SIGML resumes this semester with student presentations. There will be two students presentations on last semester UGP work and course project. CSE Y13 and MTech'15 students are encourage to attend the same. You are all cordially invited for the same. Below are the details of the talks. --------------------------------------------------------------------------------------------- *TALK I :* Title : Sparse Recovery And Optimization (UGP) Speaker : Amartya And Bhuvesh Timing : 4:00 PM - 4:30 PM , 9 January 16 Day : Saturday Abstract* : * Problem of parameter estimation with given set of constraints is central to machine learning. In several learning applications, one is required to perform estimation with far fewer data points than the number of parameters to be estimated. Though consistent estimation is impossible in these settings in general, structural constraints like sparsity or low rank make it possible to perform consistent estimation. Such constraints naturally arise in several situations where we desire "simple" solutions that can be expressed as a combination of a few "atoms" or dictionary elements. Ex. include gene expression analysis, collaborative filtering, and compressive sensing. In these cases, the optimization problem reduces to an $l_0$ minimization problem, which is a non-convex optimization problem and NP-Hard in general. There have been numerous methods proposed to "solve" this problem and all these methods provide guarantees under certain conditions. Such methods include convex relaxation to an $l_1$ minimization problem, iterative hard thresholding, and greedy approaches. In this project, we look at various sparse recovery algorithms :- GradeS, OMP, OMPR and COSAMP. We look at the problem of how accurately these algorithms can recover a high dimensional signal from a small set of measurements and provide performance guarantees on them. These methods reflect different classes of algorithms with varying requisite conditions, cost, guarantees, and different operating principles in general. We study the various mathematical properties associated with the problem and the performance of these methods with respect to various parameters like error, sparsity and number of dimensions. ---------------------------------------------- *TALK II:* Title : Diachronic Word Sense Change Identification (NLP course Project) Speaker : Raghuveer Timing : 4:30 PM - 5:00 PM , 9 January 16 Day : Saturday Abstract * : *Language has continuously changed over time. New senses for some words took birth while some old senses demised. For an example the word ‘gay’ which referred being full of joy about 2 centuries ago but is now used to refer to a homosexual; ‘economy’ referred to management of resources in the past and is now being referred to state of country in terms of productions. In this work we find words which changed their senses over time by using Word vector models trained from different time epoches and plotting their inter cosine similarity to find the words of interest. -------------------------------------------------------------------------------------------------