Sunteți pe pagina 1din 3

Deepak Pathak

Contact
Information

Senior Undergraduate,
Dept. of Computer Science and Engineering
Indian Institute of Technology, Kanpur

Research
Interests

Computer Vision, Statistical Machine Learning, Probabilistic Inference

Education

Indian Institute of Technology Kanpur


BTech. in Computer Science and Engineering

home.iitk.ac.in/~deepakp
e-mail: deepakp@iitk.ac.in
Mobile: +91 9415416525

(2010 present)

Cumulative Performance Index (CPI) of 9.9 (on a scale of 10)

Publications

Awards and
Achievements

All India Senior School Certificate Examination, CBSE India


Scored cumulative 94.8% marks in Senior Secondary School.

(2010)

All India Secondary School Examination, CBSE India


Scored cumulative 94.2% marks in High School.

(2008)

Deepak Pathak, David Rothschild and Miro Dudik, A comparison of forecasting methods: fundamentals, polling, prediction markets, and experts, submitted to International Journal of Forecasting,
2013 (currently under peer review) [paper]

Ranked first in the institute (out of 820 students) and first in the department (out of 100
students) with a CPI of 9.9/10.0 after 6 semesters.
Received Certificate of Merit for Academic Excellence (awarded to top 5% students) for the
terms 2010-11 and 2011-12.
Awarded the Summer Undergraduate Research Grant for Excellence (SURGE) 2012.
Recipient of the CBSE Merit Scholarship under AIEEE for undergraduate studies.
Ranked in Top 0.05% (amongst 0.5 million students) in IIT-JEE 2010.
Placed in Top 0.05% (amongst over 1 million students) in AIEEE 2010.
Achieved All India Rank 42 in ISAT 2010 (Indian Institute of Space Science and Technology).
Secured 12th Rank (amongst over 0.2 million students) in UPTU Examination, 2010.
Awarded Certificate of Merit in Mathematics, Chemistry and English (awarded to top 0.1%
in the country) in All India Senior Secondary Certificate Examination 2010.

Internship

Microsoft Research, New York City


A comparison of forecasting methods: fundamental, polling, prediction markets, and experts
Mentored by Dr. Miro Dudik and Dr. David Rothschild (May 2013 - July 2013)

Objective was to develop models to produce best possible domain independent forecasts using
different data sources and contrast how they compete and combine.
Developed and implemented mathematical models to efficiently translate the prediction market prices and expectation polling data so as to capture public opinion; achieving significant
plummet in the forecasting error.
Co-authored the paper which is currently under review in International Journal of Forecasting.
Explored statistical machine learning area through research talks and interacted with various
researchers, graduate students during the internship.
Forecasting Outcomes by Modelling Correlations in Presidential Prediction Market
Mentored by Dr. Miro Dudik and Dr. David Rothschild (June 2013 - present)

Project aimed at forecasting the conditional prices of sample of securities, and overall wining
probability in Prediction Market featuring US presidential elections.
Clustered states, by modeling their capacity and tendency to react to external temporal events,
in the Eigen space of prices thus abstracting out swing states.
Formulated probabilistic model using Kalman Filters to capture dynamicity of latent factors
driving prices, taking drift into account (which is ignored in other approaches namely Independent, Rank and Diffusion Models).
Currently implementing above mentioned model for Betfair data spanned across three years.

Research
Experience

Undergraduate Thesis Unsupervised Modeling of Usual Events and Detecting Anomalies in Videos
Advised by Prof. Amitabha Mukerjee, IIT Kanpur (July 2013 - Present)

Modeled the dominant and usual behavior in videos using unsupervised probabilistic topic
models, as complement of which abnormal activities are identified.
Visual words derived from location and optical flow based features are projected to latent space
using probabilistic latent semantic analysis, yielding high accuracy in surveillance videos.
Further idea is to localize the anomalous area, and incorporate HoG-HoF features to generalize
the context of visual words. Draft of the paper is available at [webpage]
Intelligent Tutoring System : Mathematical Domains (Limits, Trigonometry and Logic)
Research Project under Dr. Sumit Gulwani, Microsoft Research, Redmond and Prof. Amey Karkare, IIT
Kanpur (July 2012 - Present)

Devised efficient system for limit problems that can give solutions and generate new problem
of similar difficulty level to the given seed, optimizing search over rule space.
Extended implementation of framework for Trigonometric Domain, empirically confirming the
non-triviality of generating solution using naive search based technique.
Improved heuristic search by modeling rules through generative logistic model in a domainindependent manner for solving trigonometric proof problems.
We plan to submit paper in top tier conference in Machine Learning and Formal Methods
in early 2014.[report]
Selected
Projects

Impact of Eye Fixation Points in Human Action Recognition


Course Project for SE367 (Cognitive Science) under Prof. Amitabha Mukerjee (Jan 2013 - May 2013)
Selected amongst the top two projects in the course.

