Subit Chakrabarti

Compared 13 of the most popular dimensionality reduction methods based on their performance by,

1) Comparing their performance on benchmark data-sets.
2) Evaluating loss in SVM classifier performance following dimensionality reduction, and,
3) Implementing the Dimensionality Reduction+Clustering scheme on the RaceSpace Data set (http://www.cise.ufl.edu/~anand/RaceSpace.tar.xz)

The Algorithms compared were :- ISOMAP, Locally Linear Embedding, Laplacian Eigenmaps, Hessian… Show more

Compared 13 of the most popular dimensionality reduction methods based on their performance by,

1) Comparing their performance on benchmark data-sets.
2) Evaluating loss in SVM classifier performance following dimensionality reduction, and,
3) Implementing the Dimensionality Reduction+Clustering scheme on the RaceSpace Data set (http://www.cise.ufl.edu/~anand/RaceSpace.tar.xz)

The Algorithms compared were :- ISOMAP, Locally Linear Embedding, Laplacian Eigenmaps, Hessian Eigenmaps, PCA, Kernel-PCA, Independant Component Analysis, Self-Organizing Maps, Local Tangent Space Alignment, Laussian Process Latent Variable analysis, Stochastic Proximity Embedding, Diffusion Maps and Deep Autoencoders. Show less

Implemented the PRI disaggregatin algorithm from one of my previous publications (https://dx.doi.org/10.1109/JSTARS.2014.2315999) on a 64-core subsection of the University of Florida High Performance Computer using the OpenMPI framework and new metrics. Reduced computation times by upto 75 % when tested over the Lower La Plata basin region of South America.

Mentored a team of around 10 undergraduate students from University of Florida in building a 3-u nano-satellite to study the properties of a novel deployment mechanism in situations where the satellite has to undergo rapid retargeting maneuvers. The team was selected to fly in NASA's Reduced Gravity Experiment program in June 2014 to test the hypotheses we put forward. We successfully recorded the stress on the mechanism that deployment caused.

Compares the most performance of 11 of the most popular supervised learning algorithms of varying complexity on standard data-sets and subsequently, implements them on them on the very high volume HIGGS Boson Data set to understand their scalability.
Algorithms used : Kernel SVM, k-Nearest Neighbours, Naive Bayes, Random Forests, Recursive Neural Networks, Rate Distortion Auto-Encoders, Ada-Boost and a deep belief network.

Applied the Principle of Relevant Information (previously developed at CNEL) to Image Segmentation which performs better than state-of-the-art benchmarks.

Implemented hierarchial FIR filter banks to separate a noisy Electroencephalogram (EEG) signal to show that the most complex FIR system can be beaten by a simple adaptive Weiner filter.

Designed an area efficient implementation of the CORDIC (COordinate Rotation Digital Computer) algorithm which is an iterative algorithm for the calculation of the rotation of a two dimensional vector in linear, circular or hyperbolic coordinates using only adders and shifters.

Developed a machine learning algorithm for a virtual sensing interface between a tele-operated robotic hand and sensing glove. Performed integration of sensor data to characterize robotic hand movements based on real life glove gestures.

Modified Tesla's original designs for the Spark-Gap Coil by adding modern power electronics to light up a fluorescent bulb in a ~3 m radius without using exorbitant levels of power in the source coil.

Developed an optimized version the flood-fill algorithm with dynamic constraints for an autonomous vehicle capable of navigating through a maze in the shortest possible time, where the starting and end points are not predefined.

Built a real-time scheme to acquire the 56-bit SYNC code for 802.11g communications, in the presence of AWGN when neither timing or phase knowledge is available using intelligent energy change detection in the channel.