Deep Learning

4.1 Concept of Deep Learning

4.2 Introduction to Neural Networks
4.3 Types of Deep Learning models
4.4 Deep leaning applications
Deep learnin g is a subset of machine 4.1 Concept of Deep Learning :
learning that involves the use of artificial
The definition of Deep learning is that it
neural networ ks with multip le laye.rs
is the branch of machine learning that is
(hence "deep") to model and understand
based on artific ial neural network
complex patterns in large amounts of data.
architecture. An artificial neural network
It enables machines to learn from vast
or ANN uses layers of interconnected
datase ts by autom aticall y extrac ting
nodes called neurons that work together
features and representations, which makes
to process and learn from the input data.
it particularly effective for tasks such as
image recogn ition, natura l langua ge Deep learning is an emerging field that
processing, and speech recognition. has been in steady use since its inception
in the field in 2010. It is based on an
Deep learnin g models are designed to
artificial neural network which is nothing
learn hierarchies of features, where each
but a mimic of the working of the human
layer of the network captures increasingly
brain.
abstract representations of the input data.
This capability allows deep learning to Just like the ML model, the DL model
achieve high levels of accuracy in tasks requires a large amount of data to learn
that require the understanding of intricate and make an informed decision and is
relationships and structures within the data. therefore also considered a subset of ML.
This is one of the reason s for the
The simple machine learning algorithms
misconception that ML and DL are the
described in this chapter work very well
same. However, the DL model is based on
on a wide variety of important problems.
artificial neural networks which have the
Howev er, they have not succee ded in
capability of solving tasks which ML is
solving the central problems in AI, such
unable to solve.
as recogn izing speech or recogn izing
objects. The development of deep learning The future is driven by DL models.
was motivated in part by the failure of Without DL, Alexa, Siri, Google Voice
traditional algorithms to generalize well on Assistant, Google Translation, Self-driving
such AI tasks . . cars are not possible.

ztdf
 73
oceP Learning

Artificial lntelligen ce

achlne L

Here's a concise comparison between machine learning and deep learning: .

Feature Machine Learning Deep Learning

Definition A subset of artificial intelligence A specialized subset of machine

that enables systems to learn from learning that uses neural networks
data and improve over time with multiple layers to model
without explicit programming. complex patterns in large datasets.
Data Requiremen ts Typically requires less data to train Requires large amounts of labelled
models effectively. data for training to achieve optimal
performance.
Feature Often requires · manual feature Automatica lly extracts features
Engineering extraction and engineering by from raw data, reducing the need
domain experts. for manual intervention.
Model Complexity Generally, involves simpler models capable of capturing intricate
(e.g., decision trees, support vector relationships.
machines).
Computational Usually requires less computational Demands significant computation al
Resources power and can run on standard resources, often leveraging GPUs
hardware. for efficient processing.
Interpretability Models can be easier to interpret Models are often seen as " black
and understand, especially simpler boxes," making them harder to
algorithms. interpret due to their complexity.
Applications Used for a wide range of Commonly applied in areas like
application s like classificati on, computer vision, natural language
regression, and clustering. processing, and speech recognition.

Fund amental of Al / 2025 / IO

74 Fundamental or Al
Dl·c p lc11rnin p. i-i II fu sci 1111ti11 g orcn ol' nrtili c iul int cll1 gc11ce (Al) th at rocui;c, on
tc11chi11 gcu111p11t cr-: tn lcnm from l11rgc 111nounts of d11tn. It is II type o f machine learning,
"hk h 1tsclt" 1s II subset of A I. I he key idco hehind deep learn mg is to u<ie neural network 11 ,
"hkh nrc inspired hy how uur hrnins work

