About & Contacts


Everyone knows it's impossible. But once comes an ignorant, to whom it is unknown — and he makes a discovery.

One of the fake quotes by Albert Einstein

I am a PhD student in the Machine Learning Laboratory headed by Prof. Joachim M. Buhmann in the Department of Computer Science at ETH Zurich, Switzerland.

I recieved my diploma in Applied Math and Computer Science in 2011 from the Lomonosov Moscow State University, Russian Federation, in Bayesian Methods research group headed by Dmitry P. Vetrov. I also worked in industry for Yandex and PwC.

Currently I am working on machine learning and computer vision methods for brain imaging. Keywords for my areas of interest include: feature learning, applied statistics, optimization, graphical models, time-series analysis, causal inference, etc.

My CV: link (updated 30.09.14).

You can contact me:

  • e-mail: dlaptev@inf.ethz.ch
  • phone: +41 44 632 83 60
  • drop by: CAB E 65.2, Universitaetstrasse 6, 8092, Zurich

Research in inverse chronological order


Connectom reconstruction (current main application)

The global goal of the project is to develop image processing and machine learning methods and algorithms to detect, classify and reconstruct neuroanatomical structures like neurons, dendrites and synapses from transmission electron microscopy images.

One of the major problems is about anisotropic data segmentation. We developed different approaches to improve the segmentation using dense correspondence across sections.

The idea of decomposing the anisotropic frame into "hidden" sub-frames, described in the paper "SuperSlicing Frame Restoration for Anisotropic ssTEM and Video Data".
The pipeline that joins three sections for improved segmentation, described in the paper "Anisotropic ssTEM Image Segmentation Using Dense Correspondence Across Sections".

Supervised feature learning (current)

Developing hand-crafted features for specific computer vision applications is a very hard and time-consuming process. We propose a novel algorithm called Convolutional Decision Trees, that follows the idea of Convolutional Neural Networks and learns the set of informative features in a supervised manner to better solve the given segmentation task.

Examples of the commonly used features (left) and features learned by the proposed algorithm (right), described in the paper "Convolutional Decision Trees for Feature Learning and Segmentation".

Large-scale causality inference (current)

There is no such a thing as independent random variables in real world. Most of the things we observe are dependent, and some are even causal. As a side project I am trying to play around with different algorithms and implementations, that would estimate this causal relations through learning the Bayesian Network representing the data.

Two examples of graphical structures representing the causal relations (left) and dependencies (right). The pictures are produced with bnlearn R package.

Mining areas of interest

The idea of the project was to develop a universal tuning-insensitive algorithm for mining attractive areas from the coordinates of the photos uploaded to social networks. The goal was to provide not only points, but the whole regions that can interest tourists. The project was developed for Yandex.

The pipeline of the algorithm, based on multiple kernel density estimation. Described in the paper "Parameter-Free Discovery and Recommendation of Areas-of-Interest".

Time-series anomaly detection

The goal was to develop a system that would detect, analyse and classify abnomal behavior in the Internet market measures (ejections, changes of trend, etc.). Time-series algorithm involved independent analysis of trend, seasonal component, and noise models. The project was developed in and for Yandex.

Example of discovered anomalies in the time-series of Yandex market shares.

Short term solar activity forecast

We developed an automated algorithm for short-term solar flare activity forecast. The project was done in collaboration with Microsoft Research Cambridge.

Using images of the Sun taken in different spectrum, we analysed the presence of the Sun spots and its features and the dynamics of its changes. Based on these features we trained a short-term predictor.

Magnetogram image and the region with the sun spot, processed as described in the paper "Short-term Solar Flare Forecast".

Dual decomposition in Hidden Markov Models

The idea of the research was to add a term to HMM functional that concern global prior (i.e. total length of the signal being in each of the states).

Functional, that we finally obtained, was effectively optimized using dual decomposition approach.

Comparison of different signal segmentation techniques: true segmentation (a), Viterbi segmentation (b), DD HMM segmentation (c), described in the paper "Signal Segmentation with Label Frequency Constraints using Dual Decomposition Approach for Hidden Markov Models".

ELM-algorithm for signals and images segmentation

We incorporated global prior information into EM-algorithm. For images, global prior can be equal to the total number of pixels belonging to a segment.

The task was solved using iterative process of variational approximation and transportation problem.

Satellite photo to be segmented (a), true segmentation (b), the result of unconstrained segmentation (c) and of constrained segmentation (d), described in the paper "Variational Segmentation Algorithms with Label Frequency Constraints".

