User:Sahilvk87

Sahil Garg

Current Position

Research Assistant (with Dr. Amarjeet Singh) IIIT Delhi

Education

BE in Computer Science, Thapar University, 2009

Research

Machine Learning, Artificial Intelligence, Informative Sensing. Research work published on conferences AAAI, IJCAI, ICML, NIPS, UAI, KDD, IPSN, AISTATS is of high interest to me. Specifically, I follow work of Dr. Andreas Krause, Prof. C.E. Rasmussen, Prof. Zoubin, Dr. Fabio Ramos, Dr. Amarjeet Singh, Prof. Carlos Guestrin, Dr. Chris Paciorek, Dr. Michael Osborne, and ....

Research Journey with IIIT Delhi (April 2011 to July 2013)

In April 2011, under supervision of Dr. Amarjeet Singh, I started working upon problem: Understanding Real Phenomenon Dynamics with Mobile Robots using Gaussian Process Models. From empirical evaluation of stationary Gaussian process models on the problem, I understood that GP models, which employ variable hyper-parameters across input space, are required for more accurate modeling of real world space-time dynamics. So, I started exploring nonstationary space-time models, and thereby understood state of art models like Mixture of GPs, Process Convolution with Local Smoothing Kernels, Latent Extension of Input Space. Since non-separable space-time covariance functions are a better class for modeling complex space-time dynamics, we extended the general approach of Process Convolution for nonstationary nonseparable space-time GP modeling. The extended model, as it is the case for process convolution approach in general, was prohibitively expensive to train. So, as a preliminary contribution, we developed an efficient framework for modeling of correlation dynamics for the Process Convolution model. The extended Process Convolution model with efficient learning framework was published in AAAI-12 conference and a KDD-12 workshop. By now, I had started focusing upon the more general problem Efficient Low Cost Modeling of Correlation Dynamics Using Nonstationary Models. I applied the efficient learning framework, which we had applied on Process Convolution Models, on other intuitive nonstationary GP models, such as Mixture of GPs, Latent Extension of Input Space. Based upon extensive empirical evaluation of our porposed learning framework on the mentioned N-GP models, I discovered some unreliability aspects of the proposed framework. So, following work of Dr. Andreas Krause (ETH Zurich), I applied some of basic concepts like Active Learning, Mutual Information Gain, in the context of hyperparameter dynamics modeling to correspondingly model correlation dynamics of a real phenomenon. Application of these basic concepts, that have been extensively used in context of real dynamics modeling, is highly difficult in context of correlation dynamics modeling. Finally, we have developed an intuitive low cost efficient learning framework for a general class of nonstationary GP models, which performs significantly better than our previously proposed learning framework and other state of art frameworks. This work is submitted to be published in IJCAI-13 conference.

Publications

Sahil Garg, Amarjeet Singh, Fabio Ramos, "Efficient Space-Time Modeling for Informative Sensing", in SensorKDD'12: The 6th International Workshop on Knowledge Discovery from Sensor Data, Co-located with KDD'12 (published) [Author Version]

Sahil Garg, Amarjeet Singh, Fabio Ramos, "Learning Non-Stationary Space-Time Models for Environmental Monitoring", 26th Conference on Artificial Intelligence (AAAI), Special Track on Computational Sustainability and AI, 2012 (published) [Author Version]

Updates

We have submitted our research work for year 2012 to IJCAI-13 computational sustainability track. Decision due by April 2. Paper decision based upon preliminary reviews would be updated on Mar 6. The draft can be availed to Graduate Admissions Committee on request.

From Feb 2013 to June 2013, I would be working upon a journal draft that discusses all of intuitive nonstationary Gaussian Processes and then proposes a novel framework for efficient learning of the nonstationary GP models. This draft is targeted for publication in JAIR.

Statement of Purpose

Although, a signiﬁcant progress has been made in technology, our knowledge of real world phenomena dynamics is limited. Machine Learning (ML) and Artiﬁcial Intelligence (AI) will go a long way to improvise the learning and better understand the phenomena dynamics. Motivated by the possible impact from these two domains, over the past two years of my research experience, I have focused on developing Gaussian Process Models for eﬃciently modeling space time dynamics for multiple real world applications. I intend to further extend my understanding in ML and AI in my graduate studies while taking the motivation from real world domains, such as Vision, Robotics and Bioinformatics.

I ﬁnished my undergraduate (UG) education in 2009 in Computer Science major (B.E.-C.S.1 ) from Thapar Institute of Engineering and Technology 2 , which has been consistently ranked amongst top 20 engineering schools in India. As part of freshmen and sophomore UG curriculum, we were supposed to study diverse courses, at times completely unrelated to computer science. Since I could not get my interest into these courses, it aﬀected my overall performance for the ﬁrst four semesters. During my junior and senior years, I thoroughly enjoyed studying core and advanced computer science courses with signiﬁcantly better performance. My cumulative GPA for CS major was 8.49/10 (please refer to courses with code preﬁx CS, TR, PJ in the oﬃcial transcript). For easier evaluation of CS Major, an additional document “Unoﬃcial Additional Transcript for Computer Science Major GPA Evaluation- Thapar University- Sahil Garg- sahilvk87@gmail.com” is also submitted3 , as part of the application. I further ensured strong foundation in basic courses, such as Mathematics, Data Structures, Discrete Mathematics, Programming Languages, by studying advanced level books and correspondingly performed very well in advanced courses, such as Theory of Computation, Algorithms Analysis and Design, Artiﬁcial Intelligence.

