Research Areas

Modelling of surrogates

Modelling of surrogates

Prediction systems with kernel methods such as engineering problems with surrogate functions.

Processing of 3D data

Processing of 3D data

Processing 3D data in the areas of computer vision, brain image processing, chemoinformatics.

Processing of online data

Processing of online data

Processing of online data such as anomaly detection, forecasting, change point detection, time series.

Motivation

Often, full-scale computational experiments are expensive. In addition, there are situations where “physical” models are unavailable.

Objective

To achieve this, our objetive is to construct data-driven (surrogate) models capable of:

  • predicting behavior
  • performing model-based control and recommend future actions
  • optimizing design and performance
  • detecting anomalies and predict failures

Challenges

There are several challenges involved in the construction of surrogate models such as:

  1. The data sources have viariable degrees of fidelity (e.g. AT, computer models)
  2. Complex shapes and structure present in the data at hand. Some examples are:
    • The sample is generated on a manifold
    • The data represents an anisotropic sample distribution or includes imbalanced distribution of labels
    • The description of the data relates to a 3D object such as CAD models
    • The data set is a multidimensional time-series
  3. Local accuracy prediction is required

Tasks

  • surrogate model construction and surrogate optimization;
  • anomaly detection and rare events prediction (aka imbalanced classification);
  • adaptive design of experiments (aka active learning);
  • on-line learning for streaming data.

Theoretical studies:

  1. Conformal martingales for anomaly detection in streaming data (with MSc D. Volhonsky, V. Vovk)
  2. Efficiency of conformal kernel ridge regression (with PhD I. Nazarov, V. Vovk)
  3. Multi-hypothesis testing under sparsity conditions (with Prof. G. Golubev)

Applications

  1. Conformal anomaly detection for time-series (with PhD I. Nazarov, MSc V. Ishimtsev)
  2. Meta-learning for Bayesian optimization (with PhD N. Klyuchnikov)
  3. Anomaly detection on graphs: case of different modalities (with PhD S. Ivanov)
  4. Prediction of users’ web-sites preferences using variable fidelity gaussian processes (with PhD N. Klyuchnikov)
  5. Efficient Learning of Deep Gaussian Processes (with PhD Y. Kapushev)
  6. Distributed learning of one-class classification with privileged information. Applications to IoT (with PhD D. Smolyakov)
  7. Probabilistic Adaptive Computation Time for Object Detection (with PhD A. Notchenko)
  8. Generative Adversarial Networks for Image Steganography (with PhD D. Volkhonsky and PhD I. Nazarov)
  9. CNNs on graphs for prediction of chemical properties (with PhD A. Matveyev)
  10. Resampling Recommendation Systems for Imbalanced Classification (with MSc. A. Papanov)
  11. Generative Neural Networks for Anomaly Detection in multidimensional time-series (with MSc O. Khomenko)
  12. CNNs and fusion techniques for detection of people using thermal imagery with applications to IoT (with MSc V. Osin)

Applications jointly developed with other faculty

  1. Online learning of sleeping experts for meteo data prediction (with Prof. V. Vyugin, MSc A. Maryin, MSc N. Sviridenko)
  2. Structure Models for Chemical properties prediction (with MSc R. Kostoev, Prof. A. Shapeev)
  3. Stochastic Robust Optimization of Energy Systems (with MSc N. Gryaznov, Prof. A. Bisci)
  4. CNNs for prediction of core porosity (with MSc O. Sudakov, Prof. D. Koroteev)
  5. Imbalanced classification for oilfield productivity prediction (with MSc I. Makhotin, Prof. D. Koroteev)
  6. Mobile robot self-localization from appearance data using recent results in regression on manifolds (with A. Bernstein, Robotics CREI, and MSc A. Kvasov)
  7. Time series modeling based on manifold learning technique (with A. Bernstein and PhD student Pankaj Kumar)