I'm a cognitive scientist currently working on modeling human attention, but also broadly interested in machine learning applications and statistical analysis of large data sets. Attention is a difficult process which required optimization of the information delivery to complete multiple objectives. I previously have degrees in computer science from Indiana University and MS Ramaiah Institute of Technology.
Automatic parallel processing occurs in visual search when a target differs from other display objects on a salient visual dimension, as when a target is red and other objects are all green, and the phenomenon is termed popout. When a second display object is also perceptually salient but task irrelevant (here termed a foil), popout is hindered. The present two studies assess the amount of harm (or benefit) to popout when there are two foils, on different dimensions. In particular it answers the question: Do the interference effects cancel or add?
Markov random fields have the unique property of using neighbouring information to compute labels for the current node. In the problem of binary image segmentation; the image is segmented into object of interest and background. Various properties determine which label a specific region takes: among them are color, size and texture properties. This work explores the various characteristics of images in generating figure ground segmentations before running it through a classifier.
There are various social networks catering to different needs. Are there identifiable patterns across social networks? Is the behaviour on one social network say twitter comparable to another network? This works tries to answer this question by looking at patterns of social behaviour across twitter and flickr. The user time feeds are identified and binned to specific time windows, then the binned time series are compared across networks to find strong markers of identification. Results show based only the meta content it is very difficult to disambiguate users across social networks.
Mention classification for co-reference resolution
This work was done in collaboration with the computational linguistic department at IU to explore a naive bayes classifier for the CoNNL shared task. The algorithm was simple and used syntatic and semantic cues from the annotated text to disambiguate co-refferent text.
Knowledge acquisition in computer systems follows a very unstructured approach, in some circles we say that knowledge evolves. This project used the opencyc architecutre and imposed constraints on the knowledge framework to tie in new sources of information to existing soruces of information. A simple dialog was developed to facilitate the definition and realtionship of new terms.
In GPS based autonomous vehicles, obstacle avoidance is a problem (so is obstacle detection, but lets skip that for now). In this work we explore obstacle avoidance along pre-planned routes. On detecting an obstacle the agent makes changes to the route to facilitate smooth movement of the vehicle. This mechanism can also be used in a multiagent system since it invovles minimal changes and proposes an easy to implement behaviour.
Most clustering algorithms assume a consistent state of membeship, this works for most data but in modern cyberworld memberhips change very quickly and the systems need to adapt to this change. The current algorithm proposes a modified ant-clustering algorithm which uses pheromone based probabilities to recompute memberships. The idea behind this technique is the ability for a system to recover from a sybilNet and thus change its membership.
Zero day attacks are hard to detect,a possible solution is to use anomaly-based intrusion detection. In this work we explore a genetic algroithm approach to facilitate rapid training of the classifier. The classifier uses byte frequencies generated from packet data and learns to classify anomalous packets.
Feel free to email me to discuss about research, provide feedback, suggest new repos or to just say hello!