Cover page

Professor Naresh Dadhich was the second director of the Inter University Centre for Astronomy and Astrophysics

(IUCAA, Pune). A renowned cosmologist who could see the geometric beauty of the General Theory of

Relativity, in the way that very few others than Albert Einstein might have comprehended it. His demise on

November 6th 2025, was a signiﬁcant loss for humanity and the world of science and mathematics. We pay

homage to his memories.

Managing Editor Chief Editor Editorial Board Correspondence

Ninan Sajeeth Philip Abraham Mulamoottil K Babu Joseph The Chief Editor

Ajit K Kembhavi airis4D

Geetha Paul Thelliyoor - 689544

Arun Kumar Aniyan India

Sindhu G

Journal Publisher Details

Publisher : airis4D, Thelliyoor 689544, India

Website : www.airis4d.com

Email : nsp@airis4d.com

Phone : +919497552476

Editorial

by Fr Dr Abraham Mulamoottil

airis4D, Vol.3, No.12, 2025

www.airis4d.com

This issue of AIRIS4D pays tribute to Professor

Naresh Dadhich (1944–2025), a renowned Indian

relativist and one of the founding pillars of the Inter-

University Centre for Astronomy and Astrophysics

(IUCAA), Pune. Professor Ajit Kembhavi remembers

Professor Naresh Dadhich’s signiﬁcant contributions

to gravity theory, black hole physics, Lovelock gravity,

and brane cosmology. As a teacher and mentor, he

inspired generations of students and collaborators, many

of whom became leading scientists. He continued

publishing until his ﬁnal days. Dadhich played a

central role in building IUCAA from its inception,

promoting astronomy across India through extensive

outreach and forming strong national and international

collaborations—including with countries in Central

Asia, Africa, Iran, and Turkey. His eﬀorts helped

integrate Indian institutions into major global projects

such as SALT, the Thirty Meter Telescope, and the

LIGO-India initiative. Known for his simplicity,

rationalism, socialist ideals, and joyful personality,

Dadhich maintained deep friendships across scientiﬁc

and artistic communities. He is fondly remembered

for his clarity in explaining complex concepts and his

commitment to nurturing young talent. The issue also

features a personal reﬂection by astronomer Joe Philip

Ninan, who recounts childhood memories of Dadhich

as an inspiring teacher who sparked his early interest in

physics.

Arun Aniyan addresses “Few-Shot Learning

(FSL)”, one of the biggest limitations of traditional

deep learning—its dependence on huge labelled

datasets—by enabling models to learn new concepts

from only a few examples. Standard deep learning

struggles in real-world scenarios where data is scarce,

costly, sensitive, or rare—such as medical imaging,

scientiﬁc discovery, robotics, and personalised AI.

FSL overcomes these challenges through meta-learning,

metric-learning approaches like prototypical networks,

adaptable methods like MAML, and data-generating

techniques that expand limited datasets. By learning

how to learn, FSL allows AI systems to generalise

quickly from minimal data, making it essential for high-

impact, data-constrained domains and a signiﬁcant step

toward more human-like, eﬃcient, and generalizable

artiﬁcial intelligence.

‘The Time Arrow of Entropy” by Jinsu Ann

Mathew explains how entropy—often described as

disorder—gives direction to time by making many

processes naturally irreversible. From broken glass to

heat ﬂow, systems tend to move from order to disorder,

creating the “arrow of time.” It shows how memory in

natural and artiﬁcial systems preserves traces of the past,

slowing change and shaping future behaviour. Language

and society evolve in a similar forward-moving way,

carrying history that prevents them from returning to

earlier states. By observing how entropy ﬂuctuates,

trends, or suddenly jumps, we can understand the

rhythms and transitions in physical, biological, and

social systems. Ultimately, entropy becomes not just

a measure of uncertainty but a lens for understanding

how systems evolve.

The article “Unboxing a Transformer using

Python - Part I” by Linn Abraham introduces

how Vision Transformers (ViTs) adapt the

transformer architecture—originally built for

language translation—to image recognition tasks.

Unlike CNNs, ViTs excel at capturing long-range

dependencies in large datasets but must relearn local

pixel relationships due to fewer built-in spatial priors.

Using a PyTorch implementation, the author walks

through the structure of a ViT model for classifying

solar active regions, highlighting key components such

as patch embedding, positional embeddings, class

tokens, transformer encoder blocks, and the MLP head.

A detailed look inside the patch embedding layer shows

how images are divided into 16×16 patches, ﬂattened,

normalized, and projected into a lower-dimensional

embedding space to prepare them for transformer

processing. This ﬁrst part focuses on understanding

the model architecture, setting the stage for deeper

exploration of positional embeddings and transformer

layers in future instalments.

Dusty Plasma by Abishek explains how dusty

plasmas—ionized gases containing charged dust

grains—form a uniquely complex, multi-scale system

where grain charging, electromagnetic forces, and

Debye screening shape particle behaviour. Dust grains

acquire charge through electron/ion collection and

photoemission, leading to forces such as electrostatic lift,

ion drag, neutral drag, thermophoresis, and gravity that

determine whether grains levitate, drift, or form ordered

structures like plasma crystals. Strong coupling between

grains produces rich collective behaviour, including

phase transitions, dust acoustic waves, instabilities,

and self-organized patterns. Because grain charge

ﬂuctuates with plasma conditions, microscopic charging

dynamics strongly inﬂuence macroscopic transport and

wave phenomena. Dusty plasmas play major roles in

astrophysical environments—such as protoplanetary

disks, cometary comae, and planetary rings—as

well as in technology, where dust can both hinder

semiconductor and fusion operations and serve as a

valuable tool for diagnostics, nanoparticle synthesis,

and the study of many-body physics.

Geetha Paul provides a comprehensive overview

of diabetic retinopathy (DR), a major diabetes-related

eye disease that damages retinal blood vessels and

can lead to blindness. It explains how chronic high

blood sugar causes microvascular injury, resulting in

microaneurysms, haemorrhages, exudates, macular

oedema, and in advanced stages, the growth of fragile

new blood vessels (proliferative DR). Key associated

conditions such as drusen, choroidal neovascularisation

(CNV), and cystoid macular edema (CME) are also

described. The article highlights symptoms, risk

factors, diagnostic methods, and treatments, including

anti-VEGF injections, laser therapy, steroids, and

vitrectomy, emphasising that early detection and strict

diabetic control can prevent the most severe vision

loss. It also outlines how retinal image processing,

machine learning, and deep learning—using steps like

preprocessing, segmentation, feature extraction, and

CNN-based automated grading—play a crucial role in

detecting, classifying, and monitoring DR with high

accuracy.

iii

News Desk - Meomories Never Fade

Professor Naresh Dadhich had spend three days as guest to airis4D in November 2024 with his wife Sadhana and

Professor Kembhavi and his wife Asha. The memories of the days spend together will never fade away.

Contents

Editorial ii

1 Professor Naresh Dadhich

September 1, 1944 – November 6, 2025 1

2 My Childhood Memories of Naresh Dadhich 4

I Artiﬁcial Intelligence and Machine Learning 5

1 Few Shot Learning 6

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2 Practical Challenges of Standard Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.3 Few-Shot Learning: A Paradigm Shift in Machine Learning . . . . . . . . . . . . . . . . . . . . . 7

1.4 Core Concepts in Few-Shot Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.5 Key Approaches to Few-Shot Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.6 The Few-Shot Learning Process Workﬂow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 The Time Arrow of Entropy 11

2.1 Irreversibility: Why Time Has a Direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2 Memory in Natural and Artiﬁcial Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.3 History in Language and Society . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.4 Entropy Over Time: Patterns, Trends, and Transitions . . . . . . . . . . . . . . . . . . . . . . . . 12

2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Unboxing a Transformer using Python - Part II 14

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.2 Position & Class Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.3 The Transformer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3.4 The Final layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

II Astronomy and Astrophysics 17

1 Dusty Plasma 18

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.2 Formation of Dusty Plasma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

1.3 Forces acting on dust particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1.4 Collective behaviour of dusty plasma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1.5 Dust Acoustic Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

1.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

CONTENTS

III Biosciences 23

1 Diabetic Retinopathy- Medical Image Processing 24

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.2 Drusen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.3 CNV (Choroidal Neovascularization): . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.4 CME (Cystoid Macular Edema): . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.5 Exudates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.6 Microaneurysms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.7 Haemorrhages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1.8 Causes and Progression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.9 Symptoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.10 Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.11 Treatment and Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.12 Role of Image Processing in Diabetic Retinopathy . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.13 Automated Detection and Classiﬁcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.14 Steps in Diabetic Retinopathy Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

1.15 DR grading pipeline end-to-end . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

vii

Professor Naresh Dadhich

September 1, 1944 – November 6, 2025

by Ajit Kembhavi

airis4D, Vol.3, No.12, 2025

www.airis4d.com

Professor Naresh Dadhich passed away, at the age

of 81, on November 6, 2025 in Beijing while on an

academic visit to China. His sudden demise is a tragic

loss to the scientiﬁc community, and to the very large

number of people from diverse ﬁelds and professions,

in India and in many other countries, who were his

friends, colleagues and collaborators. He worked

mainly in general relativity and gravitation theory, and

was one of the founders of the Inter-University Centre

for Astronomy and Astrophysics (IUCAA), where he

was Director during 2003-2009. He continued to be

associated with the institute as an active scientist to the

very end.

Dadhich was born and brought up in a village near

Churu in Rajasthan, where his father was a priest.

He left home for school at a young age, got his

ﬁrst degree in mathematics from BITS-Pilani and his

M.Sc. from Vallabh Vidyanagar. He then reached the

University of Poona (now SP Pune University, SPPU)

in 1966, for research in mathematics. There he had

the good fortune of becoming a Ph.D. student of the

renowned mathematician and relativist Professor V. V.

