Page 86 - AGRANI SAMKALP SILVER JUBILEE SOUVENIR

AGRANI SAMKALP SILVER JUBILEE SOUVENIR

THE CRITICAL ROLE OF PHYSICS IN ARTIFICIAL INTELLIGENCE
Prof. Poonam Tondon
Dept. of Applied Physics

In this decade that is primarily focussing on re-defining 2. Solving the Key Challenges faced
the old paradigms and customs of the mankind, there is a in AI: In Physics we understand
sudden surge to be able to design a machine that can perform and describe in detail the formation
the mundane tasks helping us (the mankind) to reach the of a magnet during a process of iron
scales that are still intriguing. The core fundamentals of bulk cooling. This method has been
science being with us we are now trying to develop the 'so- adopted to solve two challenges in
called' Artificial Intelligence (AI) in the systems. Artificial Intelligence (AI) :
Physics being the science of all that we see around from  A priori estimation of the
the celestial bodies to the everyday occurrence of day and required dataset size to achieve
night phenomenon to the freezing of the upper layer of water a desired test accuracy.
but the lower layers of the oceans hosting a range of  The achievement of reliable decision-making under
biodiversity. Physics is thus quietly playing a critical role in a limited number of examples where each example
the development of Artificial Intelligence (AI) system. can be trained only once.
Following are some of the ways in which physics contributes
to Artificial Intelligence: 3. Physics and Deep learning: Nearly all are aware of the
physical concept of power-law scaling. This law is
1. Physics lays the foundation for Artificial Intelligence: applicable in AI and especially deep learning. The
The knowledge of Physics provides the fundamental correlation between physics and deep learning comes
understanding of how the physical world functions. This into play when physical laws and principles are used to
understanding is necessary for developing the AI inform or constrain the learning process. For example,
algorithms. Some of the commonly used algorithms are: in physics-informed neural networks (PINNs), the neural
(1) Molecular Dynamics Simulations: AI algorithms network is trained not only on data but also to satisfy
are used to simulate the physical movements of atoms physical laws described by differential equations. This
and molecules which is helpful in fields like material can lead to more accurate and generalisable models
science and biochemistry. especially in scenarios where data is scarce but the
(2) Neural Networks with Physical Constraints: Some underying physics is well understood.
neural networks are designed to respect or incorporate 4. Physics and Machine Learning: Most of the AI
the physical laws such as conservation of energy and techniques such as machine learning and optimisation
momentum, which can be particularly useful in algorithms rely on the mathematical models that describe
scientific computing and predictive modelling. the underlying physics of a system. These models can
(3) Swarm Intelligence Algorithms: these algorithms simulate, predict and optimise the physical phenomenon.
are inspired by the collective behaviour of 5. Physics in Data Analysis: By applying the principles
decentralised, self-organised systems such as bird of physics , AI systems can analyse complex data sets,
flocking fish schooling and ant colonies. They are identify patterns and make informed decisions . These
used for solving optimisation problems and systems can simulate , predict and optimise the physical
modelling complex systems. phenomenon.
(4) Poisson Flow Generative Models: The PFGM model Thus, one can summarise this article by making the reader
excels in generating complex patterns such as realistic aware of the deeply intertwined relation of physics with the
images thus helping in creating an advanced pattern development and functioning of AI systems. The fundamental
recognition. The Poisson equations from physics sciences contributes wholistically to the abilities and skills
mimics real world processes thus creating advanced of an Artificially Intelligent device such that it is able to
pattern recognition by applying it to the data. understand and interact with the physical world. 

84