Papers
Internships
Current Projects
Completed Projects
Other Experiences
Rakesh Agrawal, Sreenivas Gollapudi, Krishnaram Kenthapadi, Nitish Srivastava, and Raja Velu, Enriching Textbooks Through Data Mining, in Proceedings of the First Annual ACM Symposium on Computing for Development (ACM DEV), Association for Computing Machinery, Inc., December 2010 to appear.[link to pdf].
Philippe de Reffye, Sebastien Lemaire, Nitish Srivastava, Fabienne Maupas, and Paul-Henry Cournede. Modeling Inter-individual Variability in Sugar Beet Populations. In Proceedings of the 2009 International Symposium on Plant Growth Modeling, Simulation, Visualization, and Applications (PMA '09). IEEE Computer Society, Washington, DC, USA. [pdf]
[paper]
The aim of this project was to find the best webpages which augment the content of each section in a book chapter. This problem was about creating an “understanding” of natural language and then identifying which webpages would add to that. We used techniques from NLP to identify keywords in the text and then pruned this set using the Bing n-gram model, which is publicly available. This gave us phrases that were candidates for being key concepts. We then used Wikipedia as the knowledge base and matched these concepts to pages. The graph induced by these pages and the internal links among them can be interpreted as a concept graph that captures what is being talked about in the text. We used PageRank on this graph to find the most relevant webpages.
[SURGE report] [code] [paper] [Digiplante Project Website]
Studied propagation of uncertainity in parameters of the GreenLab plant growth model and applied various techniques of estimation of Discrete Dynamic Systems to estimate the growth parameters required to model a plant population.
Work involved formulating a mathematical framework to integrate the various sources of variability in plant growth models. Wrote C++ code to simulate plant growth taking into account parameters for environmental and genetic variability.
Studied the propagation of error as the dynamic system evolved with time. There was good agreement between the Monte-Carlo simulations and results computed by my model.
The findings have been accepted for presentation at PMA09: The Third International Symposium on Plant Growth Modeling, Simulation, Visualization and Applications -Nov. 09-13, 2009
Deep learning aims to learn in complex domains by learning a casacade of features, each successive layer of features constituting a more abstract form the input. We study a class of deep architectures called Deep Belief Networks as a dimensionality reduction technique and as an auto-encoder. In this part of the project we attempt to formulate and solve a PAC-like learnability proof for this class of deep nets. We look at Deep Belief Networks in the perspective of other dimensionality reduction techniques such as compressed sensing and random projections and try to bound reconstruction error for a given sample complexity. We have not obtained any results yet, but have identified directions to pursue further.
Metaphors are an integral part of language and thought. According to Lakoff, abstract concepts are largely analogical mappings from complex domains to physical domains, particularly space. The structure of thought processes is hence metaphorical and a grounded understanding of spatio- temporal relations allows us to understand more complex domains. There have been few attempts to build a concrete computational model based on this idea, but these models of metaphor did not attempt to learn compact representations of sensorimotor data (known as ”chunks”). In this project, we propose a simple model based on the idea of applying existing chunks in new domains. We explain the intuition and motivation for its design. One form of metaphoric extension is lexical polysemy, where one meaning of a word is applied to other situations - e.g. “tight fit”, “tight weave”, “tight cap”, “tight discipline”. In this work, we show how a model for learning the chunk related to the concept of “tight” as in tight fit, can be extended by analogy to some of its other meanings.
Bilevel programs form a class of hierarchical optimzation problems in which the constraint set is not explicit but is defined in terms of another optimization problem. Optimization problems in several real-life domains can be expressed as bilevel programs, making such problems ubiquitous. An important application is for classification model selection in Support Vector Machines involving regularization hyper-parameters. In this term paper, we analyze a trust region approach for solving bilevel programs with non-linear objective functions over a polyhedral upper-level constraint region.
The project was aimed at learning the size and structure of an interlinked hierarchy of topics which best describes a given document corpus. We estimated the number of topics in each layer of the topic hierarchy using agglomerative clustering, used generative topic models to discover topics and linked them to form an ontology. We propose a novel method of determining subsumption relations by mapping the problem to finding max-weight clique on a sparse graph. Experiments show that the method works well on real world datasets.