Narlikar, the father of Professor Jayant Narlikar. Naresh

specialised in the very diﬃcult area of general relativity,

in which he became a great expert.

Soon after his Ph. D. in the early seventies,

Naresh was appointed as a lecturer in the Department

of Mathematics in SPPU, where his scientiﬁc journey

truly began. He worked on classical and quantum

aspects of gravity, with Lovelock Gravity being one of

his favourites, brane world cosmologies, gravitational

collapse, wormholes and the astrophysics of black holes.

Over some years he gathered around him a number of

highly talented young researchers, including Sanjeev

Dhurandhar, B. S. Sathyaprakash and Patrick Dasgupta,

who had obtained their Ph.D. from leading institutes

in the country. There were also research students

including Sanjay Wagh, Ravi Kulkarni, Sucheta Koshti

and Varsha Daftardar. Some of these young people

became internationally leading researchers in their areas.

The work done during this period included the very

interesting magnetic Penrose process for the extraction

of energy from black holes. Naresh continued to publish

until his last days, and had a paper accepted just a few

days before he passed away. His essays for the Gravity

Research Foundation annual competition got Honorable

Mention several times, including in 2025, placing him

amongst the oldest persons to be so honoured. He

was President of the Indian Association of General

Relativity and Gravitation many years ago.

Naresh had very simple sounding explanations

for profound and diﬃcult to understand concepts like

the universality of space and time, constancy of the

velocity of light, the curvature of space-time and its

manifestation as gravity, derivations of Einstein’s ﬁeld

equations and so forth. He lectured on these matters to

diverse audiences, ranging from professional scientists

to students barely out of school, with the same words

and elan. Over the last few months, I have heard

him speak on these topics to early college students in

Darjeeling and Masters and research students in Silchar.

It was never clear to me how much of these lectures

were actually understood by the young people listening

to him, but they certainly enjoyed the experience of

listening to a person who looked exactly like a scient

should, and spoke from his heart, wholly believing

everything that he said.

Around 1987, Naresh’s life took a dramatic turn.

Jayant Narlikar wanted to set up a inter-university centre

for astronomy, which took form as IUCAA, and Naresh

played a pivotal role in creating the place. He was the

ﬁrst person to be appointed on the roles of IUCAA, even

before Jayant Narlikar’s appointment as the Founding

Director. Naresh worked tirelessly providing liaison

between SPPU, UGC, the government of Maharashtra

and several miniseries at the Centre. The very diﬃcult

task of setting up an institution with a complex structure

was made so much easier because of his eﬀorts. He

helped in identifying and transferring a piece of land

on the SPPU campus for the new institute.

The message that IUCAA had been created had

to be taken far and wide in the country. Naresh and I

travelled incessantly, sometimes accompanied by Jayant

Narlikar and other colleagues, to the main cities, as

well as to universities and colleges in the smaller cities

and towns in far ﬂung areas. We talked to the faculty

and students there about the facilities that IUCAA

oﬀered. We convinced them that their own specialties in

physics, mathematics and statistics could ﬁnd wonderful

applications in astronomy. Their expertise could be

harnessed to solving the new problems emerging from

the multitude of new telescopes and satellites, and

the use of emerging information technology. Soon

the tide turned, and more and more people stated

visiting IUCAA regularly to collaborate with the people

there. New groups of astronomers emerged in several

universities and colleges, and now we have a thriving

community contributing to the national astronomical

eﬀort. We were particularly successful in West Bengal,

Assam, other North-Eastern states, Kashmir and Kerala,

where there are a large number of talented young

people looking for new opportunities. Naresh and

I continued to travel to various centres until he left

for China a few weeks ago. We did that together for

36 years, acquiring in the process many collaborators

and friends. Our younger colleagues now continue the

tradition, bringing to IUCAA increasing numbers of

highly creative modern young people.

Naresh had deep and abiding friendships and

collaborations with a number of relativists and

astrophysicists in many countries. There was an unusual

element to these collaborations. While he worked with

leading scientists in Western countries, and visited

them often, he also collaborated over decades with

groups in various countries like Uzbekistan and other

Central Asian republics, Iran, Turkey, Pakistan and

South Africa. Faculty as well as students from these

countries visited IUCAA regularly, and many Ph.D.s

were produced through these interactions. His ties

with South Africa began soon after apartheid ended and

during his many visits to the country, he developed close

ties with the intellectual elite there, including scientists

from diﬀerent ﬁelds, artists, constitutional court justices,

senior science administrators, vice-chancellors and

others, and he even had a meeting with Nelson Mandela

soon after his release from prison.

Over the years, and especially when he was

Director, Naresh helped in making large astronomical

facilities available to IUCAA. He got the telescope

at Girawali going, and he made IUCAA a partner

in the Southern African Large Telescope. He also

began the process of IUCAA and other institutions

in the country becoming partners in the Thirty Meter

Telescope project, and building a LIGO gravitational

wave detector in India. Sanjeev Dhurandhar and Naresh

tried to set a up a gravitational wave detector in the

country about 25 years ago, but they were far ahead of

their time. Their pioneering eﬀorts were not in vain,

however, since they laid the foundation for the approval

of the LIGO-India project in 2016.

Naresh had intellectual convictions which went far

beyond his scientiﬁc side. In spite of his early family

background, he was a non-believer and rationalist. He

was a committed socialist and ﬁrmly believed in the

equality of all men and women, young and old, and

rich and poor. He was an activist and in spite of

his busy schedule, often participated in marches and

demonstrations on a variety of causes, including the

environment. He lived a simple life bordering on the

spartan, but he liked his fun too. He greatly enjoyed

going to parties and organising many parties himself.

Given the current rightwards march all over the world, I

used to tell him that he would soon be the last jholawala

still standing. He would of course have found a kindred

spirit in the newly elected Mayor of New York, but he

passed away before I could tell him that.

Naresh had a vast circle of friends and admirers in

his home city of Pune and elsewhere in the country. He

had very close ties with people from the performing arts

and theatre all over. For many years, great playwrights,

artists and thespians visited and performed in IUCAA,

providing a rich and multihued background to the

excellent science being done there.

Naresh is survived by his wife Sadhana, who is

a deeply committed social worker and activist, his

daughter Juee who is an entrepreneur in Information

Technology, and his son Nishith, who is a successful

producer of movies and serials.

Naresh clearly was a person of many parts. His

tall, handsome, voluble presence, his infectious laughter

and gracious company will be missed by many for long.

About the Author

Professor Ajit Kembhavi is an emeritus

Professor at Inter University Centre for Astronomy

and Astrophysics and is also the Principal Investigator

of the Pune Knowledge Cluster. He was the former

director of Inter University Centre for Astronomy and

Astrophysics (IUCAA), Pune, and the International

Astronomical Union vice president. In collaboration

with IUCAA, he pioneered astronomy outreach

activities from the late 80s to promote astronomy

research in Indian universities.

My Childhood Memories of Naresh Dadhich

by Joe Philip Ninan

airis4D, Vol.3, No.12, 2025

www.airis4d.com

My fond memories of Prof. Naresh Dadhich go

back to the astronomy workshops that IUCAA used

to conduct at Charal Mount, a hill campus near St.

Thomas College, Kozhencherry, Kerala, in the late 90s.

I was a middle school kid then, accompanying my father

to these workshops. Dadhich was a towering ﬁgure to

me—both ﬁguratively and literally! All the kids adored

him, and we would eagerly queue up to be picked up

and swung around by him. Those were the best ﬂying

experiences ever!

He was also an extraordinary teacher. He would

patiently explain the core concepts of the special theory

of relativity and even the general theory of relativity

at a level that any school kid could understand. No

matter how tired he was after lecturing to the main

workshop participants, he would happily listen to our

ıve questions and explain things from ﬁrst principles.

His excellence as a teacher was evident not only in how

he answered our queries, but also in how he encouraged

us to think more deeply about each problem. Once,

after explaining how the presence of any energy is

suﬃcient to cause curvature in the spacetime metric

and bend light, I remember asking him whether that

meant two parallel light photons travelling billions of

light years next to each other would be attracted to

one another and end up bunching together. He smiled

warmly, appreciated the question, and encouraged me

to think about it more deeply and try to come up with an

answer myself by the next day. I didn’t have an answer

then, but reﬂecting on that problem deﬁnitely sparked

a deep curiosity to learn more physics throughout my

school years.

Later in my life, I had the honour of meeting

him multiple times at IUCAA during various stages

of my career, and every interaction was insightful and

inspiring. He was my childhood hero, who played a

signiﬁcant role in inspiring me to take up a career as a

scientist.

Picture from the 1997 workshop at Charal Mount,

I am standing third from the right in this photo, with my

sister, mother and father between Prof. Ajit Kembhavi

and Prof. Narsh Dadhich.

About the Author

Joe Philip Ninan, works on exoplanets,

and astronomy instrumentation, He is currently a

faculty in the Astronomy department at Tata Institute

of Fundamental Research, Mumbai.

Part I

Artiﬁcial Intelligence and Machine Learning

Few Shot Learning

by Arun Aniyan

airis4D, Vol.3, No.12, 2025

www.airis4d.com

1.1 Introduction

One of the most signiﬁcant and pervasive

challenges in modern Artiﬁcial Intelligence, particularly

within the ﬁeld of deep learning, is the data scarcity

problem. This issue refers to the fundamental limitation

that, while incredibly powerful and capable of complex

pattern recognition, traditional deep learning models

are inherently data-hungry. They typically require an

enormous corpus of labeled examples—often scaling

into the thousands or even millions—to be trained

eﬀectively and to achieve robust, reliable generalization

to unseen data.

