Back to Talks

Seminar 8 : Multi-armed Bandits Speaker : Rohit Gurjar Date : 06:00 pm, Friday, April 15th, 2011 Venue : CS102 Abstract: Multi-armed Bandits is a problem in the field of online machine learning which has been extensively studied in last two dacades. The problem provides a natural framework for the Exploration vs Exploitation tradeoff. The problem is easy to describe. You have K levers, which give some reward on pulling. You have to explore all the arms to judge which arm is better and in the end the goal is to maximize the total collected reward. The problem has various versions. We will look at some of the important results shown by Lai and Robbins [1] and Auer et al. [2][3][4]. References : T. L. Lai and H. Robbins, "Asymptotically efficient adaptive allocation rules", Advances in Applied Mathematics, 6, pages 4-22, 1985. P. Auer, N. Cesa-Bianchi and P. Fischer, "Finite-time Analysis of the Multiarmed Bandit Problem", Machine Learning, 47(2-3), pages 235-256, 2002. P. Auer, N. Cesa-Bianchi, Y. Freund and R. E. Schapire, "The non-stochastic multiarmed bandit problem", SIAM Journal of Computing, 32, pages 48-77, 2002. P. Auer and R. Ortner, "UCB Revisited: Improved regret bounds for the stochastic multi-armed bandit problem", Periodica Mathematica Hungarica, 61(1-2), pages 55-65, 2010.

Seminar 7 : Learning grounded semantics of Hindi nouns from video surveillance and user commentary Speaker : Prabhat Mudgal and Nikhil Joshi Date : 06:00 pm, Friday, April 8th, 2011 Venue : CS102 Abstract: We attempt to learn both the semantics and the labels for classes of objects in a video, using unparsed, unconstrained textual descriptions. We first discover the object categories and their models through unsupervised clustering of segmented foreground blobs from surveillance video. We then use bottom-up attention to align objects in the video with textual user narratives describing the scene. Using no domain knowledge either during the visual analysis or the language analysis, we find the maximal associations of words in the narrative with foreground objects in attentive focus. Despite the original data and the resulting clusters being noisy, we show that nouns such as (kAr),(trak) for vehicles and (sAikal) for bicycles can be learned using this approach. We test the model by generating linguistic descriptions for novel videos in the same traffic domain, and argue for the power of such grounded symbols for general language acquisition. References : S. V. P. G. Srinath, N. Joshi, P. Mudgal and A. Mukerjee, "Learning grounded semantics of Hindi nouns from video surveillance and user commentary", ICON, 2010.

Seminar 6 : Multiagent Inductive Learning Speaker : Sujith Thomas Date : 06:00 pm, Friday, April 1st, 2011 Venue : CS102 Abstract: Artificial agents may have access to different parts of the data set. But can they still come up with a hypothesis that correctly classifies most of the data? In this talk we will present the A-MAIL framework in which the agents come up which an initial inductive hypothesis which is improved through a process of argumentation and belief revision. References : S. Ontanon and E. Plaza, "Multiagent Inductive Learning: an Argumentation-based Approach", ICML, 2010.

Seminar 5 : Finding the bias and prestige of nodes in networks Speaker : Abhinav Mishra Date : 06:00 pm, Friday, March 11th, 2011 Venue : CS102 Abstract: Recent years have seen growing interest in algorithms for identifying influential nodes in a graph. Google's PageRank [1] and Klienberg's HITS [2] are two most famous algorithms. Both of these algorithms use eigenvector calculation to rank nodes. These algorithms work on the graphs with non-negative edge weights. Such structure is common in web. We will give introduction to these algorithms. By allowing edges with negative weights, modeling gets difficult. Such graphs are common in trust-based networks. We shall discuss the difficulties arising because on negative edge weights. We will give introduction to another popular algorithm EigenTrust [3] which models such graphs. Later, we will present our work [4] in this direction. References : S. Brin and L. Page, "The Anatomy of a Large-Scale Hypertextual Web Search Engine" WWW 1998. J. Kleinberg, "Authoritative Sources in a Hyperlinked Environment", SODA 1998. S. D. Kamvar, M. T. Schlosser, and H. Garcia-Molina, "The EigenTrust Algorithm for Reputation Management in P2P Networks", WWW 2003. Abhinav Mishra and Arnab Bhattacharya, "Finding the bias and prestige of nodes in networks based on trust scores," WWW 2011. View paper here.

Seminar 3/4 : The Curse of Dimensionality in Kernel Methods Speaker : Purushottam Kar Date : 05:00 pm, Saturday, February 26th, 2011 Venue : CS102 Abstract: Kernel methods are attractive tools in non linear learning - when faced with a non-linear problem say regression, kernel methods allow one to keep using linear models with ease. However as one is dealing with more and more massive training sets today (something that is actually desirable from the point of view of getting highly accurate classifiers), one is discovering the price of this ease. The curse of dimensionality definition-ally refers to the non-intuitive distribution of volume in high dimensional hyper-spheres but has come to denote the non-graceful degradation of several algorithms when given high dimensional data (eg nearest neighbor search, quantization). In this talk we shall explore how kernel methods also have to face their very own version of this curse. We shall briefly hint at possible ways of overcoming this but will defer details to a future talk.

Seminar 2 : Evolutionary dynamics of cooperation Speaker : Balaraju Battu Date : 05:00 pm, Saturday, February 5th, 2011 Venue : CS102 Abstract: Evolutionary progress, replicating molecules had to cooperate to form first cell, construction of new features needs cells to cooperate with other cells. Human societies need to cooperate form specialized communities and built organizations. Cooperation is ubiquitous phenomena in biological and social systems. But cooperation is always vulnerable to exploitation. Cooperation demands costs to donor and benefits to recipients. This talk addresses evolutionary dynamics of cooperation infinite populations. About the Speaker : Balaraju Battu has a masters degree in Physics from the Osmania University. He has worked at the Utrecht University, the Netherlands on visual perception for three years. Currently he is working on evolutionary dynamics at the Center for Behavioral and Cognitive Sciences, Allahabad and pursuing a PhD simultaneously.

Seminar 1 : A Few Problems in Kernel Learning Speaker : Purushottam Kar Date : 03:00 pm, Saturday, January 29th, 2011 Venue : CS101 Abstract: Kernel Methods have become ubiquitous in non-generative machine learning over the past several years. More than just a tool, they have now become a learning model in themselves by providing simple yet powerful algorithms for classification, clustering, regression, ranking etc. The aim of this talk is to, both introduce kernel methods as well as take a look at ways of further exploiting the kernel learning model. The talk will start in an introductory mode - quickly giving a basic introduction to learning before moving on to kernel methods and concluding (if time permits) with a discussion on future directions of research in this model.

Back to Talks