We explore Intrinsic Motivation as a reward framework for learning how to perform complicated tasks. Most reinforcement learning tasks assume the existence of a critic who rewards the agent for its actions. However, in the case of humans, it is really the agent itself which works as the critic. We experiment with a model where the rewards are generated by the agent using a process which models the emotion of surprisal. The agent is modeled as intrinsically motivated to look for surprising events. We demonstrate how the agent learns under such a framework. Results show that even in the absence of any external reward, the agent is able to learn skills and perform complicated tasks.
This project was aimed at finding the best data mining techniques to verify worldwide compliance of the global ban on nuclear tests. The dataset consisted of Radioxenon measurements from five CTBTO monitoring sites. The key features of this dataset include imbalanced representation of the two classes. We experiment with several classifier types like J48 with adaboosting, single hidden layer neural networks, support vector machine and decision trees. We experiment with three methods for handling data imbalance: the Smote method for SVMs, changing the probabilty cutoff for classification and randomly upsampling and downsampling the classes. Using these methods we were able to achieve good performance on different metrics.
Developed a system that could be trained to extract relevant and informative portions of a group conversation given the recordings from the microphones of each of the participants. The project was worked upon by a team of 4 students. My major contributions were towards training a Structural SVM model to learn the HMM corresponding to the sequence in which speakers spoke. My SVM model also took into account various acoustic features such as the statistics of pitch, intensity and Mel-frequnecy cepstral coefficients to learn the characteristics of a relevant speech. We were able to build a cohesive system and were able to achieve a good recall of relevant portions in an unseen conversation.
The aim of this project was to study the dependence of pq-gram distance on its parameters. This distance is a recently proposed metric for measuring the distance between hierarchical data. We analyzed the sensitivity of the metric with respect to corpus paramters such as cluster radius, tree sparsity, fan-out and height. Our experiments and anlytical results give deeper insight into the nature of pq gram distance. A paper based on this work has been submitted to VLDB 2010 - 36th International Conference on Very Large Data Bases and is currently under review.
The motivation behind the project was to understand how important the role of prototypes is in the process of perception and identification; and how machine learning techniques can be used to identify prototypes and use them for classification. We used the MNIST database of handwritten digits. We use manifold learning using ISOMAP for fitting low dimensional manifolds to the set of images for each digit. The aim was to get a visualization of the data and to see if image points are naturally separated along human-observable high level features in this manifold. The tangent distance was used as the distance metric between two images.
Experimented with SLDA as a technique for object recognition and image segmentation. SLDA incorporates spatial information into the LDA model. It uses the co-occurrence of visual words in an image to extract probabilities of each word belonging to each latent topic. SLDA is a generative probabilistic model. We used a Variational Bayesian method for maximizing the expectation of obtaining the given dataset and obtained the probability distribution of each topic over the set of words.
Built a grid following robot to detect and count black colored squares on the grid. Used TSOP 1738 IR sensors and Atmega-16 microcontroller.
Implemented a compiler using lex and yacc to compile Modula-2 code for the MIPS instruction set.
Implemented Hough transforms in Java for transforming images containing straight line features after running an edge detection to emphasize boundaries. The transform was then used to extract a signature which could be used to identify geometric shapes with a good degree of accuracy. It could identify lines and circles with reasonable accuracy.
Implemented a GUI based Image Editor which could perform blurring, sharpening, posterization and edge-detection on images. It could be used to improve the quality of blurred images. It involved imposing different kernel functions at each pixel to modify images.
Wrote an application in Java that could plot curves and surfaces in 3D space given the parameters for the curves. It supports multi-angle viewing and scale change. It also incorporates time dependence to visualize the effect of time- dependent parameters on the curves.
Implemented a 32-bit CPU following MIPS instruction set on an Xilinx Spartan-3 FPGA unit using Verilog HDL.
Attended talks and tutorials by prominent researchers and was introduced to current active research directions in ML and vision.
Studied Combinatorics, Group Theory, Real Analysis and Linear Algebra at Indian Statistical Institute, Bangalore (one of the best mathematics institute in India) in the summer of 2008.
