Wednesday, April 30, 2025

R Narasimhan Centenary

 


Prof RN (April 17, 1926 – September 3, 2007) is well-known as the designer and implementor of India’s first digital computer, the Tata Institute of Fundamental Research Automatic Calculator (TIFRAC). He was also the founder of the Computer Society of India. Wikipedia has a comprehensive article about RN that is easily accessible.

However, another part of his life has been known only to professionals in the field. This part deals with language behavior, visual pattern recognition, Vision in animals and machines, and Artificial Intelligence. I will discuss a few significant questions in this research area and RNs' positions. It is incredible how his views have stood the test of time.

The role of syntax  

RN pioneered exploring syntactic methods in visual pattern recognition, which was appropriate to the context in pattern recognition with which he was first involved—bubble chamber images [RN, 1964] and [RN, 1966]. However, His interest in vision continued for years [Reddy and RN, 1971], [Barlow et al, 1972]. He was not tied down to syntactic methods.  

Soon after this, he became more interested in natural languages.

Theories of Language

Noam Chomsky’s work on transformational grammars was very influential from 1965 to 1985. Chomsky believed language is primarily a formal system governed by innate structures rather than a repository of encyclopaedic knowledge.

However, RN did not fall for the temptation to follow the stream. He recognized the importance of going beyond syntax and dealing with context, multi-sensory learning, and the pragmatics of language use. He gave special attention to language learning by children.

RN stressed phylogenetic continuity. He found it difficult to believe that language behavior suddenly appeared in homo sapiens. It is more likely to have evolved gradually from animal communication.

RN published two books collecting his work together [RN, 1998] and [RN, 2004].  

Limitations of automated logic

Symbolic AI focuses on representing knowledge in structures that can be manipulated by a machine to carry out deduction and induction. This works with toy worlds but fails to deal with the real world. Can automated deduction go beyond stored knowledge and explain common sense? RN said no. He accepted the importance of deduction and induction, but felt that common sense goes beyond these.

RN believed in paradigmatic behavior, which is driven by analogy.

Geoffrey Hinton, often called the “Godfather of AI,” has recently offered a provocative perspective on human cognition. In an interview [Hinton, 2025], he has declared that humans are fundamentally “analogy machines,” not the reasoning beings we often perceive ourselves to be.

Role of language in problem solving

As a Research Associate at the Computer Group at TIFR, I was fortunate to learn from RN’s work and publications. His special interest in language as a tool led me to explore “Language-Based Problem Solving” for my thesis, guided by Prof RN and Prof JR Isaac [Ramani, 1971]. I was working with a computer (CDC-3600) with much less memory than the cell phone I now use! The idea of language models was not well developed then, though we were writing programs that spun out stories, correct in content and grammaticality, based on ad hoc probabilities. Multi-layer neural networks and the transformer model were decades away. However, I could implement pattern matching to solve problems by analogizing from a few examples.

Learning from the Nervous System

RN was inspired by the work of Hubel and Wiesel at Harvard.

Their discovery of feature detectors in the visual cortex of cats supported his conviction that understanding perception and cognition required a multidisciplinary approach.

His paper [Barlow et al, 1972] is memorable, emphasizing his conviction that studying the human nervous system and human behavior is essential for progress in AI.

RN’s unique style:

RN never had set goals for publication for himself or those he guided. This eliminated the need to follow research fashions of the day, or to go journal hunting to find some journal that would accept whatever one had written. Integrity in research requires that you have truly significant results before you publish. These principles have made RN’s positions on research issues relevant over the decades. The coming decade is expected to be an AI decade, with highly accelerated progress in science, technology, and medicine. I wish that RN were here to enjoy the progress!

References

[RN, 1964]
Labeling schemata and syntactic descriptions of pictures, R. Narasimhan, Information and Control, Volume 7, Issue 2, 1964, Pages 151-179, ISSN 0019-9958,
https://doi.org/10.1016/S0019-9958(64)90087-7.
(https://www.sciencedirect.com/science/article/pii/S0019995864900877)

[RN, 1966]
Syntax-directed interpretation of classes of pictures. Communications of the ACM9(3), pp.166-173. 1966.

[Reddy and RN, 1972]
Reddy, V.S.N. and Narasimhan, R., 1972. Some experiments in scene analysis and scene regeneration using COMPAX. Computer Graphics and Image Processing1(4), pp.386-393.

[RN, 1981]
Narasimhan, R., 1981. A Framework for Modelling Behavior. In Modelling Language Behavior, pp.1-26.

[RN, 1998]
Narasimhan, Rangaswamy. Language Behavior: Acquisition and Evolutionary History. India: SAGE Publications, 1998.

[RN, 2004]
Narasimhan, Rangaswamy. AI And the Study of Agentive Behavior. India: Tata McGraw-Hill Publishing Company Limited, 2004.

[RN, 1997]
Narasimhan, R., 1997. Knowledge processing and commonsense. Knowledge-Based Systems10(3), pp.147-151.

[Ramani,1971]
Ramani, S., 1971, September. A Language-Based Problem-Solver. In IJCAI (pp. 463-473).

[Barlow et al, 1972] Barlow, H.B., Narasimhan, R. and Rosenfeld, A., 1972. Visual Pattern Analysis in Machines and Animals: The same principles may underlie the operation of sensory neurons, computer programs, and perception. Science177(4049), pp.567-575.

[Hinton, 2025]
Geoffrey Hinton’s interview on April 23, 2025 at the University of Toronto
https://www.youtube.com/watch?v=vpUXI9wmKLc