The 2023 Kanpur Lectures Series on Engineering and science in our world
by Prof. Sandip Tiwari
Monograph available as a pdf file
Prof. Sandip Tiwari is a Distinguished Visiting Professor in the Electrical Engineering Department at IIT Kanpur. He is Charles N. Mellowes Emeritus Professor of Engineering at Cornell University and a distinguished alumnus of our institute. He was hosted by the Samtel Centre for Display Technologies in the Jan-May semester of 2023 when he gave a series of lectures at the institute called The Kanpur Lecture Series on “Engineering and science in our world” . It consists of five lectures:
# The fourth lecture on 13th March, 2023
was held as a “K. R. Sarma
Distinguished Lecture”.
Those interested in discussing with Prof. Sandip Tiwari can reach him by email at st222@cornell.edu.
Sandip Tiwari
10/17/2022
IIT Kanpur opened my
mind’s I watching the questioning, probing and independence practiced by the
community of teachers and students. This awakening reinforced that there is a
frontier out there waiting to be
explored and to do good as one saw fit in one’s view. This has been a life lesson. Its lessons certainly apply to the technical working that
makes a living possible, but also teaches one engagement with the broader world
one dwells in Semiconductors-, computation-, and more broadly information-
centered pursuit is now pervasive. It is the engine that drives the world. This
sequence of talks explores the serious questions and challenges of the current
state—technical and worldly—with guesses for the future. The first three talks
discuss, using a perspective of complexity in electronics as a mix of
determinism and uncertainty, the questions arising in the giant integration
scales now possible with nanoscale devices where interdisciplinary interests of
computation for complex and incomplete problems is now pervasive. Next, I
follow through with a personal view of the cultural and humanist lessons of
such circumscribed science and engineering pursuits in a real world with its
open boundaries. It is a quest of learning and of progress in a dynamic world.
I will end the development of these thoughts by probing the future in which
semiconductor manufacturing is the essential physical layer for nearly every
economic and national pursuit.
The ability to control semiconductor structures at
nanometers scale and integrate in multiple dimensions has made an integration
of near-trillion scale possible using structures that are largely surfaces and
quantum-mechanical-sized material. This is a non- random statistical assembly
of near-classical objects. Information manipulation in this assembly must occur
under constraints of energy and variability that has static and dynamic
manifestation from the assembled object particles. Deterministic computing,
which is largely the present paradigm, leads to a variety of con- sequences and
constraints that set limits. Most of the modern themes—machine learning and
neural networks in practice of artificial intelligence—are still subject to
these since the implementations employ deterministic computation based on BLAS even if dealing with probabilities.
I will illustrate examples of the limitations that are far away from
thermodynamic information capacity efficiency arising
in such approaches, some common mis-understandings, and set the context
for what and which kind of problems under what constraints become
amenable to exploration of alternative
information-processing techniques.
The world is probabilistic,
whether classical arising in the incompleteness of the classical unknowns or of
the natural randomness as in quantum-mechanical fluctuations or their
spontaneous classical manifestations.
The Turing machine is a computational device that explores the extent and the
limit of what can be computed. A simple view would be that it sets limits for
implementation of deterministic logic implementation in a computing engine.
Boolean, von Neumann, for example.
Probabilities, which have within them the objective versus subjective
conundrum, not unlike the natural world we inhabit, provide a non-Turing means
to computation as one learns. The Bayesian reconstitution of the probability
with new information is the subjective tool for this learning. This makes
stochastic and probabilistic learning circuits, compact and specific, possible
that can operate rapidly and at low power in real time on real-world problems.
This talk discusses the underpinning of the computational approach and develops
and gives examples of implementation in circuits, where the probabilities are
derived using the low-power randomness from superparamagnetism.
The tension between data-guided agnostic computation and physical
principles guiding information and the evolution of the system is unresolved.
It is perhaps a mirage, a conjecture would be that they should lead to similar
guidance for a dynamic system since information in observation and the physical
laws must represent the duality of nature. Science-guided
artificial intelligence (AI) and machine learning (ML) approaches give a powerful new tool for tackling hard and
soft causality. Cross entropy, Lagrangians, and Hamiltonians as minimization approaches using Bayesian principles are
equivalent, but with different constraints due to the underlying mathematical
principles and descriptions deployed. We explore this range for real-world open
boundary problems to analyze how physics-guided AI/ML can be useful in complex
problems such as those encountered in the broader set. For a broader problem
that a human is good at, an example is of classifying or generating a specific
classical com- poser. For a narrower
problem, an example
is of integrated design with
its constraints of layout, cross talk, speed, and power. The corollary of this
view is that one can extract the physical mechanisms from the data. Such an
extraction shows the power of this information-based approach in a physical
data.
What exactly constitutes
science and what is engineering (or what is technology) is nebulous at best.
Progress in science depends on development of new tools. Experimental tools
arise from progress in technology and engineering. Theoretical tools too are
creations where one connects physical and mental worlds. The natural world is
chaotic. It is an open, dynamic and nonlinear system. Randomness, causation,
interactions, thermodynamics, etc., all matter. So, simplistic views—Pasteur’s
quadrants, Wallace-Darwin’s adaptation, Snow’s two cultures, Kuhn’s pardigm, Ockham’s razor in making simplest of choices with
least axioms, and many others—are insufficient. The conduct of science and
engineering has continuously changed since the dawn of modern science, it changes the world and the world changes
it, changes are fast and slow and non linear, local
context matters as can be seen best through commerce. How institutions practice
and succeed and evolve matters for the future trajectory. Today, most problems
need a simultaneous in-depth understanding of multiple disciplines even within
the sciences. I discuss from personal and the broader world’s experiences the
resulting conflicts: cultural such as what Snow brought up, how science and
engineering has evolved from the heydays of Bell Labs or IBM Research, and in
what shows up in the conduct of science and engineering in the world it
inhabits in the modern society, particularly in USA and Europe where I have
spent enough time experiencing the daily living. The problem is of
dimensionality reduction in complexity. In this complex world, the only rule
one can draw is the Mencken’s rule that for every complex problem, there is an
answer that is clear, simple and wrong. I will speculate based on this argument
the interesting problems for our community that the intertwined science and
engineering can fruitfully and gracefully approach.
That semiconductors have
through devices, circuits, systems, computing, communications and information
exchange made the mod- ern world possible is a sound and arguable claim. But,
we came to this point dynamically.
New inventions, new technologies, new ways of attacking the information
processing and transfer and its evolution to knowledge have all had stepping evolution. Wisdom, which follows,
is very much a particular society’s optimization—
like a minimization constraint—upon which it may act (or not). The earliest
computing companies, Burroughs to Univac do not exist, nor do those who brought
about the minicomputers such as
Digital, or microcomputers such as Sun; yet computing is the heart beat of the
society. Semiconductor manufacturing is very capital intensive, and it demands
experience and precision knowledge. Even for USA, the answer was focus on design and let TSMC build it. But, societal tensions or wars can intervene as one
sees right now. This is a very serious issue for any nation. Semiconductors are
like agriculture. One cannot be confidently independent without the ability to
build and deploy. In this broader worldly context, I would like to discuss
commercial principles that have guided the evolution of the information
enterprise, and look at the open big areas of the future, to speak to what
needs to be the broader focus of design, development, manufacturing, and associated computational developments for
the coming generations.