Internships
Current Projects
Completed Projects
Other Experiences
Papers
Rakesh Agrawal, Sreenivas Gollapudi, Krishnaram Kenthapadi, Nitish Srivastava, and Raja Velu, Enriching Textbooks Through Data Mining, in Proceedings of the First Annual ACM Symposium on Computing for Development (ACM DEV), Association for Computing Machinery, Inc., December 2010 to appear.[link to pdf].
Philippe de Reffye, Sebastien Lemaire, Nitish Srivastava, Fabienne Maupas, and Paul-Henry Cournede. Modeling Inter-individual Variability in Sugar Beet Populations. In Proceedings of the 2009 International Symposium on Plant Growth Modeling, Simulation, Visualization, and Applications (PMA '09). IEEE Computer Society, Washington, DC, USA. [pdf]
Internships
Enriching Textbooks using Data Mining
Mentors: Dr Sreenivas Gollapudi, Dr Krishnaram Kenthapadi at Search Labs, Microsoft Research Silicon Valley, Mountain View, CA, USA (May-July 2010)[paper]
The aim of this project was to find the best webpages which augment the content of each section in a book chapter. This problem was about creating an “understanding” of natural language and then identifying which webpages would add to that. We used techniques from NLP to identify keywords in the text and then pruned this set using the Bing n-gram model, which is publicly available. This gave us phrases that were candidates for being key concepts. We then used Wikipedia as the knowledge base and matched these concepts to pages. The graph induced by these pages and the internal links among them can be interpreted as a concept graph that captures what is being talked about in the text. We used PageRank on this graph to find the most relevant webpages.
Modeling Heterogeniety and Uncertainty Propagation in Plant Population Models for Parametric Identification of Sources of Inter-Individual Variability
Mentor: Dr Paul-Henry Cournede, Professor, Applied Mathematics and Systems Laboratory(MAS), Ecole Centrale Paris, France. (May-July 2009)[SURGE report] [code] [paper] [Digiplante Project Website]
Studied propagation of uncertainity in parameters of the GreenLab plant growth model and applied various techniques of estimation of Discrete Dynamic Systems to estimate the growth parameters required to model a plant population.
Work involved formulating a mathematical framework to integrate the various sources of variability in plant growth models. Wrote C++ code to simulate plant growth taking into account parameters for environmental and genetic variability.
Studied the propagation of error as the dynamic system evolved with time. There was good agreement between the Monte-Carlo simulations and results computed by my model.
The findings have been accepted for presentation at PMA09: The Third International Symposium on Plant Growth Modeling, Simulation, Visualization and Applications -Nov. 09-13, 2009
Current Projects
Deep Learning
CS 498: BTech Project - I, supervised by Dr Harish Karnick, Dept of Computer Science, IIT Kanpur. (Fall 2010) [Presentation] [Report]Deep learning aims to learn in complex domains by learning a casacade of features, each successive layer of features constituting a more abstract form the input. We study a class of deep architectures called Deep Belief Networks as a dimensionality reduction technique and as an auto-encoder. In this part of the project we attempt to formulate and solve a PAC-like learnability proof for this class of deep nets. We look at Deep Belief Networks in the perspective of other dimensionality reduction techniques such as compressed sensing and random projections and try to bound reconstruction error for a given sample complexity. We have not obtained any results yet, but have identified directions to pursue further.
Computational Modeling of Metaphors
CS 497: Special Topics in Computer Science, supervised by Dr Amitabha Mukerjee, Dept of Computer Science, IIT Kanpur. (Fall 2010) [Presentation] [Report]Metaphors are an integral part of language and thought. According to Lakoff, abstract concepts are largely analogical mappings from complex domains to physical domains, particularly space. The structure of thought processes is hence metaphorical and a grounded understanding of spatio- temporal relations allows us to understand more complex domains. There have been few attempts to build a concrete computational model based on this idea, but these models of metaphor did not attempt to learn compact representations of sensorimotor data (known as ”chunks”). In this project, we propose a simple model based on the idea of applying existing chunks in new domains. We explain the intuition and motivation for its design. One form of metaphoric extension is lexical polysemy, where one meaning of a word is applied to other situations - e.g. “tight fit”, “tight weave”, “tight cap”, “tight discipline”. In this work, we show how a model for learning the chunk related to the concept of “tight” as in tight fit, can be extended by analogy to some of its other meanings.