( Deep UeamlQQ

• Key Concepts activated or not, introducing non-linearity

into the model. Common examples include
l . Neural Ne\works
ReLU (Rectified Linear Unit) and sigmoid.
2. Layers
• How It Works
3 . Activation Functions To train a deep learning model, you
1. Neural Networks : These are structures typically follow these steps:
made up of layers of interconnected nodes 1. Collect Data: Gather a large dataset
(or neurons). Each layer processes data relevant to the problem you want to solve
and passes it to the next layer, helping the (e.g., images, text).
system learn complex patterns. 2. Train the Model: The model learns from
2. Layers: the data by adjusting its internal
parameters through a process ca ll ed
Input Layer: Where data enters the backpropagation, where it minimizes errors
network.
in its predictions.
Hidden Layers: These layers perform 3. Evaluate Performance: After training, the
calculations and transformations. Deep model's accuracy is tested on new. unseen
learning typically uses many hidden data to see how well it performs.
layers. 4.2 Introduction to Neural Networks
Output Layer: Produces the final Artificial Neural Networks (ANNs) are
results or predictions. computational models inspired by the way
3. Activatio11 Functions: These functions biological neural networks in the human
decide whether a neuron should be brain process information. They are a
l)ecp Learning 75
cornerstone of machine learning and artificial intelligence, enabling computers to recognize
patterns. make decisions. and learn from ·data.
Structure of ANNs

Bias
b

Activation Output

iJ = g(w • x + b)
Inputs

ANNs are composed of layers of functions:

interconnected nodes, or "neurons." The Weighted Sum: The neuron calculates a
main components include: weighted sum of its inputs. Each input is
- Input Layer : This is the first layer multiplied by a weight, which signifies its
where the network receives input data. importance.
Each node in this layer corresponds to • Activation Function: The weighted sum
a feature or attribute of the data. is passed through an activation function,
- Hidden Layers : These layers are which determines the neuron's output.
between the input and output layers. Common activ1:1tion functions include:
They perform computations and - Sigmoid: Produces an output between
transfonnations \m the input data. A 0 and I.
network can have one or more hidden ReLU (Rectified Linear Unit):
layers, which is what defines "deep Outputs the input directly if it is
learning" when multiple layers are positive; otherwise, it outputs zero.
present.
- Tanh: Produces an output between -1
- Output Layer : The final layer that and 1.
produces the output. The number of
• What is Neural Net ?
neurons in this layer depends on the
- A neural net is an artificial
type of task ( e.g., classification,
representation of the human brain that
regression).
tries to simulate its learning process.
• Neurons
An artificial neural network (ANN) is
Each neuron performs the following
often called a "Neural Network" or
76 Fun,lamtnhal of \I

Simpl) Neural Net tNN). /\rtilk1nl Nl·11rnl m·t\\ nrl-. 1, ,1 nct\,11 1~

l'ruditinnnlly. the word ncurnl netwnrl,.
or ~impk prnccs, 111 1,2. ckmcnh
(ncumns) whi,·h cun c,h1h11 crn11plc,
is rdi.:rrcd to a net\\ nrk. nr hinlogicnl
g lnh nl hch ovio11r. dctcrm111cd h,
neurons in the nervous system thnt
the connccti.ms hctwcc.:11 the procc ...... ,n!l
process nncl trnnsmit informntinn.
clements ond clement pnrnm1:1cr,
Artificial ncurnl network is nn
llere·s n comparison of ortilkml nc11r.1I
interconnec ted group of urti ficin l
networks (ANNs) nnd the humon hrain.
neurons that uses n mathcm nticn l
focusing on terminology:
model or computational llh)dcl for
information processi ng bused on n The structure of ortiticinl neurnl nct\\ ork.~
connccti oni st approac h to is inspirc.:d by bi ological neurons A
computation . biological neuron hos a cdl body or ~oma
to process the impul ses. dendrites to
The artificial neural networks are made
receive them. and on axon that transfer,
of interconnecting artiliciol neurons
them to other neurons. The input node"
which may share some properties of
of artificial neural net\-vorks receive input
biological neural networks.
signals. the hidden layer nodes compute

