Return to Home Page

I was introduced to the subject Pattern Recognition in 1975 when I joined Indian Statistical Institute (ISI), Calcutta, to do my PhD with Prof. D. Dutta Majumder as a CSIR Senior Research Fellow. At that time there was no text book; only a few edited volumes, mostly by Prof.  K.S. Fu, were available in our library or in the market. Another edited book that I used to consult often was “Fuzzy sets and their applications to cognitive and decision processes“ by Zadeh, Fu, Tanaka and Shimura, published by Academic Press in 1975. Apart from them I studied, though it was extremely difficult to understand, the DSc thesis of G.S. Sebestyen, entitled “Decision-making process in pattern recognition” published as a monograph by Mc Milan Press in 1962. I worked in the area of speech recognition and man-machine communication using fuzzy set theory. Prof. K.S. Fu was the foreign examiner of my PhD thesis. After that I moved to Imperial College, London, in 1979 as a Commonwealth Scholar and did another PhD in image processing using fuzzy sets.  Though there had been a possibility to work at Purdue University, USA, as a Post-doc Fellow with Prof Fu, I chose to go to Imperial College. At that time labs with complete software and hardware facilities for working in (gray) image processing were not readily available at many universities/Institutes, even in the developed nations. In 1981, I attended, as a speaker, the workshop on pattern recognition organized at Oxford University, UK, organized by NATO Advanced Studies Institute. I was thrilled as well as proud to be an invited speaker amongst several PR stalwarts including KS FU, K Fugunaga, L Kanal, P Devijver, MG Thomason and J. Kittler.

I returned from England to India in May, 1983. Then during 1986-1987, I visited the University of California, Berkeley, and the University of Maryland, College Park, as a Fulbright Fellow, and had an opportunity to work with two giants, namely, Prof. Lotfi A. Zadeh, father of Fuzzy Sets, and Prof. Azriel Rosenfeld, father of Image Processing. I also spent more than two years at the NASA Johnson Space Center, Houston, during 1990-92, and in 1994 as NRC Associate to work in the Software Technology Branch, Information Technology Division, where I came across and got interested in subjects like genetic algorithms (GA) and rough sets.

Subsequently, my research interests moved towards developing primarily the soft computing approaches comprising fuzzy sets, neural networks, genetic algorithms and rough sets, both individually and in integration. Tasks like classification, rule generation, learning with mislabeled samples, uncertainty analysis, case based reasoning, image segmentation, data mining and knowledge discovery, designing connectionist knowledge based systems, page ranking, and proving convergence were primarily considered. Methodologies with theoretical analysis, and different tools were developed. The theory of probability was also used in the process, wherever required. Real life application domains that I dealt with include speech recognition, remote sensed image analysis, medical imaging, bio-informatics, web intelligence, and social network mining.

Salient features of some of the contributions are as follows:

In the area of image analysis, we have given various definitions of image entropy based on exponential gain function, rough-fuzzy generalized entropy, and other quantitative indices for image processing tasks. The exponential gain function relies on the fact that a better measure of ignorance is 1 – p_i rather than 1/p_i (as used by Shannon), when p_i is the probability in receiving the ith event.  Rough-fuzzy entropy, on the other hand, takes care of the fuzziness involved in the boundaries of both sets and granules in the definition of rough sets, in general. In image processing it reflects the ambiguity resulting from resemblance in nearby gray levels and pixels as well as the ambiguity arising from fuzzy boundaries of regions; thereby providing a stronger paradigm for uncertainty handling. Although we have demonstrated their extensive applications mainly for image segmentation problems, one can use them for other image processing tasks, and for any other data involving uncertainty arising from fuzziness, granularity and randomness.