Trust Region Methods for Bilevel Programming
MTH 506: Optimization, Instructor Dr Joydeep Dutta, Dept of Mathematics, IIT Kanpur (Fall 2010) [Term Paper]Bilevel programs form a class of hierarchical optimzation problems in which the constraint set is not explicit but is defined in terms of another optimization problem. Optimization problems in several real-life domains can be expressed as bilevel programs, making such problems ubiquitous. An important application is for classification model selection in Support Vector Machines involving regularization hyper-parameters. In this term paper, we analyze a trust region approach for solving bilevel programs with non-linear objective functions over a polyhedral upper-level constraint region.
Completed Projects
Learning Size and Structure of Document Ontologies using Generative Topic Models
CS 397: Special Topics in Computer Science, supervised by Dr Harish Karnick, Dept of Computer Science, IIT Kanpur. (Spring 2010) [Presentation] [Report] [Final (Revised) Report (IEEE format)]The project was aimed at learning the size and structure of an interlinked hierarchy of topics which best describes a given document corpus. We estimated the number of topics in each layer of the topic hierarchy using agglomerative clustering, used generative topic models to discover topics and linked them to form an ontology. We propose a novel method of determining subsumption relations by mapping the problem to finding max-weight clique on a sparse graph. Experiments show that the method works well on real world datasets.
Exploring Effects of Intrinsic Motivation in Reinforcement Learning Agents
Course project in CS 365: Artificial Intelligence, under the guidance of Dr Amitabha Mukerjee, Dept of Computer Science, IIT Kanpur. (Spring 2010) [Presentation] [Report] [Code]We explore Intrinsic Motivation as a reward framework for learning how to perform complicated tasks. Most reinforcement learning tasks assume the existence of a critic who rewards the agent for its actions. However, in the case of humans, it is really the agent itself which works as the critic. We experiment with a model where the rewards are generated by the agent using a process which models the emotion of surprisal. The agent is modeled as intrinsically motivated to look for surprising events. We demonstrate how the agent learns under such a framework. Results show that even in the absence of any external reward, the agent is able to learn skills and perform complicated tasks.
Radioxenon monitoring for verification of the Comprehensive Nuclear Test Ban Treaty
Course project in CS 674: Machine Learning, under the guidance of Dr Harish Karnick, Dept of Computer Science, IIT Kanpur (Spring 2010) [Report]This project was aimed at finding the best data mining techniques to verify worldwide compliance of the global ban on nuclear tests. The dataset consisted of Radioxenon measurements from five CTBTO monitoring sites. The key features of this dataset include imbalanced representation of the two classes. We experiment with several classifier types like J48 with adaboosting, single hidden layer neural networks, support vector machine and decision trees. We experiment with three methods for handling data imbalance: the Smote method for SVMs, changing the probabilty cutoff for classification and randomly upsampling and downsampling the classes. Using these methods we were able to achieve good performance on different metrics.
Automated Conversation Summarization Using Speech
Winter school project at IIIT Hyderabad under the guidance of Dr Bhiksha Raj, School of Computer Science, Carnegie Mellon University, USA, Dec 2009[Presentation][Code][report][paper]Developed a system that could be trained to extract relevant and informative portions of a group conversation given the recordings from the microphones of each of the participants. The project was worked upon by a team of 4 students. My major contributions were towards training a Structural SVM model to learn the HMM corresponding to the sequence in which speakers spoke. My SVM model also took into account various acoustic features such as the statistics of pitch, intensity and Mel-frequnecy cepstral coefficients to learn the characteristics of a relevant speech. We were able to build a cohesive system and were able to achieve a good recall of relevant portions in an unseen conversation.
Indexing database of trees using pq grams
Course Project in CS 618: Searching and Indexing in Databases, under the guidance of Dr Arnab Bhattacharya, Dept of Computer Science, IIT Kanpu (Fall 2009) [Presentation] [report] [code] [paper]The aim of this project was to study the dependence of pq-gram distance on its parameters. This distance is a recently proposed metric for measuring the distance between hierarchical data. We analyzed the sensitivity of the metric with respect to corpus paramters such as cluster radius, tree sparsity, fan-out and height. Our experiments and anlytical results give deeper insight into the nature of pq gram distance. A paper based on this work has been submitted to VLDB 2010 - 36th International Conference on Very Large Data Bases and is currently under review.