Feature Artificial Neural Networks (ANN) Human Brain

Neuron Artificial Neuron (Node) Biological Neuron
Connections Weights (Synaptic Weights) Synapses
Activation Activation Function Action Potential
Input Layer Input Layer Sensory Neurons
Hidden Layer Hidden Layers Complex Neural Pathways
Output Layer Output Layer Motor Neurons / Output Signals
Learning Training (Backpropagation) Leaming (Neuroplasticity)
Memory Stored Weights (Fixed) Distributed Memory (Dynamic)
Bias Bias Tenn Biological Bias (Predispositions)
Architecture Feed forward. Convolut ional, Cortical Structures (Cerebral
Recurrent Cortex, etc.)
Data Processing Batch Processing. Epochs Continuous Processing
Error Correction Loss Function. Gradient Descent Error Correction (Feedback
Mechanisms)
Parallel Processing Parallclizable (Limited) Highly Parallel (Massively Paralkl)
Decision Making Detenninistic Outputs Probabilistic (based on
experiences)
_ l_ _ _ __ __ _ __ _ _ _ _ _ __ _ _ _ _ __
,M-llat&ffl'Tm:.;.,.;;am111e111.1'fErililllFllll'KC.:WU£i,w,o. . ._ _ _ _ _ __

l>t•,•p l.c11rnln~ 77

dendrites
...
nucleus

. ~ ~ ~ ,(
~~ body 1
cell ~I
t
axon
terminals

out
f
•
In n
bias
these input signals, and the output layer nodes compute the final output by processing the
hidden layer's results using activation functions.
Biological Neuron Artificial Neuron
Dendrite Inputs
Cell nucleus or Soma Nodes
Synapses Weights
Axon Output
Synaptic plasticity Backpropagations
Here's a comparison between Biological Neural Networks (BNN) and Artificial Neural
Networks (ANN) in tabular form :
7H Fundam ental ,,r Al

Fcuturc Artificial Neural Network-,

Rlologlcol Neural Nctwor""
(ANN)
(BNN)
Structure Composed of artific ial neuron,
Composed of biologica l neurons,
synopses, nnd complex networks. (nodes) and layers.
FuncHonallty Operates through biochemical Operates th rough mathematical
processes, electrical impulses, and functions and algorithms.
neurotransmitters.
Learning Mechanism Uses various mechanisms, including Primarily uses backpropagation
synaptic plasticity and ncurogenesis. and gradient descent for training.
Flexibility Highly adaptable and capable of Less flexible; changes are made
through defined learnin g
sel f:.organization.
algorithms.
Parallel Processing Highly parallel, with millions of Can be parallelized but typically
neurons processing information Jess efficient than bi ological
simultaneously. networks.
Energy Efficiency Highly energy-efficient, operating Generally requires more
on low energy (e.g., glucose). computational power and energy.
Development Develops through a complex Developed through programming
process influenced by genetics and and training on datasets.
environment. May fail if key components
Error Handling Can function despite damage or (neurons) are damaged or
loss of neurons. misconfigured.
Speed of Learning Learning can occur continuously Learning is typically slower, often
and rapidly. requiring many iterations.
Memory and Recall Utilizes a distributed and dynamic Uses fixed weights for memory,
memory system. which can be less adaptable.
Application Found in biological organisms, Used in various applications like
responsible for cognition and image recognition, NLP, and more.
behavior.
• McCulloch-Pitts Neuron Basic Model
The McCulloch-Pitts neuron is one of the earliest models of artificial neurons I introduced
by Warren McCulloch and Walter Pitts in 1943. This model laid the foundation for later
developments in neural networks and artificial intelligence. Here's a detailed explanation
of its structure and functionality.
I. Structure
2. Functionality
3. Logic Gates
4. Significance
5. Limitations
6. Conclusion
 Ot't' P I ,,11 r n h1,t 7CJ
1.
------------- -----------
St ru c Ill rt•
Input signals
01,H,

