Balaji Vasan Srinivasan

Exploring use of images in clinical articles for decision support in evidence-based medicine

Essential information is often conveyed pictorially (images, illustrations, graphs, charts, etc.) in biomedical publications. A clinicians decision to access the full text when searching for evidence in support of clinical decision is frequently based solely on a short bibliographic reference. We seek to automatically augment these references with images from the article that may assist in finding evidence. In a previous study, the feasibility of automatically classifying images by usefulness (utility) in finding evidence was explored using supervised machine learning and achieved 84.3% accuracy using image captions for modality and 76.6% accuracy combining captions and image data for utility on 743 images from articles over 2 years from a clinical journal. Our results indicated that automatic augmentation of bibliographic references with relevant images was feasible. Other research in this area has determined improved user experience by showing images in addition to the short bibliographic reference. Multi-panel images used in our study had to be manually pre-processed for image analysis, however. Additionally, all image-text on figures was ignored. In this article, we report on developed methods for automatic multi-panel image segmentation using not only image features, but also clues from text analysis applied to figure captions. In initial experiments on 516 figure images we obtained 95.54% accuracy in correctly identifying and segmenting the sub-images. The errors were flagged as disagreements with automatic parsing of figure caption text allowing for supervised segmentation. For localizing text and symbols, on a randomly selected test set of 100 single panel images our methods reported, on the average, precision and recall of 78.42% and 89.38%, respectively, with an accuracy of 72.02%.

A partial least squares framework for speaker recognition

Modern approaches to speaker recognition (verification) operate in a space of “supervectors” created via concatenation of the mean vectors of a Gaussian mixture model (GMM) adapted from a universal background model (UBM). In this space, a number of approaches to model inter-class separability and nuisance attribute variability have been proposed. We develop a method for modeling the variability associated with each class (speaker) by using partial-least-squares - a latent variable modeling technique, which isolates the most informative subspace for each speaker. The method is tested on NIST SRE 2008 data and provides promising results. The method is shown to be noise-robust and to be able to efficiently learn the subspace corresponding to a speaker on training data consisting of multiple utterances.

Efficient subset selection via the kernelized Rényi distance

With improved sensors, the amount of data available in many vision problems has increased dramatically and allows the use of sophisticated learning algorithms to perform inference on the data. However, since these algorithms scale with data size, pruning the data is sometimes necessary. The pruning procedure must be statistically valid and a representative subset of the data must be selected without introducing selection bias. Information theoretic measures have been used for sampling the data, retaining its original information content. We propose an efficient Rényi entropy based subset selection algorithm. The algorithm is first validated and then applied to two sample applications where machine learning and data pruning are used. In the first application, Gaussian process regression is used to learn object pose. Here it is shown that the algorithm combined with the subset selection is significantly more efficient. In the second application, our subset selection approach is used to replace vector quantization in a standard object recognition algorithm, and improvements are shown.

A Symmetric Kernel Partial Least Squares Framework for Speaker Recognition

I-vectors are concise representations of speaker characteristics. Recent progress in i-vectors related research has utilized their ability to capture speaker and channel variability to develop efficient automatic speaker verification (ASV) systems. Inter-speaker relationships in the i-vector space are non-linear. Accomplishing effective speaker verification requires a good modeling of these non-linearities and can be cast as a machine learning problem. Kernel partial least squares (KPLS) can be used for discriminative training in the i-vector space. However, this framework suffers from training data imbalance and asymmetric scoring. We use “one shot similarity scoring” (OSS) to address this. The resulting ASV system (OSS-KPLS) is tested across several conditions of the NIST SRE 2010 extended core data set and compared against state-of-the-art systems: Joint Factor Analysis (JFA), Probabilistic Linear Discriminant Analysis (PLDA), and Cosine Distance Scoring (CDS) classifiers. Improvements are shown.

