Календарь событий
семинар "Структурные модели и глубинное обучение": дополнительный семинар
Пятница 17 Июнь 2016, 16:00 - 17:30
Хиты : 127
от Адрес электронной почты защищен от спам-ботов. Для просмотра адреса в вашем браузере должен быть включен Javascript.
Коллеги!
На дополнительном заседании семинара "Структурные модели и глубинное обучение", которое состоится 17 июня (пятница), в 16.00, ИППИ РАН (http://iitp.ru/ru/contacts.htm), 6 этаж, 615 аудитория
будет представлено несколько коротких сообщений о различных интересных направлениях машинного обучения.
===================
Author: Denis Volhonskiy (HSE, IITP)
Title: Deep Convolutional Generative Adversarial Networks in Steganography
Annotation: Steganography is the way of hiding information within other information (called container). Steganalysis is the study of detecting messages hidden using steganography --- usually it is binary classifier (detect if there is some information / no information in container).
We propose a new model based on Deep Convolutional Generative Adversarial Networks for generating images, that could be used for more safety information embedding (in terms of steganalysis accuracy) using steganography algorithms. Our model allows increasing steganalysis error on a test set of generated images in comparison with an original test set.
===================
Author: Vladislav Ishimtsev (HSE, IITP)
Title: Conformalized density- and distance-based anomaly detection in time-series data
Annotation: Most of the world's data is streaming, time series data, where anomalies give pertinent information in critical situations; Examples abound in such fields as finance, IT, security, health and energy. However, detection of anomalies in time-series data is a difficult task, requiring detectors to process data in real-time, not batches, and learn while simultaneously making predictions.
We consider new approaches to detect anomalies in time-series data using conformalized density- and distance-based anomaly detection algorithms.
Testing and comparison of algorithms will be done on Numenta Anomaly Benchmark (NAB), which attempts to provide a controlled and repeatable environment of open-source tools to test and measure anomaly detection algorithms on streaming data.
===================
Author: Albert Matveev (HSE)
Title: Experts Aggregation for Time Series Prediction
Annotation: Time series prediction is one of the most significant problems in applied mathematics. Estimation of a future outcome of some sequence is desired in many scientific and practical fields, especially in finance. Regression-type forecasters are usually used in order to solve this problem. However, it is essential to note that for different regimes we need to use different methods. It is obvious that without knowing true future value of a time series, the learner cannot select a method and corresponding forecast, which would provide the best accuracy among available ones. Therefore, we consider a set of independent base forecasting algorithms which we call experts and formulate a problem of prediction with expert advice. We would like to construct an aggregating algorithm that will adaptively aggregate available expert forecasts so that aggregated forecast is close to the best expert in terms of loss process.
In this talk the main aggregating schemes and loss bounds for them will be presented. These aggregating algorithms will be compared on several datasets and the results of computational experiments will be provided.
===================
Author: Oleg Maslennikov (HSE)
Title: Deep Features based Image Retrieval
Annotation: Computer vision tasks are common in today's world, for example, OCR tasks (optical character recognition), license plate recognition with traffic cameras, image processing in medicine and others. One of these tasks is CBIR - content base image retrieval. Examples of such tasks are the image search in Google, recently appeared service “Findface”, which allows finding a person in social networks by his/her photo.
In the presentation we will review approaches to deep learning based solution of CBIR tasks, discuss various architectures of corresponding neural networks, as well as the choice of a metric distance and its influence on obtained results.
--
С уважением,
Евгений Бурнаев
На дополнительном заседании семинара "Структурные модели и глубинное обучение", которое состоится 17 июня (пятница), в 16.00, ИППИ РАН (http://iitp.ru/ru/contacts.htm), 6 этаж, 615 аудитория
будет представлено несколько коротких сообщений о различных интересных направлениях машинного обучения.
===================
Author: Denis Volhonskiy (HSE, IITP)
Title: Deep Convolutional Generative Adversarial Networks in Steganography
Annotation: Steganography is the way of hiding information within other information (called container). Steganalysis is the study of detecting messages hidden using steganography --- usually it is binary classifier (detect if there is some information / no information in container).
We propose a new model based on Deep Convolutional Generative Adversarial Networks for generating images, that could be used for more safety information embedding (in terms of steganalysis accuracy) using steganography algorithms. Our model allows increasing steganalysis error on a test set of generated images in comparison with an original test set.
===================
Author: Vladislav Ishimtsev (HSE, IITP)
Title: Conformalized density- and distance-based anomaly detection in time-series data
Annotation: Most of the world's data is streaming, time series data, where anomalies give pertinent information in critical situations; Examples abound in such fields as finance, IT, security, health and energy. However, detection of anomalies in time-series data is a difficult task, requiring detectors to process data in real-time, not batches, and learn while simultaneously making predictions.
We consider new approaches to detect anomalies in time-series data using conformalized density- and distance-based anomaly detection algorithms.
Testing and comparison of algorithms will be done on Numenta Anomaly Benchmark (NAB), which attempts to provide a controlled and repeatable environment of open-source tools to test and measure anomaly detection algorithms on streaming data.
===================
Author: Albert Matveev (HSE)
Title: Experts Aggregation for Time Series Prediction
Annotation: Time series prediction is one of the most significant problems in applied mathematics. Estimation of a future outcome of some sequence is desired in many scientific and practical fields, especially in finance. Regression-type forecasters are usually used in order to solve this problem. However, it is essential to note that for different regimes we need to use different methods. It is obvious that without knowing true future value of a time series, the learner cannot select a method and corresponding forecast, which would provide the best accuracy among available ones. Therefore, we consider a set of independent base forecasting algorithms which we call experts and formulate a problem of prediction with expert advice. We would like to construct an aggregating algorithm that will adaptively aggregate available expert forecasts so that aggregated forecast is close to the best expert in terms of loss process.
In this talk the main aggregating schemes and loss bounds for them will be presented. These aggregating algorithms will be compared on several datasets and the results of computational experiments will be provided.
===================
Author: Oleg Maslennikov (HSE)
Title: Deep Features based Image Retrieval
Annotation: Computer vision tasks are common in today's world, for example, OCR tasks (optical character recognition), license plate recognition with traffic cameras, image processing in medicine and others. One of these tasks is CBIR - content base image retrieval. Examples of such tasks are the image search in Google, recently appeared service “Findface”, which allows finding a person in social networks by his/her photo.
In the presentation we will review approaches to deep learning based solution of CBIR tasks, discuss various architectures of corresponding neural networks, as well as the choice of a metric distance and its influence on obtained results.
--
С уважением,
Евгений Бурнаев
Место ИППИ РАН, аудитория 615
