[PhD in CS @ NYU] [Doctorat info @ NYU]
[Résumé (en)] [CV (fr)] [Projects] [Projets]
[Publications (en)] [Publications (fr)]
[Passions (en)] [Passions (fr)]
[Photos (en)] [Photos (fr)]

Dynamic Factor Graphs

Dynamic
Factor
Graphs

Gene regulation networks

Learning
genetic
regulation
networks

Epileptic
seizure
prediction

Epileptic seizure prediction

Dynamic
topic
models

Dynamic topic models
Propagation dynamics of epileptic seizures

Propagation
dynamics
of epileptic
seizures

Scale from oriented quasi-stationary images

Scale from oriented quasi-stationary images

Prediction
of power
transformer
failures

Prediction of power transformer failures

Stereoscopic
vision for
mobile
robotics

Stereo vision for robots
Belief propagation stereo vision

Belief
propagation
stereo
vision

Portfolio optimization in the energy sector

Portfolio
optimization
in the
energy
sector

Hodgkin-Huxley
neuronal
simulator

Hodgkin-Huxley biological neuronal network simulator

Presentations at the CBLL weekly meetings:

# November 3, 2009: "Learning Realistic Human Actions", Ivan Laptev, Marcin Marszalek, Cordelia Schmid, Benjamin Rozenfeld
# May 6, 2009: "Gaussian Process Dynamical Models", Jack Wang, David Fleet, Aaron Hertzmann
# September 24, 2008: "Epileptic Seizure Prediction from EEG": review and our new successful technique
# February 6, 2008: "Temporal Lobe Epilepsy Model", based on research by Fabrice Wendling
# November 22, 2006: "Pyramidal Methods in Image Processing and Recognition and Pyramid Matching Kernels"
# October 5, 2005: "MEG Source Localization Using an MLP with Distributed Output Representation", Sung Chan Jun, Barak Pearlmutter, Guido Nolte

Dynamic Factor Graphs for time series modeling
This is my thesis research project under Prof. Yann LeCun's supervision, published in Lecture Notes in Artificial Intelligence and presented at ECML-PKDD 2009.
A DFG includes factors modeling joint probabilities between hidden and observed variables, and factors modeling dynamical constraints on hidden variables. The DFG assigns a scalar energy to each configuration of hidden and observed variables. A gradient-based inference procedure finds the minimum-energy state sequence for a given observation sequence. Because the factors are designed to ensure a constant partition function, they can be trained by minimizing the expected energy over training sequences with respect to the factors' parameters. These alternated inference and parameter updates can be seen as a deterministic gradient-based EM-like procedure.
Using smoothing regularizers, DFGs have been shown to reconstruct chaotic attractors and to separate a mixture of independent oscillatory sources perfectly. DFGs outperformed the best known algorithm on the CATS competition benchmark for time series prediction. DFGs also successfully reconstruct missing motion capture data. For more details, download the article, or follow the video lecture at ECML-PKDD 2009, with corresponding slides.

Epileptic seizure prediction using convolutional networks on patterns of EEG synchronization
This is my core applied research project, for which I have received a Google Student Award at the 2008 Machine Learning Symposium of the New York Academy of Sciences and a Young Investigator Award at the 2009 International Workshop on Seizure Prediction IWSP4. It is a collaboration between NYU's Computational and Biological Learning Lab, led by Prof. Yann LeCun, and Drs. Deepak Madhavan and Ruben Kuzniecky from the University of Nebraska Medical Center and the NYU Comprehensive Epilepsy Center.
Our patent-pending methodology achieved the best results so far on the 21-patient EEG public dataset from the University of Freiburg. Our best algorithm, consisting in convolutional networks applied to patterns of wavelet-based phase-locking synchrony succeeded in predicting all test seizures with no false positives on 15 patients out of 21. See the article published and presented at IEEE Machine Learning for Signal Processing 2008 (which you can download here) and our recent article in Clinical Neurophysiology (which you can download here), as well as the talk at IWSP 4.

Discovery of genetic regulatory networks
I contribute to a large project in collaboration with Prof. Dennis Shasha and the NYU Plant Biology lab, aimed at discovering gene regulation networks from gene/protein expression micro-array data. We employ machine learning algorithms based on dynamic factor graphs, and have so far applied these techniques to the Arabidopsis response to nitrogen (forthcoming paper).

Convolutional Networks for the prediction of the propagation of epileptic seizures
This is one of my epilepsy research projects at NYU's Computational and Biological Learning Lab, in collaboration with Dr. Deepak Madhavan from NYU Comprehensive Epilepsy Center, and Prof. Yann LeCun.
We have investigated learning the dynamics and interdependence of EEG channels at the onset of epileptic seizures, in order to help localize the epileptogenic focus. See the abstract as well as the poster (left, right, add-on) from the December 2006 American Epilepsy Society meeting, the abstract from the July 2007 AAAI conference, and the abstract from the December 2007 American Epilepsy Society meeting.

Retrieving scale from oriented quasi-stationary images
This image processing research project was done at Schlumberger-Doll Research, and has since been published in Pattern Recognition Letters (download the article). Our algorithm manages to quantify scale changes in large-scale "quasi-stationary" patterns. Instead of coming up with an orientation-independent measure, we use the local orientation field to guide the scale measure. This algorithm has been applied to Multiple-Point Geostatistics.

Robotic stereo vision: camera calibration, stereo rectification, depth matching and 3D reconstruction
(Mobile Robotics class, Fall 2006 with Prof. Yann LeCun, implemented using Lush and OpenCV library.
We hacked a Roomba (iRobot) vacuum cleaner by plugging it to a Minimac computer and 2 FireWire libdc1394 cameras).
I implemented the stereo rectification from scratches using either rotation from extrinsic camera properties or by matching SIFT features, and wrote a real-time SAD depth matching algorithm (see references and wiki pages).

Belief propagation stereo vision
(Computer Vision class, Spring 2006 with Prof. Davi Geiger). The cost function was based on image measurements: intensity accumulation and derivatives (homework 1).
That intensity derivative attribute yielded good results with edgels-based and dynamic programming-based contour detection (homework 2).
Using belief propagation, disparity maps were constructed from cost functions based on intensity derivatives, then the depth maps of the stereo images were used for 3D reconstruction (homework 3).

Biological neural network simulator based on the Hodgkin-Huxley model
(Mathematical Aspects of Neurophysiology class, Fall 2006 with Prof. Charles Peskin). It is a computer simulation of a network of conductance-based single-compartment neurophysiological model neurons, stimulated either with deterministic or stochastic external current, or by EPSP and IPSP resulting from deterministic synaptic transmission (see the project report and simulation results).

Retrospective evaluation of portfolio allocation in the energy sector
(Quantitative Risk and Portfolio Management class, Fall 2007, taught by Prof. Attilio Meucci). Starting from NYSE daily stock prices and Brent/Oklahoma oil spot prices, I conducted a retrospective portfolio allocation study on a 4-week prediction horizon, estimating missing prices (Expectation-Maximization), detecting outliers (Minimum Volume Ellipsoid), computing the Normal or Student-t distributions of invariants (weekly compounded returns), and performed a mean-variance optimal portfolio allocation. Matlab code and project report and results are available here.