Designed action recognition pipeline using HOG3D feature descriptors derived at points of
fixation and saccadic motions of human gaze.
Devised array of words approach in order to maintain temporal sequentiality of visual
words, thus examining against standard bag of words approach.
Concluded that entire foveated region is not quite useful, though its sub-regions have potential
interest points; this is in coherence with the covert attention in humans.[webpage]
Human Action Recognition Using 3-D Convolutional Network
Course Project for CS365 (Artificial Intelligence) under Prof. Amitabha Mukerjee (Jan 2012 - May 2012)

Proposed and implemented a 3-D convolutional neural network model which can learn spatiotemporal features and classify human actions without any prior knowledge.
Idea involved the extension of 2-D CNN algorithm, applied on MNIST image database, to 3-D
CNN where third dimension is ascribed to the temporality of frames.
Evaluated the algorithm on Weizmann Dataset obtaining an accuracy of 91.07%, significant
among the neural/deep net based techniques for video action classification.[webpage]
Improving Handwritten Character Recognition
Course Project for CS771 (Machine Learning) under Prof. Harish Karnick (July 2013 - Nov 2013)

Project aimed at recognition of 62 classes of hand written English characters and numerals.
Proposed novel hybrid approach combining strong classifier (SVM) with deep-belief networks.
Utilized features extracted at intermediate layers of shared-weighted networks, for providing
input to discriminative classifier; this led to significant improvement in accuracy.[report]
Goal Directed Navigation in Dynamic Obstacle Space and its extension to 2-Stage Search
Research Project under Prof. Arnab Bhattacharya, IIT Kanpur (May 2012 - July 2012)
Internship under Summer Undergraduate Research Grant for Excellence Program 2012

Proposed and designed an optimal path-cost and time efficient heuristic algorithm for target
oriented navigation in completely dynamic obstacle (both increasing and decreasing) space.
The time-space curve for proposed algorithm showed a sub-linear variation with respect to size
of matrix, thus giving high time efficiency.
Further designed and implemented the extension of this algorithm to 2-stage search i.e. searching in a space if future target is known and then empirically analyzed the deviation of the
proposed solution from optimal answer.[report]

Inference in Markov Networks


Term paper for CS698Q (Probabilistic Inference) under Prof. Somenath Biswas, IIT Kanpur (July 2013present)

Surveyed foundation of Markov networks from the perspective of log-linear models, motivating
conditional random fields and Ising Models.
Analyzed exact and approximate inference methods for undirected graphical models in correspondence to those in Bayesian Networks, and their wide application in computer vision.
Implementation ML-Haskell: Machine Learning Library in Haskell
Projects
Course project in CS653 (Functional Programming) under Prof. Piyush P. Kurur (Jan 2013 - May 2013)

Project aimed at implementing standard machine learning algorithms utilizing techniques of


functional paradigm.
Implemented generalized library for clustering, SVM classifier and KNN classifier, optimizing
the computational time by leveraging Haskell specific features and functionalities.
Other algorithms implemented in the project involved HMM, LSA, Naive Bayes and Logistic/Linear regressions.[repository]
Relevant
Courses

Statistical ML: Machine Learning, Probabilistic Inference, Mathematics in Machine Learning*, Statistical Simulation and Data Analysis*, Convex Optimization in Sp/COM** *, Representation and
Analysis of Random Signals (Probability Theory)**
AI: Computer Vision and Image Processing*, Artificial Intelligence Programming, Introduction to
Cognitive Science, Special Topics in Computer Science
Algorithms: Data Structures and Algorithms, Advanced Algorithms, Randomized Algorithms
Computer Science: Discrete Mathematics, Theory of Computation, Functional Programming,
Computer Networks, Database Management Systems, Compiler Design, Operating Systems, Introduction to Computer Organization, Principles of Programming Languages, Programming Tools and
Techniques
Mathematics Courses: Linear Algebra, Real Analysis, Complex Analysis, Differential Equations,
Mathematical Logic
* ongoing , ** audited

Technical
Skills

Programming Languages - C, C++, Java, Python, F#, C#, Haskell, SQL, PHP
Other Tools - R Stats, MATLAB, Weka, GNU Octave, OpenCV platform, LATEX, Lex, Yacc

Positions of
Responsibility

Asst. Coordinator, Institute Counselling Service - Led a 132 member team to organize various
activities for the institute like the Orientation Programme for 830 freshmen, workshops and confidence
building sessions by the counsellors.
Link Student - Mentored the assigned academically deficient students and provided them academic
tutoring for various courses on an individual basis. Took remedial classes for Programming.
Student Guide - Helped the new students in adapting to the environment of the Institute.
Secretary, Robotics Club - Guided students for the robotics competitions and organized workshops.
Took introductory robotics lectures for the first year students.

S-ar putea să vă placă și