In neural computation, we have pioneered the synergistic integration of fuzzy sets with artificial neural networks (ANNs), introducing the notion of fuzziness at various levels. We have developed a family of generic models encompassing fuzzy multilayer perceptron (FMLP), fuzzy logical MLP, knowledge-based FMLP, rough-fuzzy MLP, fuzzy decision tree based knowledge-based MLP, and modular evolutionary rough-fuzzy MLP — particularly for classification/clustering and rule mining, under uncertainty. The proposed fusion of fuzzy sets with ANNs has the effect of enhancing the capability of latter for handling classes with overlapping or complex boundaries and accepting non-numeric inputs, and hence their performance and application domains.

In granular computing, my research has shown that rough information granules derived from the data can improve the learning efficacy, dimensionality reduction and classification performance, and reduce significantly the computation time for various tasks, e.g., inferring the architecture of networks, estimating initial value for k-means algorithms, generating cases or prototypes, and retrieving. The symbiotic integration of this concept with other heterogeneous paradigms such as fuzzy-neural networks and genetic algorithms with variable mutation probability results in a highly effective tool for mining structures, patterns and more meaningful rules, and dealing with knowledge discovery aspect in very large and heterogeneous data sets.

Among my other contributions related to data mining, I would like to mention the fast, efficient and powerful data condensation and dimension reduction methods that have been very useful for developing speedy, accurate, and scalable data mining tools.

In web intelligence research, we have defined a new metric that generalizes Kendall's Tau distance for comparing two page rank vectors, or more generally, any pair of score vectors. The metric is found, theoretically and experimentally, to be powerful in determining the stopping time of the power method (thereby saving the computation time massively) and measuring how well ranks represent scores. The probabilistic surfer model, developed maintaining the continuity of topics, can simultaneously rank and categorize web pages, while the fuzzy surfer results in a robust version of PageRank, called FuzzRank.

The aforesaid research in pattern recognition and machine intelligence has led to the emergence of several modern disciplines involving fuzzy sets with other computational paradigms, as evident from the literature. This has also led to the incorporation of rough sets as a component of soft computing; Zadeh’s original definition had components like fuzzy logic, neurocomputing, GA, and probabilistic reasoning. This augmentation by rough sets has enhanced the computational intelligence property of soft computing and triggered its multifarious applications. 

My current research interest is primarily pivoted on social network mining, video tracking, and computational theory of perceptions involving granular computing, among others.

For furtherance of research in the said topics in a consolidated manner, our Institute created a new unit in March 1993 called Machine Intelligence Unit under my headship. As recognition of my research contributions, the Department of Science and Technology, Govt. of India established the nation’s first research center in Soft Computing in October 2004 under my leadership at ISI, Calcutta. The international conference PReMI (Pattern Recognition and Machine Intelligence) is my brain child. The conference is unique in the sense that it provides a platform to exchange ideas regularly between these two communities for mutual benefit. The first edition was in Calcutta in 2005, and the fourth one in Moscow in 2011; and all of them were co-sponsored by IAPR. It will remain incomplete if I do not acknowledge my PhD students and collaborators. Most of my work was done with PhD students and young colleagues/collaborators. I have learnt a lot from my students who keep me always young.  My foreign collaborators where I pay (or used to pay) several visits are in Warsaw University, Poland; Hong Kong Poly University, Hong Kong; and the University of Naples Parthenope, Italy. Through the Warsaw collaboration with Prof. Andrzej Skowron, I met Prof. Z. Pawlak (father of rough sets) in the mid 90s, who unfortunately passed away in 2008.

“Pattern recognition” has not only led me to win several prestigious awards and honors in India and abroad and attain the highest administrative and academic position (namely Director, of my Institute), but also gave me opportunities to visit more than thirty countries for academic purposes and, more importantly, to know the common people and culture there and make new friends. I enjoy my research work very much and love publishing in peer reviewed archival journals. I look into every problem mostly from a PR perspective. I would like to remain active as long as I can, and continue to do what I love to.

FEATURE

Getting to know…Sankar K. Pal, IAPR Fellow

By Sankar K. Pal, IAPR Fellow (India)

Professor Sankar K. Pal, IAPR Fellow

ICPR 2002, Quebec City, Quebec, Canada

For contributions to pattern recognition and image processing

using soft computing.