Publications in inverse chronological order


My Google Scholar profile: link

Selected publications

2014

  • D. Laptev, J.M. Buhmann. Convolutional Decision Trees for Feature Learning and Segmentation. German Conference in Pattern Recognition (GCPR), 2014, honorable mention, pdf, slides
  • D. Laptev, A. Vezhnevets, J.M. Buhmann. SuperSlicing Frame Restoration for Anisotropic ssTEM. ISBI, 2014, pdf, poster, link

2012

  • D. Laptev, A. Vezhnevets, S. Dwivedi, J.M. Buhmann. Anisotropic ssTEM Image Segmentation Using Dense Correspondence Across Sections. MICCAI, 2012, pdf, poster, link

2011

  • V. Chernyshov, D. Laptev, D. Vetrov. Short-term solar flare forecast. Graphicon 2011 (21th International Conference on Computer Graphics and Vision), 2011, pdf

2010

  • D. Kropotov, D. Laptev, A. Osokin, D. Vetrov. Variational Segmentation Algorithms with Label Frequency Constraints. International Conference on Pattern Recognition and Image Analysis (PRIA), 2010, pdf, link

All publications

2014

  • D. Laptev, A. Tikhonov, P. Serdyukov, G. Gusev. Parameter-Free Discovery and Recommendation of Areas-of-Interest. ACM SIGSPATIAL GIS, 2014, to be published
  • D. Laptev, J.M. Buhmann. Convolutional Decision Trees for Feature Learning and Segmentation. German Conference in Pattern Recognition (GCPR), 2014, honorable mention, pdf, slides
  • D. Laptev, A. Vezhnevets, J.M. Buhmann. SuperSlicing Frame Restoration for Anisotropic ssTEM. ISBI, 2014, pdf, poster, link
  • ... extended in ...
  • D. Laptev, J.M. Buhmann. SuperSlicing Frame Restoration for Anisotropic ssTEM and Video Data. Neural Connectomics Workshop, ECML-PKDD, 2014, pdf, slides

2013

  • K. Nekrasov, D. Laptev, D. Vetrov. Automatic Determination of Cell Division Rate Using Microscope Images. PRIA, 2013, pdf

2012

  • D. Laptev, A. Vezhnevets, S. Dwivedi, J.M. Buhmann. Anisotropic ssTEM Image Segmentation Using Dense Correspondence Across Sections. MICCAI, 2012, pdf, poster, link

2011

  • V. Chernyshov, D. Laptev, D. Vetrov. Short-term solar flare forecast. Graphicon 2011 (21th International Conference on Computer Graphics and Vision), 2011, pdf
  • D. Laptev. Searching for informative features in Magnetogram Solar images. Diploma thesis, Lomonosov Moscow State University, 2011, pdf
  • D. Laptev. Searching for active regions in the image. Lomonosov 2011 (International student, postgraduate and young scientist conference), 2011, pdf

2010

  • K. Nekrasov, D. Laptev, D. Vetrov. Automatic Detection of Cell Division Intensity in Budding Yeast. International Conference on Pattern Recognition and Image Analysis (PRIA), vol. 2, pp. 335-339, 2010, pdf
  • D. Kropotov, D. Laptev, A. Osokin, D. Vetrov. Signal Segmentation with Label Frequency Constraints using Dual Decomposition Approach for Hidden Markov Models 8th International Conference Intellectualization of Information Processing, pp. 403-406, 2010, pdf
  • D. Kropotov, D. Laptev, A. Osokin, D. Vetrov. Variational Segmentation Algorithms with Label Frequency Constraints. International Conference on Pattern Recognition and Image Analysis (PRIA), vol. 20, pp. 324-334, 2010, pdf, link

2009

  • D. Laptev. Automatic detection of the number of dividing cells based on the microscope images. Graphicon 2009 (19th International Conference on Computer Graphics and Vision), 2009 First prize in “Young scientists” nomination, pdf

Experience


Conferences & workshops (publication activity not included)


Industry jobs

  • Yandex. 04.2011 - present time, Analyst-Mathematician.
  • Development of predictive and explaratory ("what-if") models. Analysis of trends, events and influencing factors. Segmentation and description of markets and users. Automated spatial and temporal data analysis.
  • PricewaterhouseCoopers. 10.2010 - 03.2011, Systems and Process Assurance Consultant.
  • Business process description and formalization. Audit of processes and IT infrastructure. Automated outliers detection.

Teaching and thesis supervision at ETH Zurich

  • Computational Intelligence Lab SS12, SS13, SS14
  • Introduction to Machine Learning AS12, AS13, AS14
  • Master thesis "Automated segmentation of neuronal structures using 3d Markov Random Fields with biologically-driven shape priors" by Roland Bärtschi

Reviewing for

Code & Software


Robust PCA implementation for MATLAB (with examples)


A very compact and illustrative implementation of ADMM optimization method for Robust PCA problem. Three applications are covered in the examples: simple matrix decomposition, image inpainting (image recovery) and video decomposition (separating foreground from background).


Variance Gamma module for python


A package for Python to work with Variance Gamma distribution. Contains probability density and cumulative distribution functions, random point generator, and two methods for parameter fitting (method of moments and maximum likelihood method).

Other activities


I try to help other people and ask for help using StackExchange. Check my profiles in CrossValidated and in StackOverflow. And here is my flairtm badge:


I regularly attend Machine Learning and Data Science Meetup orginised by our Institute, but gaining more popularity outside of Academia. You are welcome to join!

Machine Learning and Data Science Meetup


I appreciate the idea of ResearchGate to openly share knowledge. So please...

Dmitry Laptev


I think that Bitcoin is an amazing concept. Science and geeks can perfectly work together, slowly changing the world.

Bitcoin


One of my hobbies is mobile development for fun. So far I have developed: StereoCrack for iOS and Android - a game for your imagination and brains.