Lecture by Dmitry Vetrov on big data math: tensors, neural networks, Bayesian inference
Today is the lecture of one of the most famous machine learning specialists in Russia, Dmitry Vetrov , who heads the department of big data and information retrieval at the Faculty of Computer Science, working at HSE with the support of Yandex.
How can I store and process multidimensional arrays in linear in memory structures? What does the training of neural networks from trillions of trillions of neurons give and how can it be implemented without retraining? Is it possible to process information “on the fly” without saving the incoming data consistently? How to optimize a function in less time than it takes to calculate it at one point? What does learning from poorly defined data provide? And why is it necessary to know mathematics well to solve all the problems listed above? And the next one.
')
People and their devices began to generate such a large amount of data that even the computational power of large companies could not keep up with their growth. And although it is impossible to work with data without such resources, people make them useful. Now we are at a stage when there is so much information that traditional mathematical methods and models become inapplicable. From Dmitry Petrovich’s lecture, you will learn why you need to know mathematics well for working with machine learning and data processing. And what kind of "new mathematics" you need for this. Presentation slides - under the cut.
Dmitry Vetrov graduated from Moscow State University, Candidate of Physical and Mathematical Sciences. Author of more than 120 scientific publications. Dmitry Petrovich has developed the “Bayesian machine learning methods” and “Graphic models” courses that he reads at Moscow State University and at the Yandex Data Analysis School. Took part in several interdisciplinary research projects on the development of new methods of machine learning and probabilistic inference (cognitive sciences, medicine, inorganic chemistry, etc.). Leads the Bayesian research team.
How can I store and process multidimensional arrays in linear in memory structures? What does the training of neural networks from trillions of trillions of neurons give and how can it be implemented without retraining? Is it possible to process information “on the fly” without saving the incoming data consistently? How to optimize a function in less time than it takes to calculate it at one point? What does learning from poorly defined data provide? And why is it necessary to know mathematics well to solve all the problems listed above? And the next one.
')
People and their devices began to generate such a large amount of data that even the computational power of large companies could not keep up with their growth. And although it is impossible to work with data without such resources, people make them useful. Now we are at a stage when there is so much information that traditional mathematical methods and models become inapplicable. From Dmitry Petrovich’s lecture, you will learn why you need to know mathematics well for working with machine learning and data processing. And what kind of "new mathematics" you need for this. Presentation slides - under the cut.
Dmitry Vetrov graduated from Moscow State University, Candidate of Physical and Mathematical Sciences. Author of more than 120 scientific publications. Dmitry Petrovich has developed the “Bayesian machine learning methods” and “Graphic models” courses that he reads at Moscow State University and at the Yandex Data Analysis School. Took part in several interdisciplinary research projects on the development of new methods of machine learning and probabilistic inference (cognitive sciences, medicine, inorganic chemistry, etc.). Leads the Bayesian research team.
Source: https://habr.com/ru/post/261245/
All Articles