This profound reliance on vast, meticulously

curated datasets presents a major and often

insurmountable bottleneck in a myriad of critical real-

world scenarios. Consider domains such as medical

imaging, where obtaining millions of examples of a rare

disease is simply infeasible due to patient privacy, cost,

or the sheer infrequency of the condition. Similarly, in

specialized industrial applications, robotics, or natural

language processing for low-resource languages, the

cost and eﬀort of manually labeling suﬃcient data are

prohibitive. The data scarcity problem is not merely

an inconvenience; it actively hinders the deployment of

deep learning solutions in areas where data acquisition

is inherently diﬃcult, expensive, sensitive, or slow,

thereby creating an urgent need for more data-eﬃcient

learning paradigms.

1.2 Practical Challenges of Standard

Deep Learning

Standard deep learning models are fundamentally

rooted in the principle of statistical learning. This

paradigm dictates that a model’s intricate network

of parameters is ﬁne-tuned through an iterative

optimization process, speciﬁcally by minimizing a

deﬁned loss function across an expansive dataset. The

eﬃcacy of this process hinges on a critical trade-

oﬀ: the volume of data must be immense enough

to simultaneously prevent overﬁtting—the detrimental

scenario where the model merely memorizes the

idiosyncrasies and noise of the training data—and

to guarantee robust generalization—the model’s true

measure of success, which is its ability to perform

accurately and reliably on entirely unseen, out-of-

sample data. The prerequisite for these massive

datasets, however, introduces several signiﬁcant and

often prohibitive practical diﬃculties across various

real-world applications:

The Financial and Temporal Burden of Data

Acquisition

The process of acquiring raw data is often just

the beginning. The subsequent step—labeling

or annotating this data with accurate

ground-truth information (e.g., classifying an

image, transcribing audio, or delineating

boundaries)—is an incredibly resource-intensive

endeavor. It demands signiﬁcant ﬁnancial

investment to hire and train human annotators and

consumes substantial amounts of time, creating

a bottleneck that delays the development and

1.3 Few-Shot Learning: A Paradigm Shift in Machine Learning

deployment of new AI systems. This cost scales

linearly, or often super-linearly, with the desired

size and complexity of the dataset.

Constraints Imposed by Privacy and Ethical

Considerations

In highly sensitive and regulated domains, such

as healthcare (e.g., patient records, medical

scans) and ﬁnance (e.g., transaction histories,

personal credit data), the volume of accessible

data is severely restricted. Strict government

regulations (like GDPR, HIPAA) and corporate

policies, driven by critical ethical concerns over

individual privacy and data security, mandate

data anonymization, aggregation, or outright

restriction, making the assembly of massive,

comprehensive datasets practically impossible.

Challenges in Handling Novelty, Rarity, and

Extreme Class Imbalance

Deep learning struggles profoundly when faced

with categories that are newly emerging or

extremely rare. By their nature, these categories

preclude the possibility of gathering the large,

statistically representative datasets required by

standard models. Examples include:

Scientiﬁc Discovery: A new, never-

before-seen viral strain or an anomaly in

astronomical data.

Ecology: A rarely observed or newly

discovered animal or plant species.

Security: A novel, sophisticated zero-day

cyber threat or a rare type of equipment

failure. The lack of historical data for

these ’outlier’ events makes the statistical

learning approach ineﬀective, as the model

has little to no data to learn from, making

accurate prediction and classiﬁcation nearly

impossible.

In summary, the statistical appetite of standard

deep learning forms a powerful barrier to entry for

many critical applications, necessitating innovative

approaches—such as Few-Shot Learning—that can

learn eﬀectively from minimal data.

1.3 Few-Shot Learning: A Paradigm

Shift in Machine Learning

Few-Shot Learning (FSL) represents a critical

evolution in the ﬁeld of machine learning, emerging as

a vital solution to the pervasive challenge of data scarcity.

At its core, FSL is a specialized subﬁeld dedicated to

creating models capable of learning eﬀectively from

only a handful of labeled examples, often as few as

one or ﬁve. This approach stands in stark contrast to

traditional deep learning methodologies, which require

massive, meticulously curated datasets (often thousands

or millions of examples) to achieve high performance.

FSL aims to mimic the remarkable eﬃciency of human

learning, where a person can typically recognize and

categorize a new concept—like a novel type of bird

or a new gadget—after seeing just a single or a few

instances.

The fundamental objective of FSL is not simply

to achieve high accuracy with less data, but to instill

a deeper, more adaptable form of intelligence in AI

systems. This is often achieved through meta-learning,

where the model is trained on a vast array of tasks to

learn how to learn quickly, rather than just learning a

single task. The knowledge gained from a diverse set

of ”training tasks” is then rapidly transferred to a new,

data-poor ”test task.”

1.3.1 The Critical Need for FSL in

Real-World Applications

The signiﬁcance of Few-Shot Learning is

most pronounced in real-world scenarios where the

conditions for traditional big-data AI are diﬃcult or

impossible to meet. In these domains, data collection

is inherently:

Expensive and Time-Consuming: The manual

labeling and veriﬁcation of specialized data

require expert human labor and signiﬁcant

ﬁnancial investment.

Ethically Constrained or Private: Certain data,

such as individual health records or biometric

proﬁles, are protected by privacy laws, severely

limiting the size of available datasets.

1.4 Core Concepts in Few-Shot Learning

Atypical or Rare: By deﬁnition, some

phenomena occur infrequently, making large-

scale data collection fundamentally impossible.

FSL is therefore critical for enabling AI adoption

across several high-impact sectors:

Medical Imaging and Diagnostics: This is a

particularly crucial application. FSL allows for

the training of diagnostic models for rare diseases

or novel strains of viruses where only a minute

number of conﬁrmed, expertly-labeled cases exist

globally. It provides a path to deploy life-saving

AI tools even when the data is extremely scarce.

Scientiﬁc Discovery and Conservation: In

ﬁelds like biology and astronomy, FSL can be

used to rapidly identify and classify new species

of plants, animals, or celestial objects based on

initial, limited observations. This accelerates the

pace of discovery and aids in conservation eﬀorts

for endangered or newly found organisms.

Robotics and Autonomous Systems: For robots

to be truly autonomous, they must be able to

rapidly adapt to new, unseen objects, tools, or

environmental conditions without needing hours

of pre-training. FSL allows a robot to learn

the function of a new grip or the layout of an

unfamiliar room with minimal demonstration.

Personalized AI and User Experience: FSL

enables the development of user-speciﬁc models

with exceptional eﬃciency. This is vital for

personalization, where an AI assistant might

learn a user’s unique speech patterns, handwriting

style, or speciﬁc product preferences with just

a few examples, enhancing privacy and user

experience.

By empowering models to glean meaningful insights

from limited data, Few-Shot Learning represents a

fundamental step toward achieving Artiﬁcial General

Intelligence (AGI). It pushes AI closer to the

adaptability, eﬃciency, and cognitive ﬂexibility that

deﬁne human learning.

1.4 Core Concepts in Few-Shot

Learning

Few-Shot Learning typically involves two main

phases: the Meta-Training Phase and the Few-Shot Task

Phase.

1.4.1 Meta-Training Phase (Training the

’Learner’)

In this phase, the model is trained on a large base

dataset composed of many diﬀerent classes. The goal

is not to learn to classify these base classes perfectly,

but to learn a robust and generalizable learning strategy

or a good feature extractor.

Data Structure: The base dataset is split into

numerous ”episodes” or ”tasks.” Each episode

is a mini few-shot problem itself, designed to

simulate the ﬁnal few-shot task.

Objective: The model learns a universal

mechanism that can quickly adapt to a new task

with minimal examples.

1.4.2 Few-Shot Task Phase (Applying the

’Learner’)

This is the real test. The model is presented with a

novel dataset containing classes it has never seen before

during meta-training. This dataset is structured into

a support set and a query set as shown in Table 1.1.

This structure is often deﬁned as an

-way

-shot

Table 1.1: Task Description

Set Name 1 Purpose 2 Example Size

Support set S Contains the few labeled examples of the new classes. 2 N classes, K examples each

Query Set Q Contains unlabeled examples of the new classes for prediction.

classiﬁcation task with:

N-way: The number of novel classes.

-shot: The number of labeled examples per

class (i.e., the ”few shots”).

An example would be

way

shot task, which

means if

N = 5

and

K = 1

, the model sees only one

labeled example for each of the ﬁve new classes in

the support set and must then correctly classify new

examples in the query set.

1.6 The Few-Shot Learning Process Workﬂow

1.5 Key Approaches to Few-Shot

Learning

Several innovative techniques have been developed

to tackle the FSL problem. The primary goal is either

to ﬁnd a better starting point for learning or to improve

the classiﬁcation process itself.

1.5.1 Metric-Learning Approaches

Metric-learning methods aim to learn an

embedding space (a feature representation) where

examples of the same class are clustered closely together,

and examples of diﬀerent classes are far apart.

Prototypical Networks: This is a popular and

intuitive approach. For a given few-shot task, the

model calculates a prototype (the mean vector)

for each class in the support set. New query

examples are classiﬁed based on the distance

(e.g., Euclidean distance) to the closest prototype.

Relation Networks: Instead of just measuring

distance, these networks learn a non-linear

similarity metric. They take a pair of images

(a support image and a query image) and output

a score indicating how likely they belong to the

same class.

1.5.2 Model-Agnostic Meta-Learning

(MAML)

MAML is a highly inﬂuential technique that

focuses on ﬁnding a set of initial model parameters

that are very sensitive to change. This means that with

only a few gradient descent steps and a few examples

(

-shots), the model can quickly adapt and perform

well on a new task.

The meta-training objective is to minimize the ﬁnal

loss after performing one or a few inner-loop gradient

steps. It learns a universal initialization point for the

learner.