As a starting point of my professional career, I preferred to join a software start up, Commdel Consulting Services. As a solo developer, I developed a core component of a product, EPayment, for parsing binary packets, sent over TCP network from a terminal device, as per conﬁgurable ISO8583 speciﬁcation, into business objects. Currently, EPayment is used to process ﬁnancial transactions worth approx. 1 billion USD every year, and played a major part in the growth of Commdel (20 software engineers in year 2009 to 110 currently). Working closely with founders of Commdel, besides the software skills, I began to highly appreciate the value of hard work, consistency, patience, risk taking, and developed aspirations for bigger goals. In June 2010, I joined a startup Snowpal Software Services as a founding developer. Snowpal was a team of three, all working remotely, including two other co-founders. I developed a RESTful API, including database design, for a product in education domain. In parallel, I joined IIIT Delhi as a Research Assistant of Dr. Amarjeet Singh in January, 2011 to work upon the problem of Nonstationary Gaussian Process Modeling for Eﬃcient Monitoring of Real World Dynamics Using Mobile Robots. Within a few months, I developed a deep interest in the ML/AI aspect of the problem. As a result, I joined IIIT Delhi as a full time researcher in December 2011. From my experience in Snowpal, I learned to solve problems independently. Unfortunately, since all of us considered Snowpal work as a secondary commitment, it could not accomplish its goals as per our expectations. It also gave me a lesson to fully commit myself to a single task to achieve the goals I set for myself.

Since April 2011, my research focus, under the supervision of Dr. Amarjeet Singh (IIIT Delhi) and remote supervision of Dr. Fabio Ramos (University of Sydney), has been to reduce the learning cost for Non-stationary Gaussian Processes (N-GPs), such as Process Convolution with Local Smoothing Kernels (Higdon, 1998), Heteroscedastic GP (Goldberg, 1997), Gaussian Process Product Model (Adams, ICML, 2008), Spatial Deformations (Schmidt, 2003), Mixture of GPs (Volker, NIPS, 2001), Latent Extension of Input Space (Pﬁngsten,2006). These models employ local hyperparameters across the entire input space so as to model correlation dynamics of a nonstationary phenomenon. Learning hyper-parameters across the entire input space is prohibitively expensive. So, as a preliminary contribution, we proposed a low cost eﬃcient framework for learning of the Process Convolution model, that was published in the proceedings of AAAI-2012 conference. The framework models the local hyper-parameter dynamics as a latent Gaussian process and learn the latent space dynamics on a sparse set of induced latent locations, selected using information gain or marginal likelihood maximization. An extension of this work, that proposed to model the local hyper-parameter dynamics under a process convolution based dependent latent GP framework, was published in the proceedings of a KDD-2012 workshop, SensorKDD-12 (see CV for details). Further, we have developed an algorithm that adaptively senses the latent space dynamics using information gain. Adaptive sensing using the proposed algorithm signiﬁcantly improves upon accurate learning of the N-GP while also reducing learning cost to a constant factor (in terms of number of hyperparameters under a joint optimization). This work is in submission for publication in Computational Sustainability track of the IJCAI-13 conference. We have also developed an algorithm for learning N-GP on very large datasets by intelligently exploiting local inﬂuence of local hyper-parameters of an N-GP using mutual information. Currently, we are working on a journal draft (JAIR or JMLR) that would present all of the N-GP models that are intuitive especially from the perspective of environmental dynamics modeling. The applicability of the proposed algorithms for learning each of the intuitive N-GPs would also be presented in the draft. Intuitively, for an environmental dynamics sensing application, we believe that both the latent space dynamics (i.e. correlation dynamics) and the real observable dynamics should be sensed in parallel in an adaptive and collaborative manner. For sensing the real observable dynamics, a physical mobile robot sensor is used, whereas, in parallel, the latent space dynamics are sensed by performing optimizations on a remote computing machine using our proposed learning algorithms. An algorithm for this speciﬁc application would be presented in the journal draft.

The research experience under the guidance of Dr. Amarjeet, helped me in 1) Signiﬁcantly improving my writing skills for technical drafts; 2) Developing a long term vision for a research problem and suitably dividing it into multiple near term objectives; 3) Developing a basic idea into a solution for a complex research problem; 4) Learning from the failures like rejection of a research paper, and using the feedback to achieve success in the future; and 5) Taking important decisions, like prioritizing sub-problems or solutions for a problem, independently. In general, from the research experience, I have now understood that computer science research is one of the best career path for me.

Considering the broad impact of ML/AI in multiple domains, I would love to take the opportunity to further work on research problems that help in improvising the state of the art in these domains while being motivated by real world challenges.