/\c.tlvn l lon
fun ction
~ ut
q,(. ) i----Yi
Summing junctio n
I> W k4

Synaptic Welnhts
rh..: McC ulloch-Pitts neuron consists of I lcrc, wrcprescnts the weights, x,thc
several key components: Input values, ond the sum is taken over
Inputs : The model accepts multiple oil inputs.
binury inputs (0 or l ), which represent - Binary Nnture: The model operates in
signals from other neurons or sensory a/binary fashion, meaning it can only
inputs. produce o utputs of O or I. T hi s
Weights : Each input is ussociatcd characteristic aligns with the idea of
with a weight, although in the originul simple decision-making processes.
mode l, these weights arc impl icitly 3. Logic Gates
considered as being equal (typically The McCulloch-Pitts neuron can be used
binary: either present or absent). to model basic logical fun ctions. By
C:
Summation : The inputs are summed configuring the inputs and thresholds, it
together, resulting in a total activation can represent various logic gates:
value. - AND Gate: Re4uircs all inputs to be
.
Threshold : A threshold value is
I to produce an output of I .
defin ed. Jf the summed inputs exceed - OR Gnte: Requires at least one input
this threshold, the neuron "fires" and to be I to produce an output of 1.
produces an output of I; otherwise, the - NOT Gntc: Inverts the input; if the
output is 0. input is 0, the output is I, and vice
2. Functionality versa.
The McC ull och-Pitts neuron operates 4. SIGNIFICANCE
based on the following logic: - Foundutlon of Neural Networks :
- Activation Function: The output y is The McCulloch-Pitts model was pivotal
determined by the equation: in establishing the concept of neurons
us compututionul units, inlluencing later
y - {I if l:( w, . x1 ) threshold}
~.
development s in urti fici ul neural
() otherwi se networks.
 Fundamental of Al
80
6. Conclusion :
- Theoretical Framework : It provided
The McCulloch-Pitts neuron is a funda-
a simple framework to understand how
mental concept in the history of artifi cial
networks of neurons could be used to
intelligence and neural netwo rk s. By
perform complex computations.
simulating basic logical operati ons and
5. Limitations introducing the idea of neuron s as
While the McCulloch-Pitts neuron was computational units, it paved the way for
ground-breaking, it has several limitations: more advanced models and techniques in
- Binary Inputs and Outputs : The the field . Despite its s impli city, th e
binary nature restricts its ability to principles established by thi s model
model more complex, continuous-valued continue to influence modern neural
functions. network d~signs.
- Static Weights : The original model • Types of Artificial Neural Networks
does not incorporate learning or weight Artificial Neural Networks (ANNs) come
adjustment mechanisms. in various architectures, each suited to
- Lack of Complexity : It cannot speci fie tasks and data types. Below is a
represent more complex functions detailed overview of the most common
requiring non-linear combinations of types of ANNs, their structures, and their
inputs. applications.

1. Feedforward Neural Networks (FNN)

outputs
inputs

input layer hidden layer output layer

• Structure: • Key Features:

Composed of an input layer, one or Architecture: Each neuron in one
more hidden layers, and an output layer is connected to every neuron in
layer. the next layer (fully connected).
Data flows in one direction- from the Activation Functions: Common
input layer to the output layer- activation functions include Sigmoid.
without any cycles or loops. ReLU (Rectified Linear Unit), and
Tanh.
J)ccr Lcnrnin~ R1
• Applications:
Used for tasks . . ·such as c Iass1· fi1cat1on.
·
regression. and pattern rccognit1on.
Common 111 simple applications t·k d. · ·· · · I ·ri ·
- -=-:;---:-;-:--;:-;-::-:~ ~ : - - - - - - - --- - 1 ....:c...., 1g1t rccog111t1on and basic image c ass1 1cat1on.
t. Feedback Networks
2. Recurrent Neura l Netw orks (RNN)
2. Feedback Networks :