1.5.3 Data Augmentation and Generative

Models

Another strategy is to increase the amount of data

by synthesizing new, realistic examples. While standard

data augmentation (e.g., rotation, cropping) helps, FSL

often uses more sophisticated generative models (like

GANs or VAEs) to create synthetic examples of the few-

shot classes, eﬀectively turning the

-shot problem

into a richer-shot problem.

1.6 The Few-Shot Learning Process

Workﬂow

Few-Shot Learning involves a systematic, multi-

step process. The steps below outline a typical metric-

learning workﬂow:

1. Meta-Training

Train a feature encoder network on episodes from

the base dataset to learn a robust feature space.

2. Task Formulation

Sample an

-way

-shot task, creating a

Support Set

and a Query Set

with novel

classes.

3. Feature Extraction

Use the trained encoder to map all images in

and Q into the embedding space.

4. Prototype Calculation

For each class in

, calculate the class prototype

(e.g., the average of the feature vectors).

5. Prediction (Inference)

Classify each query example by ﬁnding the

prototype it is closest to in the embedding space.

1.7 Conclusion

Few-Shot Learning (FSL) is not merely an

incremental improvement over existing deep learning

techniques; it represents a paradigm shift toward more

eﬃcient, adaptable, and human-like AI. We have seen

that FSL directly addresses the most crippling limitation

of standard deep learning—its insatiable demand for

data—by enabling robust model performance from

only a handful of examples. Through meta-learning

1.8 References

strategies, metric learning, and sophisticated data

synthesis, FSL allows models to acquire the crucial

skill of learning how to learn.

The immediate and long-term potential of FSL is

transformative, promising to unlock AI deployment in

high-stakes, data-scarce domains previously considered

intractable:

Accelerating Scientiﬁc Discovery: FSL is

poised to signiﬁcantly accelerate research in areas

like medical diagnostics for rare diseases, real-

time identiﬁcation of novel pathogens, and rapid

classiﬁcation in ecological and astronomical

surveys, where data inherently exists in small

batches.

Enabling Adaptive Robotics: For autonomous

systems and robotics, FSL provides the necessary

cognitive ﬂexibility for quick adaptation to

new tools, environments, and tasks, moving

robots closer to true independence and utility

in unpredictable human settings.

Fostering Personalized AI: FSL allows for the

creation of hyper-personalized models that learn

an individual’s speciﬁc preferences, biometric

data, or interaction styles with minimal data

points, greatly enhancing user privacy and the

quality of user-AI interaction.

In essence, FSL moves Artiﬁcial Intelligence

beyond the brute-force statistics of ”big data” toward

a model of genuine intelligence characterized by

eﬃciency and generalization. For the next generation

of AI researchers, Few-Shot Learning represents a

crucial, active frontier—a fundamental step in realizing

the dream of Artiﬁcial General Intelligence, where

machines can learn new concepts as rapidly and

eﬀectively as humans. The ongoing research into more

sophisticated meta-learning architectures and generative

models promises to further solidify FSL’s role as the

cornerstone of data-eﬃcient machine learning.

1.8 References

https://towardsdatascience.com/few-shot-

learning-a-new-paradigm-in-ai-a5170f438a95

https://arxiv.org/abs/1904.05046

https://lilianweng.github.io/posts/2018-11-20-

meta-learning/

About the Author

Dr.Arun Aniyan is leading the R&D for

Artiﬁcial intelligence at DeepAlert Ltd,UK. He comes

from an academic background and has experience

in designing machine learning products for diﬀerent

domains. His major interest is knowledge representation

and computer vision.

The Time Arrow of Entropy

by Jinsu Ann Mathew

airis4D, Vol.3, No.12, 2025

www.airis4d.com

Entropy is usually described as a measure of

disorder or uncertainty. But real systems—whether

physical, biological, or social—do not stay still. They

move, change, and carry memories of their past. When

we look at entropy through the lens of time, a deeper idea

appears: change has a direction. Some things naturally

happen in one way but not the reverse—heat spreads

out, languages evolve, societies grow, and information

slowly gets lost.

This one-way movement is known as the arrow

of time, and entropy is what gives it direction. By

understanding how entropy changes over time, we can

see why systems age, why they resist going backwards,

and how their past shapes what they become.

2.1 Irreversibility: Why Time Has a

Direction

Irreversibility means that some things in life

happen in only one direction. Even though the basic

rules of physics can run forward or backward, the world

we see doesn’t behave that way. Time feels like it moves

forward because certain changes naturally go from order

to disorder and almost never the other way around. This

is why we remember yesterday but not tomorrow, why

things age instead of becoming new again, and why

many processes cannot simply “undo” themselves.

A simple example is a glass falling and breaking.

When the glass is whole, all its pieces are arranged

in a very neat, organized way. Once it breaks, the

pieces scatter in many diﬀerent positions. It’s easy for

a whole glass to break because there are many ways

for the pieces to land, but it’s almost impossible for the

(image courtesy:AI-generated image)

Figure 1: The thermodynamic arrow of time

scattered pieces to jump back and form the original glass

again. So in everyday life, we only see the breaking,

not the un-breaking.

Another familiar example is heat. If you place

a hot spoon in cold water, the heat spreads into the

water until both become warm. But the warmth never

suddenly leaves the water to make the spoon hot again.

The mixing of heat happens naturally, but the reverse

doesn’t. These everyday experiences show us why time

feels like it goes in only one direction: some changes

happen easily, while their opposites are so unlikely that

we never see them. That is what gives time its direction.

2.2 Memory in Natural and Artiﬁcial

Systems

Memory in a system means that its past aﬀects

how it behaves today. Some systems hold on to their

2.4 Entropy Over Time: Patterns, Trends, and Transitions

previous states for a long time, while others forget

quickly. This “memory” doesn’t always mean conscious

remembering—it can simply be a pattern, a structure,

or a change that stays even after the original cause is

gone. Because of this, systems do not start fresh at

every moment. Their history continues to shape what

they do next.

In nature, many things show this kind of memory.

A magnet keeps its direction even after the external

force that aligned it is removed. Trees remember past

seasons by forming growth rings that reﬂect dry or rainy

years. Even the human body carries memory—your

immune system “remembers” past infections and reacts

faster when it encounters the same germ again. These

are all examples where the past leaves a mark that

inﬂuences the future. Artiﬁcial systems, like machines

and computers, also have memory. A simple thermostat

“remembers” the previous temperature setting and

adjusts the room based on that. A trained neural network

remembers patterns from the data it learned and uses

them to make predictions. Even phone keyboards have

memory: they learn the words you type often and

suggest them more quickly. In all these cases, the

system’s past experience guides how it behaves now.

Memory slows down how quickly a system changes.

Instead of jumping immediately into randomness or

disorder, the system holds on to what it has learned or

experienced. This is why memory plays an important

role in how systems evolve over time—it keeps the past

alive and shapes the direction of future behaviour.

2.3 History in Language and Society

History plays a powerful role in shaping both

language and society. Neither of them starts from

scratch at any moment; instead, they grow on top

of what came before. This means that the way

people speak, behave, and organize themselves today

is strongly inﬂuenced by choices, habits, and patterns

that developed over many years. Because of this deep

inﬂuence of the past, changes in language or society

usually happen slowly and rarely return to earlier forms.

In language, history leaves clear traces. Words

change their meanings over time, but they do not

suddenly become what they were centuries ago.

Grammar also shifts gradually, building on earlier usage

rather than reinventing itself. For example, English no

longer uses words like “thou” and “thee,” and it will

not suddenly bring them back into everyday speech.

New words and styles appear, but they grow out of

older ones. The language we speak today is a layered

record of how people spoke in the past. Society shows

the same pattern. Customs, laws, and social norms

evolve over time, shaped by many years of shared

experience. Traditions—such as festivals, greetings, or

family roles—continue because they have been passed

down through generations. Even when societies change,

they do so gradually, carrying forward parts of their

history. For instance, cities develop new technologies

and lifestyles, but old neighbourhoods, cultural habits,

and social expectations often remain. The past acts like

an anchor, inﬂuencing how people behave and how they

make decisions.

Because language and society carry their history

with them, they rarely move backward. Change happens

in a forward direction, building on what already exists.

This is why history becomes an active force: it keeps

certain patterns stable, guides how new ones form, and

shapes the way communities grow and communicate

2.4 Entropy Over Time: Patterns,

Trends, and Transitions

As we follow entropy through time, we begin to

notice the patterns in how systems change. Entropy

does not stay ﬁxed; it shifts in response to the small and

large events happening inside the system. Sometimes

these changes are gentle and gradual, and sometimes

they are sudden. Watching these changes helps us

understand the rhythm of a system’s life—how it grows,

adapts, or stabilizes.

In many systems, entropy moves in small daily

ﬂuctuations. These little rises and falls come from

ordinary variations: a slight change in how people speak,

a tiny shift in temperature, or a brief change in social

behaviour. These small movements may seem trivial,

but they show that the system is active and constantly

2.5 Conclusion

responding to its environment. Over longer periods,

clearer trends appear. Some systems slowly become

more uncertain or diverse—for example, a language

steadily adding new words or a community becoming

more interconnected over years. Other systems might

become more organized, as when a group settles into

new routines or when a learning system gradually

becomes more conﬁdent and consistent. These long-

term trends show the direction a system is heading.

At times, however, entropy does not change slowly

at all. Instead, it jumps. These jumps often signal major

turning points—a sudden shift in public mood, a rapid

change in technology, or a natural system reaching a

threshold and reorganizing itself. These transitions are

important because they mark moments when the usual

patterns break and something new begins. By looking

at entropy over time, we gain a clearer picture of how

systems behave. It becomes easier to see when they are

stable, when they are slowly drifting, and when they

are on the edge of change. Entropy, viewed this way,

becomes not just a measure of uncertainty but a tool for

understanding how systems move and evolve.