The UMD-JHU 2011 speaker recognition system

In recent years, there have been significant advances in the field of speaker recognition that has resulted in very robust recognition systems. The primary focus of many recent developments have shifted to the problem of recognizing speakers in adverse conditions, e.g in the presence of noise/reverberation. In this paper, we present the UMD-JHU speaker recognition system applied on the NIST 2010 SRE task. The novel aspects of our systems are: 1) Improved performance on trials involving different vocal effort via the use of linear-scale features; 2) Expected improved recognition performance in the presence of reverberation and noise via the use of frequency domain perceptual linear predictor and cortical features; 3) A new discriminative kernel partial least squares (KPLS) framework that complements state-of-the-art back-end systems JFA and PLDA to aid in better overall recognition; and 4) Acceleration of JFA, PLDA and KPLS back-ends via distributed computing. The individual components of the system and the fused system are compared against a baseline JFA system and results reported by SRI and MIT-LL on SRE2010.

Will your facebook post be engaging

Social media has become the ideal platform for promotional activities of organizations. However, due to the volatility of social media, the wrong message posted at the wrong time can result in significant damage to hard-built brand image. This calls for a mechanism to gauge the reactions a post will evoke from a given social community. The community can vary from customers of a particular brand to brand loyalists interacting through its social pages (for example, on Facebook). In this paper, we focus on learning the communitys reaction from past posts and providing a predictive model for gauging the reaction of the community before the post is published. This helps the marketer take better-informed decisions. Short-text posts in social media leads to a sparse feature space, we propose additional meta-features that improve reaction modeling. Given the feature representation, we discuss the possibility of casting the underlying problem under different paradigms - classification, regression and learning to rank. We study the performances of each of these paradigms on real data from Facebook. We will discuss the challenges involved, and ways to mitigate them, in addition to our observations, results and insights.

Kernelized Rényi distance for speaker recognition

Speaker recognition systems classify a test signal as a speaker or an imposter by evaluating a matching score between input and reference signals. We propose a new information theoretic approach for computation of the matching score using the Rényi entropy. The proposed entropic distance, the Kernelized Rényi distance (KRD), is formulated in a non-parametric way and the resulting measure is efficiently evaluated in a parallelized fashion on a graphical processor. The distance is then adapted as a scoring function and its performance compared with other popular scoring approaches in a speaker identification and speaker verification framework.

Preventing Inadvertent Information Disclosures via Automatic Security Policies

Enterprises constantly share and exchange digital documents with sensitive information both within the organization and with external partners/customers. With the increase in digital data sharing, data breaches have also increased significantly resulting in sensitive information being accessed by unintended recipients. To protect documents against such unauthorized access, the documents are assigned a security policy which is a set of users and information about their access permissions on the document. With the surge in the volume of digital documents, manual assignment of security policies is infeasible and error prone calling for an automatic policy assignment. In this paper, we propose an algorithm that analyzes the sensitive information and historic access permissions to identify content-access correspondence via a novel multi-label classifier formulation. The classifier thus modeled is capable of recommending policies/access permissions for any new document. Comparisons with existing approaches in this space shows superior performance with the proposed framework across several evaluation criteria.

Community reaction: from blogs to Facebook

Online social media is all pervasive in this digitally connected world. It provides a great platform to share information and news, and have public discussions on these topics. These interactions happen on owned-sites as well as on earned social media. But it is reasonable to hypothesize that the communities on the various platforms will engage differently. We explore this hypothesis using data collected from a blog as well as their Facebook page. We observe significant differences between the blog and Facebook as two mediums. First, we see how the topic of the post leads to differences in how posts are received on the two mediums. We next characterize the distribution of the time to comments, displaying how blog and Facebook differ. Finally, we describe how the sentiment of posts and popular entities differ across mediums. In conclusion, reactions across the two mediums are so diverse that it calls for different strategies across different social mediums, we provide our recommendations towards this goal.

Topic-based targeted influence maximization

Social Networks play a very important role as a medium to propagate information among people. Marketers use this to campaign for their products and influence customers. However, it is not practically possible for a marketer to reach out to each and every individual prospective/existing customer due to the sheer size of the networks (in the orders of millions or billions). Therefore, marketers reach out to a small set of people (influencers) who have the potential to further influence/reach out to the targeted customers. Practically, it is not just enough if these influencers have a large following, they also need to have to be able to influence people in the topic that is relevant to the marketer and the influencer must be able to address the target segment that the marketer is targeting. In this paper, we first analyze various edge weighting mechanisms to incorporate influencing probability and utilize this to propose an algorithm to find influencers to maximize the spread to a specified set of targets.