top of page

The effects of noise on the neuronal interaction (2019-2021)

In neuroscience, a correct mapping of the connections between the neurons is necessary to better understand how human brain works. This is still a very challenging task and many approaches have been tried to accomplish this task. Here, we look at the interaction between the neurons under noise signals originating from unknown physiological phenomena. We develop and apply novel information-theoretic methods to analyze such effects. Among these, mutual information and transfer entropy are very powerful tools to understand the degree to which each neuron is connected with another. Moreover, by using transfer entropy, we can also analyze the cause and effect type of relationships among these neurons. In addition to the information-theoretic approaches, we make use of other statistical tools, such as the copulas and methods of machine learning.

Speech and Audio Biometrics (2016-present)

In this project, we seek betters ways of predicting some of the human properties, such as height and age, using audio records. Some of this work has been on the effects of micro features of speech, such as the Voice Onset Time and Voice Offset Time. During this work, different predictors, such as linear regression (LR), random forest regression (RF), Gaussian process regression (GPR), support-vector regression with a linear kernel (SLK), and a KNN regression (KNN) are compared using TIMIT database. In addition to this, the detection of voice disguise using formant information constitutes another challenging aspect of our project. This is a following research where the topic is conceived during the collaboration at CMU.


Digital forensics applications using computer vision and machine learning (2014-2016)

As a Postdoctoral Fellow at the Carnegie Mellon University, I worked on the development of a software package for Cyber Mission Readiness and Expertise Evaluation. It has been designed to provide an automated evaluation of cyber-security personnel at mission critical positions. I have worked on designing the computer vision methods to identify predetermined images in a recorded video stream. For this task, the relevant images are extracted using template-matching algorithms on a workstation composed of Graphical Processing Units (GPU).


Design of experiments and analysis of industrial data using advanced analytics (2012-2014)

The main focus of the work was on the design of experiments, where data were collected using different combinations of different materials used in Aluminum production at Alcoa facilities. The outputs of the experiments were interpreted statistically and chemically. With this project, our research team has been awarded three international patents so far. In addition to this, analysis of different data sets pertaining to tribology, milling and ingot casting processes obtained in the research facilities of Alcoa were analyzed by various statistical tools.


Applications of machine learning for remote sensing of aerosols (2011-2012)

At The University of Texas, I worked on identifying the reasons of biases in the Aerosol Optical Depth (AOD) measurements obtained by satellite based Moderate Resolution Imaging Spectroradiometer (MODIS) instruments and the ground-based Aerosol Robotic Network (AERONET) system. To determine factors, we designed a methodology incorporating the Artificial Neural Networks (ANN). After training the network, we utilized it to predict the AOD values and demonstrated the success of the approach by the estimation of information-theoretical quantities, such as the Mutual Information.


The assessment of climate feedback processes (2009-2011)

In order to better understand the climate feedback processes, I have designed advanced statistical tools for the Remote Sensing of Climate Group (RSCG) of NOAA-CREST at CCNY. As the traditional linear approaches are not satisfactory to elucidate the relationships between climate variables, I have proposed accurate estimation techniques of an information-theoretical quantity, called Transfer Entropy, to determine the cause and effect dependencies among different variables. I have provided an approach where three methods, namely the Kernel Density Estimation, the Piecewise-constant Bayesian model and Adaptive partitioning method, are utilized for an accurate estimation from data. I have demonstrated the successful results on the highly nonlinear and chaotic Lorenz equations.


Identifying relationships among Earth climate data (2007-2009)

This work focused on developing information-theoretic techniques to identify relevant climate variables and to quantify the spatial and temporal aspects of their interactions. I have developed a novel Bayesian approach to quantify the uncertainties in the estimations from data. I applied these techniques on data taken from the International Satellite Cloud Climatology Project (ISCCP) and analyzed the relationship between the Pacific Sea Surface Temperatures (SST) and the cloud coverage around the globe.


Characterization of interstellar organic molecules (2007-2009)

I have developed Bayesian source separation techniques to identify complex organic molecules in interstellar clouds by analyzing their infrared spectra. Separated source concentrations are provided with their error bars, illustrating the uncertainties involved in the estimation process, unlike the traditional Nonnegative Least Squares method. The approach is demonstrated on synthetic spectral mixtures using spectral resolutions from the Infrared Space Observatory (ISO) and the performance of the method was tested for different noise levels.


Sequential Bayesian modeling of non-stationary non-Gaussian processes and its application to seismic signal modeling (2002-2007)

I developed general, flexible Bayesian methods utilizing Particle Filtering, to model stochastic processes and the statistical dependencies between them. I have worked on modeling seismic signals through Bayesian modeling of time-varying autoregressive (TVAR) processes with Alpha-stable distributions. I have also developed similar approaches to enhance Synthetic Aperture Radar (SAR) images, model time-varying statistical dependencies between the components of a Vector Autoregressive (VAR) process with non-Gaussian probability distributions and to separate hidden, correlated and time-varying autoregressive sources from their mixtures.


System identification: Acoustic echo cancellation (1999-2000)

I worked on parallel structured adaptive filters to identify unknown systems. I have applied these parallel adaptive filters to remove acoustic echoes in closed environments.

bottom of page