Identification of Prototypical and Peripheral Handwritten Digits using Manifold Learning
Course Project in SE 367: Introduction to Cognitive Science, under the guidance of Dr Amitabha Mukerjee, Dept of Computer Science, IIT Kanpur (Fall 2009) [Final Presentation] [report]The motivation behind the project was to understand how important the role of prototypes is in the process of perception and identification; and how machine learning techniques can be used to identify prototypes and use them for classification. We used the MNIST database of handwritten digits. We use manifold learning using ISOMAP for fitting low dimensional manifolds to the set of images for each digit. The aim was to get a visualization of the data and to see if image points are naturally separated along human-observable high level features in this manifold. The tangent distance was used as the distance metric between two images.
Spatial Latent Dirichlet Allocation
Course project in CS 676 (Audit): Computer Vision and Image Processing, under the guidance of Dr Simant Dube, Dept of Computer Science, IIT Kanpur. (Fall 2009)[Presentation]Experimented with SLDA as a technique for object recognition and image segmentation. SLDA incorporates spatial information into the LDA model. It uses the co-occurrence of visual words in an image to extract probabilities of each word belonging to each latent topic. SLDA is a generative probabilistic model. We used a Variational Bayesian method for maximizing the expectation of obtaining the given dataset and obtained the probability distribution of each topic over the set of words.
DetectorX: Grid following robot
For a robotics competition in Techkriti 2008, the annual technical festival of IIT Kanpur (Feb 2008) [Competiton abstract][circuit pic]
Built a grid following robot to detect and count black colored squares on the grid. Used TSOP 1738 IR sensors and Atmega-16 microcontroller.
A Compiler for Modula-2 for the MIPS architecture
Course project in CS335: Principles of Compiler Design, under the guidance of Dr Sanjeev K Aggarwal, Dept of Computer Science, IIT Kanpur (Spring 2010) [Code]Implemented a compiler using lex and yacc to compile Modula-2 code for the MIPS instruction set.
Extraction of image signatures for identifying straight line features using Hough Transforms
Independent project, Sep-Oct 2008 [code]Implemented Hough transforms in Java for transforming images containing straight line features after running an edge detection to emphasize boundaries. The transform was then used to extract a signature which could be used to identify geometric shapes with a good degree of accuracy. It could identify lines and circles with reasonable accuracy.
Implementation of Image Editor in Java
Project at the Programming Club in IIT Kanpur, May-June 2008. [code]Implemented a GUI based Image Editor which could perform blurring, sharpening, posterization and edge-detection on images. It could be used to improve the quality of blurred images. It involved imposing different kernel functions at each pixel to modify images.
Plotting parameterized curves in 3D space
ESC101 project under the guidance of Prof. Mainak Choudhuri, Dept of Computer Science, IIT Kanpur, Fall 2007. [code]Wrote an application in Java that could plot curves and surfaces in 3D space given the parameters for the curves. It supports multi-angle viewing and scale change. It also incorporates time dependence to visualize the effect of time- dependent parameters on the curves.
Implementation of MIPS processor on FPGA unit
Course project in CS220:Computer Organization, guided by Prof. SK Mehta, Dept of Computer Science, IIT Kanpur, Fall 2008Implemented a 32-bit CPU following MIPS instruction set on an Xilinx Spartan-3 FPGA unit using Verilog HDL.
Other Experiences
Participation in 2010 Winter School on Machine Learning and Computer Vision
Organized by Microsoft Research India and Canadian Institute for Advanced Research (CIFAR) at Indian Institute of Science, Bangalore, Jan 2010[Program website]Attended talks and tutorials by prominent researchers and was introduced to current active research directions in ML and vision.
Participation in the NBHM Nurture Programme
Managed by Prof TSSRK Rao at ISI Bangalore under the aegis of the National Board for Higher Mathematics(NBHM)[Program Website]Studied Combinatorics, Group Theory, Real Analysis and Linear Algebra at Indian Statistical Institute, Bangalore (one of the best mathematics institute in India) in the summer of 2008.