2.5 Conclusion

Looking at entropy through the lens of time helps

us understand why systems move the way they do.

Whether we study physical processes, languages, or

societies, we ﬁnd the same pattern: the past shapes the

present, and changes tend to move in a forward direction.

Irreversibility, memory, and history each show us how

systems carry traces of what they were, and how these

traces guide what they become. By watching how

entropy shifts—slowly, quickly, or suddenly—we can

see not just the current state of a system but the path it

is taking. Entropy, in this sense, becomes more than

a measure of uncertainty. It becomes a way to read

the story of a system’s evolution, showing how small

actions build into larger patterns and how the arrow of

time leaves its mark on everything that changes.

References

Entropy - Physics

Entropy Is Universal Rule of Language

Hidden Information, Energy Dispersion and

Disorder: Does Entropy Really Measure

Disorder?

Where Does Our Arrow Of Time Come From?

Can Time really heal anything?

About the Author

Jinsu Ann Mathew is a research scholar

in Natural Language Processing and Chemical

Informatics. Her interests include applying basic

scientiﬁc research on computational linguistics,

practical applications of human language technology,

and interdisciplinary work in computational physics.

Unboxing a Transformer using Python - Part

by Linn Abraham

airis4D, Vol.3, No.12, 2025

www.airis4d.com

Figure 1: Position and Class embedding in ViT

3.1 Introduction

In the ﬁrst part of this article, we looked into the

patch embedding layer part of the Vision Transformer.

In this article, we continue from there and look into

how the position and class information are learnt and

then onto how the transformer is implemented.

3.2 Position & Class Embedding

By creating patches and ﬂattening them out, we

have destroyed the position information of the patches

in the original 2D grid. Also, since these patches are

processed in parallel by the attention mechanism, even

the 1D ordering of the patches is meaningless. But

the spatial position is a physical information that we

want the model to learn. Additionally, since this is

a classiﬁcation model, we also want the class label

to be learnt by the model. Thus we want to to

have placeholders for these two important pieces of

information in the embedding and then let the model

learn it.

To see how how this is implemented, note that

we ﬁrst embed the class token. The class token is

deﬁned and initialized as a learnable parameter of

shape

[1, 1, dim]

(see, line 7 in the code). The

repeat

function takes this class token and repeats it along

a dimension as speciﬁed by the pattern. The new

variable b which deﬁnes the number of repetitions to

be made along the zeroth axis is passed as an input

to the function. These are then concatenated to the

patch embeddings along its 1st dimension which is

the sequence dimension with size equal to number of

patches in the image. Thus the new shape becomes

[b, n + 1, dim]

with the ﬁrst token in each sequence

being the class token followed by the patch tokens. The

positional embedding that we deﬁned earlier is now

added to the output of the class-token embedded tensor.

Note that this is an element-wise addition and not a

concatenation as with the case of the class token. Also

note that there is broadcasting happening in the zeroth

dimension (the 1 gets broadcasted to the batch size).

Other than this, there is no change in the output shape

as a result of this addition. Also we require n+1 in the

sequence dimension (which might be to allow for cases

where the number of patches are variable or something

like that).

3.3 The Transformer

Let’s see what has happened so far. We started

with a batch of images, which have been made into a

batch of sequences with each sequence containing the

image patches, that have been ﬂattened along the height,

REFERENCES

Figure 2: Deﬁnition of the transformer in ViT

width and channel dimensions, and with an extra class

token added to it. What remains to be explained is the

transformer block, which we partly saw in the beginning

of this article.

After the class and position embedding steps,

a dropout layer is added before being fed into the

transformer block. Within the transformer block we ﬁrst

see a normalization layer. After this comes the

to qkv

function which is does a linear projection to a dimension

which is three times the inner dim. This inner dimension

is the product of the dimension per attention head times

the number of attention heads. So this is where the Q,

K and V matrices are learnt. After this it is again split

into the three diﬀerent matrices. And then reshaped to

split into each individual head so that the dot product

of the q and k vectors can be taken. The dot product is

implemented as a matrix multiplication with one vector

transposed. The transpose is taken by swapping the

last two dimensions. At this stage the output is also

scaled for stability. This is followed by another dropout

layer. After this a weighted addition of the value vectors

with the attention scores are taken using regular matrix

multiplication. The ﬁnal rearrangement converts the

single head into the multi-head output. If more than

one head is used add another projection from the inner

dimension to the input dimension otherwise just use an

identity operation.

3.4 The Final layers

After the transformer block we have a pooling

layer. Two valid options are mean or class. If mean, all

the tokens are averaged else the class token is taken as

the representation for the image. This is followed by the

MLP head which is a linear layer with size equal to the

Figure 3: Deﬁnition of Attention in ViT

number of classes. In between we have a

to latent

layer which is just an identity operation in this case.

This is done so that it can be swapped with something

else if required without changing the forward function.

References

[Goodfellow et al.(2016)Goodfellow, Bengio, and Courville]

Ian Goodfellow, Yoshua Bengio, and Aaron

Courville. Deep Learning. Adaptive Computation

and Machine Learning. The MIT press, Cambridge,

Mass, 2016. ISBN 978-0-262-03561-3.

[Dosovitskiy et al.(2021)]

Alexey Dosovitskiy, Lucas

Beyer, Alexander Kolesnikov, Dirk Weissenborn,

Xiaohua Zhai, Thomas Unterthiner, Mostafa

Dehghani, Matthias Minderer, Georg Heigold,

Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby.

AN IMAGE IS WORTH 16X16 WORDS:

TRANSFORMERS FOR IMAGE RECOGNITION

AT SCALE. 2021.

[Chollet(2025)]

Francois Chollet. Deep Learning

With Python, Third Edition. MANNING

PUBLICATIONS, S.l., 2025. ISBN 978-1-63343-

658-9.

[Alammar()]

Jay Alammar. The Illustrated

Transformer. https://jalammar.github.io/illustrated-

transformer/.

REFERENCES

About the Author

Linn Abraham is a researcher in Physics,

specializing in A.I. applications to astronomy. He is

currently involved in the development of CNN based

Computer Vision tools for prediction of solar ﬂares

from images of the Sun, morphological classiﬁcations

of galaxies from optical images surveys and radio

galaxy source extraction from radio observations.

Part II

Astronomy and Astrophysics

Dusty Plasma

by Abishek P S

airis4D, Vol.3, No.12, 2025

www.airis4d.com

1.1 Introduction

Plasma is a distinct state of matter that arises when

a gas becomes highly ionized, producing a mixture of

free electrons and ions. Although ﬁlled with charged

particles, plasmas are generally quasi-neutral, meaning

that positive and negative charges balance out on large

scales. A key feature of plasma is Debye shielding,

where electric ﬁelds are screened over a characteristic

distance known as the Debye length [1]. This prevents

long-range electrostatic interactions from dominating its

behaviour. Remarkably, plasma makes up about 99.9%

of the visible universe, encompassing stars, interstellar

gas, and the solar wind. Plasma can be classiﬁed into

several types based on their temperature, ionization,

density, collisions, and magnetic properties. The main

categories include cold plasma, warm plasma, hot

plasma, ultracold plasma, fully ionized plasma, partially

ionized plasma, collisional plasma, non-collisional

plasma, neutral plasma, non-neutral plasma, high/low

density plasma, magnetic plasma, and dusty plasma

[2,3]

Dusty plasma, also called complex plasma, is a

type of plasma that contains tiny solid particles ranging

from nanometres to micrometres in size that become

electrically charged and interact with the surrounding

ions and electrons. These dust grains can originate

naturally, such as in interstellar clouds, planetary rings,

and Earth’s mesosphere, or be artiﬁcially introduced

in laboratory experiments. Once charged, the dust

particles signiﬁcantly alter the plasma’s behaviour by

disturbing the balance of charged particles, leading to

new phenomena not seen in ordinary plasmas. For

example, dusty plasmas can form plasma crystals,

where dust grains arrange themselves into ordered

lattice structures due to strong electrostatic coupling.

Depending on conditions, they may behave like a gas,

liquid, or solid, making them a fascinating system for

studying phase transitions and collective eﬀects. Dusty

plasmas are important in astrophysics because they help

explain processes in star formation, comet tails, and

cosmic dust clouds, while in technology they appear in

semiconductor manufacturing, fusion devices, and even

spacecraft propulsion [2]. Their study bridges plasma

physics, materials science, and astrophysics, oﬀering

insights into how matter behaves under extreme and

complex conditions.

1.2 Formation of Dusty Plasma

Dusty plasma formation begins with the

introduction or creation of solid particles (dust grains)

in a gas that can be ionized; these particles may

be naturally present (cosmic dust, cometary ejecta,

planetary rings, mesospheric aerosols) or produced in

situ by mechanical, chemical, or sputtering processes

in laboratory and industrial discharges. When the

background gas is ionized by thermal energy, electric

discharges, photoionization, or energetic particle ﬂux

the resulting free electrons and ions collide with and

are collected by the dust grains, so that grains rapidly

acquire net charge (typically negative, because electrons

are far more mobile than ions). Additional charging

channels such as photoemission (from ultraviolet

illumination) and secondary electron emission (from

energetic ion or electron impacts) can modify the

1.4 Collective behaviour of dusty plasma

sign and magnitude of the grain charge, producing

spatial and temporal charge variations. As grains

charge, they perturb the local plasma: mobile electrons

and ions rearrange to screen the grain potential over

a Debye length, so interactions between grains are

screened Coulomb (Yukawa) interactions rather than

pure Coulomb forces [4]. In many laboratory and

space settings, the balance of forces on a grain-

electrostatic lift from sheath or ambipolar ﬁelds, gravity,

ion drag, neutral gas drag, and thermophoretic forces

determines whether grains levitate, drift, or settle;

in plasma sheaths above electrodes, for example,

electric ﬁelds can levitate negatively charged grains

against gravity, enabling stable suspensions and the

formation of ordered structures. As the dust density and

coupling strength increase, strong electrostatic coupling

between grains can drive transitions from gaseous to

liquid-like and even crystalline arrangements (so called

plasma crystals), while collective modes unique to

dusty plasmas such as dust acoustic waves emerge

because the massive, charged grains introduce new

low-frequency dynamics. The charging process itself

is often time-dependent: ﬂuctuating plasma conditions,

variable illumination, and grain motion cause charge

to vary on timescales comparable to or longer than

grain dynamical times, coupling microscale charging

kinetics to macroscopic transport and instabilities.

