computer Science, Yandex and Harbour. Space University organize the Olympiad with the support of Sberbank.