Department of Computer Science & Engineering, IIT Madras, Chennai


Error message

Notice: Undefined offset: 1 in counter_get_browser() (line 70 of /opt/lampp/htdocs/aidb/sites/all/modules/counter/

AI - Planning and Constraint Satisfaction

Associated Faculty :Deepak Khemani

AI Planning is a sub-field of AI which focuses on the development of algorithms that can model such goal-oriented behavior by synthesizing a course of action that can achieve the desired objectives. These algorithms serve the needs of intelligent agents that require some kind of independent, autonomous existence capability.

NLP and Text Mining

Associated Faculty :B. Ravindran , Sutanu Chakraborti

Our group has been looking at the marriage of both semantic and statistical approaches to text mining tasks. Much of the work in the group has been driven by specific problems. We have worked on the problem of text summarization [SRR2006a, 2008ARb, 2008ARa] categorization [SRR2008, SRR2006b], clustering [JDR2008, JDRS2007], etc. Some of the current projects are focused on (auto | financial | micro) blog analysis (part of the work is funded by General Motors, India Science Labs), resume processing (funded by Burning Glass Technologies), representative document set mining, question answering systems, etc.

Mining and Indexing of Dynamic Networks

Associated Faculty :Sayan Ranu

Querying and mining graph datasets have been extensively studied and continue being one of the most active research areas. However, an overwhelming majority is centered on analyzing static graph properties. Recent advancements in our scientific and technological capabilities have spurted an unprecedented growth in both the data volume and data complexity, which static graphs fail to characterize adequately. In social networks, for example, new nodes get added every second. Links between nodes change as old acquaintances get forgotten and new friendships are forged. The content at each node (such as a Facebook wall post) change with time and propagate through the network. In road networks, the volume of traffic changes every minute. While some parts of the network can cope with higher traffic, others get bogged down by congestions, which in turn, alters the typical commuting behavior resulting in the congestion further spreading to other parts of the network. In biological graphs, such as gene interaction networks, two genes are connected if they jointly co-regulate a biological function. While the structure is static, the node labels, denoting gene expression levels, vary from human to human. As we move towards personalized medication and treatment, it is critical to understand the interplay between disease, gene expression levels, and the network structure to devise cost-effective treatments.

In essence, techniques based on static graph theory fail to incorporate the dynamic aspects inherent in a wide variety of graph datasets including social networks, road networks, and biological networks. What are the “laws" governing the evolution of these dynamic networks? If we partially observe a trend, can we predict its cascading effect? Can we mine patterns that highlight trends deviating from the expected behavior? Given a training data of gene expression levels and the occurrence of a disease, can we mine the smallest subgraph in the gene interaction network that makes a person susceptible to that disease? These are some of the fundamental questions that we study to better understand the evolution of dynamic networks.

Trajectory Analytics

Associated Faculty :Sayan Ranu

The last decade has witnessed an unprecedented growth in the availability of devices equipped with location-tracking sensors. Examples of such devices include cellphones, in-car navigation systems and weather monitoring gadgets. The widespread usage of these devices has resulted in an abundance of data that are in the form of trajectories. Querying and mining these trajectories is critical to extract the knowledge hidden in the raw datasets and to design intelligent spatio-temporal data management systems.

How do you compute similarity between trajectories when they are noisy and the sampling rate is low? How do you predict congestions in a road network from a stream of trajectory data? Given a budget to update the road network infrastructure and knowledge about historical trajectory data, what is the optimal allocation of the budget?

Reinforcement Learning

Associated Faculty :B. Ravindran

One of the main focus of research in our group is on building situated learning agents that can incrementally solve larger and larger problems using structures built from prior experience. We look at this both from a spatial and temporal perspective, with motivations drawn from cognitive theories of representation.

From a spatial, or a representation, perspective we want agents to be able to incorporate only those aspects of their enivornment that are crucial for the task at hand. We build on the notion of MDP homomorphisms proposed in my thesis and have explored various extensions to this basic framework[RB 2004, RBM 2007, NR 2007, NR 2008].

These architectures work with an already abstract representation of the world. Closer to the sensory level we are looking at the learning of visual routines that can tell an agent what aspects of its visual input to focus on. These visual routines then act as the feature extraction units that the higher levels in the architecture choose from. This work is part of a joint project with Profs. Jeremy Wyatt and Richard Dearden from University of Birmingham and Dr. Anurag Mittal and myself, from this Department, funded by the British Council under the UKIERI program.