In many natural environments (protoplanetary disks,

cometary comae, interstellar clouds) coagulation,

charging, and plasma drag together inﬂuence grain

growth and dynamics, while in technological plasmas

(semiconductor processing, fusion edge plasmas) dust

formation and charging can degrade performance or

seed new instabilities. The net result is a self-consistent,

multi-scale formation process in which ionization,

grain charging, screening, and force balance produce

a complex medium whose structure and waves diﬀer

qualitatively from ordinary two-component plasmas.

1.3 Forces acting on dust particles

Dust particles immersed in an ionized gas rapidly

acquire charge through electron and ion collection,

photoemission, and secondary electron emission, so that

electromagnetic forces often dominate their dynamics;

the electrostatic force F

E produced by local

electric ﬁelds can levitate negatively charged grains in

sheaths or drive them along ﬁeld gradients, while gravity

g pulls them downward and sets a baseline for

levitation height. Streaming ions transfer momentum to

grains producing ion drag (with collection and orbital

components) that can push grains in the ion ﬂow

direction and modify equilibrium positions; collisions

with neutral gas molecules produce neutral (viscous)

drag that damps grain motion and controls relaxation

timescales. Temperature gradients in the neutral gas

give rise to a thermophoretic force that drives grains

from hot to cold regions, and radiation pressure or

photoelectric eﬀects can both exert direct momentum

transfer and change grain charge, altering electrostatic

responses. Charged grains interact with one another

through Debye-screened Coulomb (Yukawa) potentials,

so interparticle electrostatic forces can be strongly

repulsive and, at high coupling, produce ordered

“plasma crystal” lattices or liquid-like behaviour; the

eﬀective coupling is governed by grain charge, spacing,

and the Debye screening length [5,6]. Crucially,

charging is often time dependent ﬂuctuations in

plasma density, temperature, illumination, or grain

motion cause the grain charge to vary on timescales

comparable to dynamical times, coupling microscopic

charging kinetics to macroscopic transport, instabilities,

and wave phenomena such as dust acoustic waves.

In laboratory setups, the balance of electrostatic

lift, ion drag, neutral drag, thermophoresis, and

gravity determines whether grains levitate in sheath

regions or settle, and microgravity experiments reveal

three-dimensional structures otherwise distorted by

gravity. In natural environments (cometary comae,

planetary rings, interstellar clouds, mesosphere) the

same forces, together with coagulation and plasma

chemistry, control grain growth, transport, and radiative

eﬀects. Understanding dust dynamics therefore

requires treating electromagnetic, collisional, and

external forces simultaneously, using kinetic, ﬂuid,

and molecular-dynamics approaches to capture the

multi-scale, strongly coupled behaviour unique to dusty

plasmas.

1.5 Dust Acoustic Waves

1.4 Collective behaviour of dusty

plasma

Dusty plasma collective behaviour arises when

many charged grains interact through screened Coulomb

(Yukawa) potentials[7], so that the ensemble can

behave like a strongly coupled many-body system

whose macroscopic properties depend on grain charge,

interparticle spacing, and the Debye screening length;

under strong coupling the dust component can undergo

phase transitions from gas-like to liquid-like and

even to ordered solid-like states known as plasma

crystals, where grains form lattice structures and support

phonon-like excitations. Because dust grains are

heavy, they introduce low-frequency collective modes

such as the dust acoustic wave, whose dispersion

and damping are governed by dust mass, charge,

neutral drag, and ion dynamics; these modes couple

to ion-acoustic and sheath oscillations, producing

modiﬁed wave spectra and multi-scale interactions.

Charging dynamics and charge ﬂuctuations play a

central role: time-dependent collection of electrons

and ions, photoemission, and secondary emission

cause grain charge to vary, which feeds back on

interparticle forces and can seed instabilities, mode

coupling, and anomalous transport. External ﬁelds

and ﬂows electric ﬁelds in sheaths, ion streaming,

neutral gas ﬂows, and temperature gradients drive

directed transport through ion drag, thermophoresis,

and electrostatic forces, producing layered levitated

structures, voids, and ﬂow-induced ordering; ion

streaming in particular can destabilize ordered states

and excite dust-density waves. Collisional damping

with neutrals and frictional forces set relaxation

times and determine whether collective excitations

are underdamped or overdamped, while strong

coupling enhances correlation eﬀects such as caging,

long-range order, and non-Newtonian rheology in dusty

plasma liquids. Nonlinear phenomena are common:

solitary structures, shocks, vortices, and self-organized

patterns emerge from the interplay of long-range

screened interactions, dissipation, and external forcing.

Laboratory experiments exploit the visualisability

of individual grains to study many-body physics

directly tracking particle trajectories reveals melting,

crystallization, defect dynamics, and transport at the

single-particle level while microgravity experiments

remove gravitational distortion and expose true

three-dimensional collective phases. In astrophysical

and technological contexts, collective dusty plasma

behaviour inﬂuences coagulation and growth of grains

in protoplanetary disks, the formation of spokes and

structures in planetary rings, and contamination and

instability issues in plasma processing and fusion edge

plasmas. Overall, the collective dynamics of dusty

plasmas combine strong coupling, screened interactions,

time-dependent charging, and multi-scale coupling to

produce waves, phase transitions, instabilities, and

self-organization that make them a uniquely accessible

platform for studying complex plasma phenomena

1.5 Dust Acoustic Waves

Dust acoustic waves arise in a dusty plasma when

the massive charged dust grains act as the primary

inertial component and the lighter plasma species

(electrons and ions) supply the restoring pressure,

producing a new low-frequency branch of collective

oscillation distinct from ordinary ion-acoustic modes

[8]. Key properties include a phase velocity that

is typically much smaller than ion thermal speeds,

a dispersion relation that depends explicitly on dust

charge, dust mass, dust density, and the Debye screening

length, and a frequency range often in the Hz to kHz

band in laboratory conditions. Because dust grains

are heavy, the characteristic frequency is low and

the waves are easily observable by direct imaging

of particle motion, which has made dust acoustic

waves a powerful diagnostic in experiments. Damping

and modiﬁcation of the wave arise from neutral gas

friction (viscous drag), ion drag, and especially dust

charge ﬂuctuations: time-dependent charging processes

(electron/ion collection, photoemission, secondary

emission) change the grain charge on timescales

comparable to the wave period and can either damp or

destabilize the mode depending on plasma conditions

[9]. In ﬂowing plasmas, ion streaming relative to the

1.6 Conclusion

dust can drive instabilities that amplify dust acoustic

waves or produce shock-like structures; nonlinear

evolution yields solitary waves, shocks, and vortices

in strongly driven regimes. Theoretical descriptions

typically start from ﬂuid or kinetic models that

include dust continuity and momentum equations

coupled to Poisson’s equation with Debye screening,

and reﬁnements incorporate collisional absorption,

charge variation, strong coupling (Yukawa interactions),

and ﬁnite grain size eﬀects. Experimentally, dust

acoustic waves were predicted and then observed in

laboratory discharges where controlled dust injections

and sheath electric ﬁelds produce levitated dust layers;

modern experiments have visualized wave propagation,

measured dispersion relations, and demonstrated

nonlinear solitary structures and cylindrical solitons.

In space and astrophysical settings, dust acoustic–

like modes can inﬂuence dust transport, coagulation,

and structure formation in cometary comae, planetary

rings, and dusty regions of the interstellar medium.

Overall, dust acoustic waves are a hallmark collective

phenomenon of dusty plasmas [10], linking microscopic

charging physics to macroscopic wave dynamics and

enabling direct study of many-body and nonlinear

processes

1.6 Conclusion

Dusty plasmas ﬁnd broad applications across

both natural-science and engineering domains:

in astrophysics they help explain processes in

protoplanetary disks, cometary comae, planetary rings,

and interstellar clouds where charged grains aﬀect

coagulation, radiative transfer, and structure formation;

in industry, dust formation and charging are central

concerns in semiconductor manufacturing and plasma

processing because contaminant particles degrade

device yield, while in fusion devices dust from wall

erosion inﬂuences plasma purity and safety and must

be managed; in laboratory research dusty plasmas

serve as an accessible tabletop platform for studying

strongly coupled many-body physics, enabling direct

visualization of phenomena such as plasma crystals,

melting, defect dynamics, and wave propagation (e.g.,

dust acoustic waves) that inform both basic science

and applied diagnostics. Key technological uses

include using controlled dusty plasmas for nanoparticle

synthesis and surface modiﬁcation, employing dust

as a diagnostic tracer to map electric ﬁelds and

ﬂows, and exploiting ordered dust structures to study

phase transitions and transport at the single-particle

level. Operational risks notably particle contamination,

discharge nonuniformity, and dust-driven instabilities in

processing and fusion environments mean that practical

deployment requires active monitoring (optical imaging,

laser scattering, probes), cleanliness protocols, and

mitigation strategies such as controlled gas ﬂows,

electrostatic conﬁnement, and periodic cleaning to limit

performance loss and safety hazards.

