Sergey Demyanov

PhD, Machine Learning Specialist, Researcher

I am interested in methods of deep learning and artificial intelligence. I obtained my PhD from Melbourne Uni. Now a PostDoc at IBM Research Australia. Check out my Worklab for Tensorflow.

Publications

04/2016

  • Demyanov, Sergey, et al. "Classification of Dermoscopy Patterns Using Deep Convolutional Neural Networks." International Symposium on Biomedical Imaging. IEEE, 2016.
    I evaluated the ability of CNN to detect skin dermoscopy patterns, related with melanoma, and compared it with other classification algorithms.

12/2015

  • Demyanov, Sergey. "Regularization methods for neural networks and related models." PhD Thesis, 2015.
    I described the existing methods of regularization of neural networks for supervised learning, proposed a new method to achieve local invariance, presented a new way of hyperparameter tuning for SVM, and considered one problem of video classification.

11/2015

  • Demyanov, Sergey, et al. "Detection of Deception in the Mafia Party Game." Proceedings of the 2015 ACM on International Conference on Multimodal Interaction. ACM, 2015.
    I described a new dataset with videos of truthful and deceptive people and presented an automatic system of feature extraction from facial cues of deception, that obtains a significant accuracy in predicting deception. These features are based on movements of the eyebrows and mouth, detected using image registration techniques.

02/2015

11/2012

Working Experience

IBM Research Australia

Post Doctoral Researcher.

09/2015 - current

07/2015 - 09/2015
(intern)

Employing the methods of deep learning for medical image analysis.

  • developing the design of experiments, conducting experiments, writing articles
  • coordinating the conduction of experiments within the group
  • proposing the ideas for patents, writing patents
  • leading the ML reading group to study advancements in the area

Glowbyte Consulting

Business Analyst

01/2011 - 09/2011

Incorporating of a decision making software package (SAS RTDM) for credit assessment in the IT system of a large bank.

  • analysing business requirements from a technical perspective
  • writing scope statements for developers
  • implementing the core parts of the proposed algorithms
  • communicating with multiple bank stakeholders on a regular basis

APITech

Business analyst, freelancer

08/2009 – 11/2010

Analysing of web-traffic sources. Developing marketing strategies.

Anycode

PHP programmer

01/2008 – 06/2008

Developing a tourism website.

Education

The University of Melbourne

Department of Computing and Information Systems

09/2011 - 09/2015

Postgraduate student

Research topics:

Machine learning techniques for automatic deception detection from facial cues, deep learning models, neural nets, support vector machines, model selection.

Supervisors:

  • James Bailey - Associate Professor, CIS Department
  • Ramamohanarao Kotagiri - Professor, CIS Department
  • Christopher Leckie - Associate Professor, CIS Department

Moscow State University

Faculty of Computational Mathematics and Cybernetics
Department of Operations Research

09/2004 - 06/2009

Undergraduate student

GPA: 4.6/5.0

Mathematical courses:

mathematical analysis, linear algebra and geometry, ordinary differential equations, discrete mathematics, methods of optimization, numerical analysis, theory of probability and mathematical statistics, optimal control theory, game theory, etc.

Computer science courses:

theory of computation, algorithms and algorithmic languages, parallel processing, computer graphics and etc.

Extra courses:

theory of optimization, digital methods of optimization, additional methods of operations research and econometric analysis of time series.

Supervisors:

  • Alexander A. Vasin - Professor, Department of Operations Research, CMC, MSU
  • Alexander A. Belolipetsky - chief of the Department of Mathematical Modelling Systems, Computing Centre of RAS

Courseworks:

  • Search for the optimal probabilities of inspection in the tax model with corruption. Marked excellent.
  • Analysis of the reaction-diffusion equation behavior near the bifurcation point. Marked excellent.

Diploma:

Analysis of the dissipative structures in a one morphogenesis problem. Marked excellent. This work were advised for the diploma works competition.

Software

Worklab for Tensorflow

Tensorflow-worklab is a set of classes and scripts that allow to easily setup the models, import pretrained models, and setup the parameters of reading, training and testing. A good set of tools for a practitioner.

ConvNet toolbox for Matlab

The toolbox written on C++ and CUDA with Matlab interface. It was one of the first open source neural network toolbox for Matlab, that can be run on a GPU. The toolbox allows training neural networks of any depth using a wide range of layer types and parameters. It has over 180 stars on GitHub.

Happy or Not

Happy-or-not is the iOS application, which uses neural networks to detect facial expressions. The application uses a set of weights from the winning submission to a Facial Expression Kaggle competition.

Skills

Programming languages

Python, Matlab, C/C++/Objective C, CUDA, SAS

Toolboxes

Tensorflow, Caffe

Other skills

GNU/Linux, SQL, Git, HTML, LaTeX

Natural languages

  • English - fluent
  • Russian - native speaker

Social Activities

Melbourne University Russian Society

President

11/2013 – 09/2014

CIS Postgraduate committee, Melbourne University

Secretary

03/2013 – 03/2014

CMC Faculty, Moscow State University

Activity organizer

01/2005 - 06/2007