From a temporal perspective we have addressed issues of representations and sub-tasks - the question of what is an adequate representation for the sub-task at hand [RB 2003b, RB 2003c, RBM 2007]. We are currently working on using cognitively motivated representation schemes for hierarchical task architectures. On the question of building temporal hierarchical architectures, we are exploring issues regarding incremental learning of the hierarchy, combining learning and planning; and looking at more cognitively motivated representations of tasks and sub-tasks.

We are also exploring the realizations of these algorithms on real-robot platforms. This requires us to address issues that are closer to the sensory and motor control level, including visual cognition, and other localization mechanisms [BRK 2009]. We are also looking at transfer of learning from one robot to another under several constrained settings as well as the problem of learning from instructions.

Data Mining

Associated Faculty :B. Ravindran

My interest in data mining has been motivated partly by my desire to understand the use of statistical models in mining, partly by my drive for doing something of immediate relevance, and partly by the connections between my work on abstraction and some of the problems in mining. One of the ongoing projects is on opthalmic data mining with Sankara Nethralaya^ [RKR2008, CR2008]. We work closely with doctors from their Vision Research Foundation on building better screening tools, disease incidence prediction, risk factor analysis, etc.
We are also looking at telecom data analysis, funded by Ericsson R & D, India. The problems we focus on is trying to understand customer behavior better through various analytics tools, including social network analysis [KR2009, AR2010].

Case Based Reasoning

Associated Faculty :Sutanu Chakraborti

A key area of focus of our research is in case based reasoning (CBR). One of the main problems that rule based expert system approach faced was that of knowledge acquisition. The mechanism of reasoning with expert rules is sound. But while in theory there is no difference between theory and practice, in practice there is. The bottleneck in the use of rule-based systems is in getting the domain experts to articulate their knowledge in the form of rules. Often an expert would be able to take knowledgeable decisions, but be unable to describe the rules behind the decision-making. CBR emerged as an alternative where complete experiences are stored in lieu of the modular abstract rules. CBR is different from conventional rule based systems in its ability to integrate contextual knowledge as captured in actual experiences. Since CBR stores actual experiences, one immediate consequence is its deployment as a tool to build a corporate memory. Every experience encountered is stored as a case, and the case base becomes a repository of all experiences. On encountering a problem needing a solution, the CBR system uses inexact matching techniques to retrieve from memory a set of similar experiences, in the process also producing possible solutions to the present problem. In this way CBR can enhance a database to serve as a knowledge base.

Data based Systems

Associated Faculty :P. Sreenivasa Kumar

Apart from the work described above the AI&DB lab has done work in the following areas: Natural language generation, with the aim of building articulate systems that will be able to give instructions or explanations in human languages. This could be for pedagogical systems, and also for retrieving and conveying information from databases, for example medical histories, and processes for example weather forecasting: Data warehousing and mining, to extract useful information from large amounts of typically operational data: And we are actively looking at XML based semi-structured databases that will be useful in building heterogeneous knowledge bases with wider access.

Indexing Schemes for XML

Associated Faculty :P. Sreenivasa Kumar

The Web has undoubtedly become the prime medium for information delivery, information collection, electronic commerce applications, distance education and a host of other applications. Web data is a mix of structured, semi-structured and unstructured data. eXtensible Mark-up Language (XML) is fast replacing the traditional Hypertext Mark-up Language (HTML) as the data exchange standard. The unique feature of semi-structured data is that the data is self-describing. The data contains the schema or structure information as part of the data. Several interesting issues such as storage of semi-structured data, devising query languages and efficient query processing in such databases that arise in this context are under active investigation in the AI-DB laboratory.

Semantic Web Technologies and Ontologies

Associated Faculty :P. Sreenivasa Kumar

In this area, we are mostly working on modularity of ontologies, ontology based question answering, question driven ontology authoring, ontology based knowledge enrichment and federated SPARQL query processing. Recent advancement of semantic web technologies (SWTs), paved way to facilitate SWTs in many industry and in-house projects.

Recommender Systems

Associated Faculty :Sutanu Chakraborti

Recommender systems play an extremely important role in finding items of interest from a huge number of choices for a particular user. With the proliferation of mobile devices and penetration of internet to masses has led to torrential data flow on the web. Recommender systems filter this massive volume of information to give personalized suggestions.

Few of the current focus areas of our group are modeling human recommenders, intelligently reduce the cognitive load on the users, knowledge based recommenders, modeling utility and offline evaluation of recommendation systems.