Dusty (complex) plasmas are both a challenge

and an opportunity: they complicate many plasma

technologies by introducing additional charge, mass,

and transport channels, yet they also provide a

uniquely visible and tuneable system for exploring

collective phenomena and for enabling novel material

and diagnostic techniques. Understanding and

controlling dust charging, screening, and force balances

is essential to harness beneﬁts while minimizing

harms; progress depends on integrated approaches that

combine experiments (including microgravity studies),

kinetic and ﬂuid modelling, and molecular-dynamics

simulations to capture multi-scale coupling, charge

dynamics, and strong-coupling eﬀects. The ﬁeld

therefore sits at the intersection of astrophysics,

materials science, and plasma engineering, oﬀering both

fundamental insights and concrete applications when

dust is measured, modelled, and managed eﬀectively.

References

Francis, F, Chen., (2015). “Introduction to

Plasma Physics and Controlled Fusion (3rd ed.)”.

Springer Cham.

Mendis, D.A. (1979). “Dust in cosmic plasma

environments.” Astrophys Space Sci 65, 5–12

Shukla, P. K., & Mamun, A. A. (2015).

“Introduction to dusty plasma physics.”- CRC

press.

1.6 Conclusion

Ignatov, A. M. (2005). “Basics of dusty plasma.”

Plasma physics reports, 31(1), 46-56.

Shukla, P. K., & Eliasson, B. (2009).

“Colloquium: Fundamentals of dust-plasma

interactions.” Reviews of Modern Physics, 81(1),

25-44.

Lampe, M., Joyce, G., Ganguli, G., &

Gavrishchaka, V. (2000). “Interactions between

dust grains in a dusty plasma.” Physics of

Plasmas, 7(10), 3851-3861.

Thomas, H., Morﬁll, G. E., Demmel, V., Goree,

J., Feuerbacher, B., & M

ohlmann, D. (1994).

“Plasma crystal: Coulomb crystallization in a

dusty plasma.” Physical Review Letters, 73(5)

Rao, N. N., Shukla, P. K., & Yu, M. Y.

(1990). “Dust-acoustic waves in dusty plasmas.”

Planetary and space science, 38(4), 543-546.

Rosenberg, M., & Kalman, G. (1997). “Dust

acoustic waves in strongly coupled dusty

plasmas.” Physical Review E, 56(6), 7166.

10.

Merlino, R. L. (2014). “25 years of dust acoustic

waves.” Journal of Plasma Physics, 80(6), 773-

786.

About the Author

Abishek P S is a Research Scholar in

the Department of Physics, Bharata Mata College

(Autonomous) Thrikkakara, kochi. He pursues

research in the ﬁeld of Theoretical Plasma physics.

His works mainly focus on the Nonlinear Wave

Phenomenons in Space and Astrophysical Plasmas.

Part III

Biosciences

Diabetic Retinopathy- Medical Image

Processing

by Geetha Paul

airis4D, Vol.3, No.12, 2025

www.airis4d.com

1.1 Introduction

Diabetic retinopathy (DR) is a progressive eye

disease and a leading cause of vision impairment and

blindness among individuals with diabetes mellitus.

This microvascular complication primarily aﬀects the

retina, the light-sensitive tissue lining the back of

the eye, by damaging its small blood vessels due to

prolonged exposure to high blood glucose levels. DR

can develop in both type 1 and type 2 diabetes and

typically advances over the years, often without initial

symptoms. As diabetes prevalence grows globally,

diabetic retinopathy has become a major public health

concern, particularly impacting adults of working age.

The primary risk factors include the duration of diabetes,

poor glycemic control, hypertension, and dyslipidemia.

Early identiﬁcation and management, through regular

eye examinations and optimised diabetic care, are

essential to reduce the burden of visual disability caused

by this disease.

The pathogenesis of diabetic retinopathy involves

continuous injury to the retinal microvasculature.

Chronic high blood sugar levels induce metabolic

and biochemical changes, resulting in thickening of

the capillary basement membrane, loss of pericytes,

and the formation of microaneurysms. As the disease

progresses, vascular permeability increases, resulting

in ﬂuid leakage, retinal swelling (macular oedema),

and eventually ischemia due to capillary non-perfusion.

In advanced stages, marked by proliferative diabetic

retinopathy, abnormal new blood vessels proliferate

on the retinal surface in response to ischemia. These

vessels are fragile, prone to bleeding, and can precipitate

complications such as vitreous haemorrhage or retinal

detachment, ultimately risking permanent vision loss if

untreated.

Clinically, DR is classiﬁed into non-proliferative

(NPDR) and proliferative diabetic retinopathy (PDR)

stages. NPDR is characterised by microaneurysms,

haemorrhages, hard exudates, and sometimes

macular oedema, while PDR is distinguished by

neovascularisation. The onset and progression of DR

may be insidious; therefore, annual comprehensive

dilated eye examinations are recommended for all

diabetic patients. Treatment options include optimising

systemic risk factors, intravitreal pharmacotherapy (e.g.,

anti-VEGF agents), laser photocoagulation, and surgical

intervention for severe cases. Despite advances in

therapy, prevention remains the cornerstone of reducing

the impact of diabetic retinopathy by achieving strict

glycemic, blood pressure, and lipid control.

2. Non-proliferative diabetic retinopathy

Non-proliferative diabetic retinopathy (NPDR) is

the early stage of the disease in which symptoms will be

mild or nonexistent. In NPDR, the blood vessels in the

retina are weakened. Tiny bulges in the blood vessels,

called microaneurysms, may leak ﬂuid into the retina.

This leakage may lead to swelling of the macula.

3. Proliferative diabetic retinopathy

Proliferative diabetic retinopathy (PDR) is the

more advanced form of the disease. At this stage,

1.2 Drusen

Figure 1: High quality to low quality of DR Retinal

images,

Image courtsey:https://www.nature.com/articles/s41598-021-93632- 8?fromPaywallRec=false

Figure 2: (A) Illustration of the parts of the eye, (B)

Retina of the eye, (C) Drusen are yellow deposits under

the retina, (D) CNV (Choroidal Neovascularisation)

is abnormal blood vessel growth beneath the retina,

causing leakage; and CME (Cystoid Macular Oedema):

is ﬂuid accumulation causing retinal swelling.

Image courtesy: https://www.nature.com/articles/s41467-021-23458-5

circulation problems deprive the retina of its oxygen

supply. As a result, new, fragile blood vessels can

begin to grow in the retina and into the vitreous, the

gel-like ﬂuid that ﬁlls the back of the eye. The new

blood vessels may leak blood into the vitreous, causing

vision to become cloudy

. Diabetic retinopathy is a leading cause of

blindness resulting from damage to retinal blood vessels

caused by diabetes. Image processing plays a central

role in detecting, grading, and monitoring diabetic

retinopathy using retinal fundus photographs.

Figure 3: Illustration of Drusen and its formation

beneath retina.

Image courtsey: https://www.brightfocus.org/resource/why-is-my-doctor-always-talking-about-dru

sen/

1.6 Microaneurysms

1.2 Drusen

These are small yellowish deposits of cellular

debris, proteins, and lipids that accumulate between

the retina and the underlying layer called Bruch’s

membrane. Drusen are a hallmark feature, especially in

age-related macular degeneration, but can also be seen

in other retinal conditions. They appear as small spots

beneath the retina and indicate waste accumulation

resulting from dysfunction of the retinal pigment

epithelium.

1.3 CNV (Choroidal

Neovascularization):

This is the abnormal growth of new blood

vessels from the choroid layer beneath the retina into

the subretinal space. CNV is often associated with

advanced forms of macular degeneration and can cause

leakage and bleeding, leading to vision loss. It is

a serious complication that can be seen in diabetic

retinopathy as well.

1.4 CME (Cystoid Macular Edema):

This refers to swelling in the macula (central

retina) caused by ﬂuid accumulation in cyst-like spaces.

CME is a common consequence of diabetic retinopathy

that results from vascular leakage and inﬂammation. It

leads to central vision impairment and is a major cause

of vision loss in DR.

1.5 Exudates

Exudates in the context of diabetic retinopathy

are deposits composed primarily of lipids (fats) and

proteinaceous material that leak out from damaged

blood vessels in the retina. These deposits accumulate

in the outer layers of the retina and appear as yellow or

white patches on retinal images.

Figure 4: Exudates in Dibetic Retinopathic eye.

Image courtsey: https://pmc.ncbi.nlm.nih.gov/articles/PMC4568614/

Figure 5: Microaneurysms in Diabetic Retinopathic

eye

Image courtsey: https://onlinelibrary.wiley.com/doi/10.1155/2023/1305583

They are caused by the breakdown of the blood-

retina barrier due to damaged retinal capillaries,

allowing serum proteins and lipids to escape and settle

in the retinal tissue. Exudates often indicate leakage

and swelling in the retina, and their presence, especially

near the macula (central retina), can lead to signiﬁcant

visual impairment.

1.6 Microaneurysms

Microaneurysms are tiny bulges or swellings in

the walls of the small blood vessels (capillaries) of the

retina, occurring as an early sign of diabetic retinopathy.

They form when high blood sugar levels caused by

diabetes weaken the retinal blood vessels, causing them

to balloon out locally. These microaneurysms can leak

ﬂuid or blood into the retina, potentially leading to

vision problems.

1.7 Haemorrhages

Haemorrhages in diabetic retinopathy refer to

bleeding that occurs when the fragile blood vessels

in the retina break and leak blood into the retinal tissue.

They can appear in various forms depending on their

location and size:

Dot and blot haemorrhages: Small, round

haemorrhages found in the deeper layers of the retina,

typically resulting from the rupture of microaneurysms.