outputs
inputs con11>etition/in hi bit ion

feedback

• Structure: - Dynamic State: Capabl e of

Also known as recurrent network s. maintaining a state or memory, which
feedback networks allow connections all ows the netwo rk to process
between neurons in a way that feedback sequences of inputs.
is possible (i.e ., a neuron can send • Applications:
information back to previous layers). Commonly used in systems requmng
• Key Features : continuous feedback, such as control
- Feedback Loops: Neuron s can systems and certain adaptive systems.
influence their own future states by - Limited usage compared to more
sending outputs back to earlier layers. advanced recurrent networks.

3. Recurrent Neural
Networks (RNN)

I
' -- --- ----
Input L.ayer
-----
Hidden Layer Output L.ayer
neurons x, neurons YI neurons :Zk

Fu n<1amcntal or Al / 2025 / 11
 fg & ' ,_, w
◄
1 C •

Ml Fundum cnhel of Al

• Strnrtun•: I. lmn~ e nncl Video Proccsslnl( :

- lmn ~c C ln ~" lfl cn tlon ANN '!,
RNN s nrc dc~ ig 11e d lo hnncll c
~cqucnt lnl dnt11 by inco rporn ting cycles p11r tic ulnr ly Con volu lio nal Ncurul
m the network . Network'! (CNNs), urc cxtcrn1ive ly 1111cd
for clas'li ly ing imoges in cntcgoric,
Ench neuron cnn send outputs buck to ohje ctc; rn
(e .g., ide ntify in g
itsel f or oth er neuron s in previous
photogrnphs) .
lnyc rs, crcn ting a fcedhnek loop.
• Key Features: - Obj ect Detection : ANN s can detect
and loca te objects with in image<i or
Temporal Dep endencies: RNN s cnn vi deos , used in ap pl ica lio ns like
captu re depe ndencies in seq uenc es, auto nomous vehi cles and surveillance
mnking them suitnhle for tnsks where syste ms.
cont ext is ossc ntiul.
- lmn gc Segm entation : CNN s arc
Variations : Includes Long Short-Term em ploy ed to segm ent imag es into
Memory (LST M) networks and Gated di ffcrc nt parts for tasks such as medical
Recu rrent Units (G RU) to add ress imag ing analysis.
issues like vani shin g grad ients .
2. Natu ral Lnngungc Processing (NLP)
• Applications :
- Text Classification: ANNs are used to
Wi de ly used in natu ra l lang uage class ify lext into predefined categories
processing (e.g., lang uage modellin g, (e.g., spam detection in emails).
text generatio n), speech rccognili on,
and time series analysis. - Sentiment Annlysls: Analyzing socia l
medi a or prod uct reviews to determine
Effec tive in tasks invo lving sequences, the senl ime nt (pos ilive, negative,
such as pred icti ng the next word in a neut ral).
sente nce or gene rating music.
- Machine Tran slation: Neural networks
Artificial Neural Networks (ANNs) have arc used in trans latio n syste ms (e.g..
a wide range of appl icatio ns across various Google Translate) to convert text from
lield s due to their abi lity to learn from one lang uage to another.
data and reco gnize patte rns. I !ere arc some
of the key applicati ons: - Chatbot s and Virtual Assistants:
ANNs power conversat ional agents that
I . Image and Video Processin g :
understand and respond to user queries.
2. Natu ral Language Processin g (NI ,P)
3. Speech Recognition 3. Speech Recognition :
4. Healthcare - Voic e A11s lstnnts : A NNs enable
5. Financial Services systems like Siri and Alexa to under-
6. Autonomous Systems stand und process spoken com mands ·
7. Rcco mme rn.l,'11 ion Syslems - Tran scription Services : Convertin g
8. Time Series Prcd ic1ion spoken language into writt en text for
9. Anomaly Dcteclion appl icati ons in medi co! transcription or
10. Oaming uu1o matcd cu<;tomcr service.
--
l)ccp Learnin g

4. Henlthcarc
Medica l Diagno sis: ANNs assist in
8. Time Series Prediction
· Stock Price Predic tion : AN Ns
83

diagno sing diseas es by analyz ing

analyze historical stock data to forecas t
medical images (e.g., X-rays, MRis) or
future price movem ents.
patient data.
· Weather Forecasting : Used in models
. Drug Discov ery: Used to predict how
that predict weathe r patterns based on
different compou nds will interact with
historic a l climate data.
biologi cal system s, speedin g up the
drug develop ment process . 9. Anoma ly Detection :
Person alized Medici ne: ANNs help Netwo rk Securi ty : ANNs mon itor
tailor treatme nts based on individual networ k tra ffic to identi fy unus ual
patient data and genetic profiles. pattern s that may indicat e security
s. Financial Service s : breaches.

Fraud Detect ion : ANNs analyz e - Manu factur ing : U sed to d etect
transac tion pattern s to identif y defects or anomalies in products during
potenti ally fraudul ent activities in real- the produc tion process .
time. 10. Gamin g :
. Credit Scorin g : Assess ing the - Game AI : ANNs are employ ed to
creditw orthine ss of individ uals by create intelligent agents that can adapt
analyzi ng historic al data. to players ' strategies in real-tim e.
• Algorithmic Tradin g : ANNs are used - Proced ural Conte nt Gener ation :
to develop trading strategies based on Used to generate game environ ments or
market data. levels based on learned pattern s.
6. Autono mous System s : 4.3 Types of Deep Learni ng Models
· Self-D riving Cars : ANNs process Deep learning encomp asses a variety of
sensor data to unders tand the model types, each suited for specific tasks
enviro nment and make driving and data types. Here's a detailed overvie w
decisions. of the main types of deep learning models :
- Roboti cs : Used in robotic systems for 1. Feedforward Neural Networ ks (FNNs )
naviga tion, manip ulation , and 2. Convolutional Neural Networ ks (CNNs )
interact ion with objects . 3. Recurrent Neural Networ ks (RNNs )
7. Recom menda tion System s 4. Generative Adversarial Networ ks (GANs )
· E-com merce : ANNs analyz e user 5. Autoen coders
behavior and preferences to recommend
6. Transformers
produc ts ( e.g., Amazo n, Netflix ).
7. Graph Neural Networ ks (GNNs )
· Conten t Recom menda tion : Used by
platform s like YouTu be and Spotify to 8. Deep Reinfor cement Leamin g Models
suggest videos or music based on user 1. Feedforward Neural Networks (FNNs )
interests. Struct ure : The s imples t ty pe o f
IU l· uncfamrn1 .. 1 ,,t Al

ncur.il r\clwurk where data llow'i 111 one d1rcl.llon lrom 1npu1 11, 1>u1pu1 w1th,,ul l) c.ll:
or loop,
u~age : l omnu,nl y used tor hJ'iH.: cltt \'il licat,on and rcgrc'i'ito n ta"lkfi. 'iuch a'i pred1u1ng
hou, ing prn.c:, or diJ'.','tl I> mg 1111,1µc-,

r\ 0 . D
, J
.. 0 ,.
.,.
,,.
~

(J / ()
0 .. ..
0 ,..,. _,,., .(_)
0
Input layer
0 D output layer

hidden laye r I hidden laye r 2

2. Convol ul ional Neural Networks (CNNs) Key Components

Structure : 01.:, 1gncd , pec,ficall> lor Convolutional Layers : Apply filters
proc.c,, ing f! rid - like data. such a"> to input data Lo ex tract features like
image, CNN<i use convolutional layers edges or textures.
to automat ica ll y detect c; patial Pooling Layers : Reduce the dimcn•
hierarchies tn data. sionality of the data while retaining
important features.