Flame-shaped hemorrhages: Occur in the

1.11 Treatment and Management

Figure 6: Haemorrhages in Diabetic Retinopathic eye.

Image courtsey: https://www.goodeyes.com/diabetic-retinopathy/

superﬁcial layers of the retina, resembling ﬂame shapes

along the retinal nerve ﬁber layer

Haemorrhages are a sign of worsening diabetic

retinal damage and are associated with increased

vascular permeability and vessel wall weakening. In

severe diabetic retinopathy, especially proliferative

diabetic retinopathy, new fragile blood vessels may form

and bleed into the vitreous humour, causing vitreous

haemorrhage, which can lead to vision loss. The

presence and extent of haemorrhages are important

clinical indicators used in diagnosing and staging

diabetic retinopathy.

1.8 Causes and Progression

High blood sugar from diabetes damages retinal

blood vessels, causing increased permeability and blood

vessel loss, which leads to retinal ischemia and swelling.

The retina reacts by growing abnormal new vessels that

are fragile and prone to bleeding. Vision loss occurs

particularly when macular edema develops or in the

advanced proliferative stage with retinal detachment

risks.

1.9 Symptoms

Common symptoms include blurry or cloudy

vision, ﬂoaters, dark spots or areas in the vision, and

eventual vision loss. Early stages may be asymptomatic,

making regular screening important for people with

diabetes.

1.10 Diagnosis

Diagnosis involves dilated eye exams using

ophthalmoscopy or retinal imaging to detect

microaneurysms, haemorrhages, exudates, and new

vessel growth. Regular screening is crucial since early

diabetic retinopathy can be managed before signiﬁcant

vision loss.

1.11 Treatment and Management

Treatment aims to manage diabetes (maintaining

blood sugar, blood pressure, and cholesterol control)

and directly address eye damage.

Treatments include Anti-VEGF injections to

reduce abnormal blood vessel growth and macular

swelling, Steroid injections to reduce inﬂammation

and swelling, Laser photocoagulation to seal leaking

vessels and prevent bleeding, and Vitrectomy surgery

for severe cases with vitreous haemorrhage or retinal

detachment.

Strict glycemic control and routine eye exams are

essential to prevent progression. Early detection and

intervention can prevent 90% of severe vision loss cases.

Overall, diabetic retinopathy requires a combination

of systemic disease management and targeted ocular

treatments for best outcomes

1.12 Role of Image Processing in

Diabetic Retinopathy

Retinal imaging techniques, such as fundus

photography and optical coherence tomography (OCT),

provide high-resolution images of the retina.

Image processing enhances these images through noise

reduction, contrast adjustments, and colour channel

extraction to highlight key retinal structures.

Segmentation algorithms isolate anatomical features

like blood vessels, the optic disc and detect lesions

(microaneurysms, haemorrhages, exudates), which are

signs of DR.

1.13 Automated Detection and

Classiﬁcation

Machine learning and deep learning models,

especially convolutional neural networks, analyse

processed images to automatically detect and classify

1.15 DR grading pipeline end-to-end

Figure 7: Illustration of the steps in DR -Image

processing

Image courtsey: https://www.nature.com/articles/nrdp201612

DR stages.

These systems recognise subtle changes in retinal

features associated with early and advanced DR that

might be missed in manual examination.

Automated grading and early detection improve

screening eﬃciency, allowing timely intervention to

prevent severe vision loss.

1.14 Steps in Diabetic Retinopathy

Image Processing

Preprocessing: Involves green channel extraction

for higher contrast, histogram equalisation for

brightness normalisation, resizing images, and

denoising to enhance image clarity.

Feature Extraction: Quantiﬁes biological

features such as exudates, microaneurysms,

haemorrhages, blood vessel area, and bifurcation points.

These features are critical markers of retinopathy and

its severity.

Figure 8: The local fundus image dataset was randomly

divided into training and validation datasets. Seventy

per cent of the total images in the dataset were used

for training the image quality assessment sub-network.

The lesion detection sub-network was trained using

gradable images with annotations of retinal lesions.

Then, the total gradable images in the training set

were used to train the DR grading sub-network. All

images in the local validation dataset were used to test

the image quality sub-network. Finally, the gradable

images labelled with retinal lesions were used to test the

lesion detection sub-network. DR, diabetic retinopathy.

Image courtesy:https://www.nature.com/articles/s41467-021- 23458-5

Segmentation: Localises regions of interest (optic

disc, fovea, lesions) using thresholding and edge

detection algorithms.

Classiﬁcation: Machine learning and deep

learning models (like CNNs, SVMs, and RSG-Net)

automatically detect and classify stages of diabetic

retinopathy by analysing extracted features.

Grading: Automated systems grade images into

retinopathy stages (none, mild, moderate, severe,

proliferative) to guide clinical decision-making.

1.15 DR grading pipeline end-to-end

Data Collection and Preparation

Acquire a large annotated dataset of retinal fundus

images graded by DR severity levels (e.g., from public

datasets like Kaggle’s EyePACS).

Preprocess images by resizing (commonly to 224x224

pixels), normalisation, noise reduction, and enhancing

contrast.

Utilise data augmentation techniques (such as rotation,

ﬂipping, and zooming) to enhance training data

diversity and mitigate overﬁtting.

Model Architecture and Training

Choose a robust convolutional neural network

1.15 DR grading pipeline end-to-end

(CNN) architecture such as ResNet-50 or a specialized

model like RSG-Net for feature extraction.

Apply transfer learning by ﬁne-tuning a pretrained

backbone on DR data to leverage prior knowledge.

Train the model for multiclass classiﬁcation

corresponding to DR grading stages (no DR, mild,

moderate, severe, proliferative).

Use suitable loss functions (categorical cross-entropy

for multiclass) and optimisers like Adam or SGD with

tuned learning rates.

Implement regularisation techniques like dropout and

batch normalisation to prevent overﬁtting.

Monitor validation metrics to apply early stopping or

learning rate adjustments during training.

Model Evaluation and Validation

Evaluate performance using accuracy, AUC score,

sensitivity, speciﬁcity, and confusion matrices.

Validate the model on separate test data to ensure

generalisation.

Utilise robust metrics to evaluate misclassiﬁcations

between adjacent DR grades and mitigate class

imbalance.

Pipeline Automation and Deployment

Build a workﬂow to handle data input,

preprocessing, prediction, and grading output.

Optionally integrate a user interface for uploading

images and displaying results.

Containerise the model using Docker and deploy on

cloud platforms for scalability.

Implement continuous monitoring and retraining based

on new data.

This pipeline encapsulates automated DR detection with

high accuracy and practical usability, enabling timely

diagnosis and referral for diabetic retinopathy patients.

References

https://www.nature.com/articles/s41467-021-2

3458-5

https://eyewiki.org/Age-Related Macular Degen

eration

https://www.sciencedirect.com/org/science/articl

e/pii/S1546221824007951

https://pmc.ncbi.nlm.nih.gov/articles/PMC10411

652/

https://www.health.harvard.edu/newsletter artic

https://biomedpharmajournal.org/vol10no2/dia

betic-retinal-fundus-images-preprocessing-and-featu

re-extraction-for-early-detection-of-diabetic-retinop

athy/

About the Author

Geetha Paul is one of the directors of

airis4D. She leads the Biosciences Division. Her

research interests extends from Cell & Molecular

Biology to Environmental Sciences, Odonatology, and

Aquatic Biology.

About airis4D

Artiﬁcial Intelligence Research and Intelligent Systems (airis4D) is an AI and Bio-sciences Research Centre.

The Centre aims to create new knowledge in the ﬁeld of Space Science, Astronomy, Robotics, Agri Science,

Industry, and Biodiversity to bring Progress and Plenitude to the People and the Planet.

Vision

Humanity is in the 4th Industrial Revolution era, which operates on a cyber-physical production system. Cutting-

edge research and development in science and technology to create new knowledge and skills become the key to

the new world economy. Most of the resources for this goal can be harnessed by integrating biological systems

with intelligent computing systems oﬀered by AI. The future survival of humans, animals, and the ecosystem

depends on how eﬃciently the realities and resources are responsibly used for abundance and wellness. Artiﬁcial

intelligence Research and Intelligent Systems pursue this vision and look for the best actions that ensure an

abundant environment and ecosystem for the planet and the people.

Mission Statement

The 4D in airis4D represents the mission to Dream, Design, Develop, and Deploy Knowledge with the ﬁre of

commitment and dedication towards humanity and the ecosystem.

Dream

To promote the unlimited human potential to dream the impossible.

Design

To nurture the human capacity to articulate a dream and logically realise it.

Develop

To assist the talents to materialise a design into a product, a service, a knowledge that beneﬁts the community

and the planet.

Deploy

To realise and educate humanity that a knowledge that is not deployed makes no diﬀerence by its absence.

Campus

Situated in a lush green village campus in Thelliyoor, Kerala, India, airis4D was established under the auspicious

of SEED Foundation (Susthiratha, Environment, Education Development Foundation) a not-for-proﬁt company

for promoting Education, Research. Engineering, Biology, Development, etc.

The whole campus is powered by Solar power and has a rain harvesting facility to provide suﬃcient water supply

for up to three months of drought. The computing facility in the campus is accessible from anywhere through a

dedicated optical ﬁbre internet connectivity 24×7.

There is a freshwater stream that originates from the nearby hills and ﬂows through the middle of the campus.

The campus is a noted habitat for the biodiversity of tropical Fauna and Flora. airis4D carry out periodic and

systematic water quality and species diversity surveys in the region to ensure its richness. It is our pride that the

site has consistently been environment-friendly and rich in biodiversity. airis4D is also growing fruit plants that

can feed birds and provide water bodies to survive the drought.