Fully
Connec ted
Co nvo lution
0 ~.
Poo ling Output
Input

LJ
·o ·-
~---- - - - - - - - y - - - ~ ) \...____y____)
Feature Extract ion Classifi cation
 Lcurninf,! HS
occP
~Usngc : Primarily used in computer Key Variants :
v is ion tas k s . s uc h as ima g e
Long Short-Term Memory (LSTM):
recog niti o n. o bj e ct de te cti o n, a nd
Addresses th e v a ni s h ing g rn d ic n l
segmentation.
prob lem, enabl ing better lea rn ing of
Recurrent Neural Networks (RNNs)
3. long-range dependenc ies.
Structure: Des ig ned for sequen tial
data. RNNs have loops allowing them Gated Recurrent Unit (GRU): A
to maintain a mem ory of prev ious simplified vers ion o f LSTM tha t is
in uts. often faster and requ ires less memory.

Recurrent network

output layer
input layer \ (class/target)
y
hidden layers : "deep" if> 1

Usage: Ideal for time series analysis, natural language processing (NLP), and speech
recognition.
4. Generative Adversarial Networks (GANs)
Structure: Comprises two neural networks- the generator and the discriminator- that
compete against each other.

H,gh
O,o nens,o nal
Sam ple
um
lf Jl
Space Real
·l lmuges
D 1scn min;:it 1w •
Re..)I

Nctw orh. 0
Lovr.1 Gener ative
D l_,
lJ1mer)s1on.:il N etwork Generated Fa ... e
Laten t Fake Images J
~pace G
 ,
86 Fundamental of AI

Functioning: The generator creates effectively.

fake data, while the discriminator Key Components : Multi-head
evaluates the authenticity of the data. attention, feedforward neural networks,
Over time, both networks improve, and positional encoding.
leading to the generation of highly Usage : Dominant in NLP tasks,
realistic data. including translation, summarizati on,
Usage : Used for image generation, and language modeling (e.g., BERT,
v ideo synthesis, and data GPT).
augmentation. Graph Neural Networks (GNNs)
7.
5. Autoencoders : Structure : Designed to work with
Structure : Consists of an encoder data structured as graphs (nodes and
that compresses input data into a edges), GNNs capture relationships
lower-dimensional representation and a between entities.
decoder that reconstructs the original
Usage : Applied in social network
data from this representation.
analysis, recommendation systems, and
• Key Variants : molecular chemistry for predicting
Denoising Autoencoders : Trained to molecular properties.
reconstruct clean data from noisy Deep Reinforcement Learning Models
8,
inputs.
Structure : Combines deep learning
Variational Autoencoders (VAEs):
with reinforcement learning principles,
Incorporate probabilistic elements,
where agents learn to make decisions
allowing for generative tasks.
by interacting with an environment.
Usage : Employed in tasks like data
Usage: Widely used in game AI (e.g.,
compression, image denoising, and
AlphaGo), robotics, an~ autonomous
anomaly detection.
systems, where agents learn optimal
6. Transformers : policies through trial and error.
Structure : Utilizes self-attention
Here's a comparison of Convolutional
mechanisms to weigh the importance
Neural Networks (CNNs) and Recurrent
of different parts of the input data,
Neural Networks (RNNs) in tabular form:
allowing for parallel processing and
handling long-range dependencies
 Le11rninJ? 87
oet'fl

Feature Convolutional Neurnl Networks Recurrent Neural Networks

(CNNs) (RNNs)

-----
purpose Primarily used for spati a l data
analysis (e.g., images).
Designed for sequential data and
time-series analysis.
_Architecture Consi sts of convoluti onal and Composed o r recurrent layers with
pooling layers, followed by fully loops to maintain memory o f
connected layers. previous inputs.
oatn Handling Processes data with a fixed input Handles variable-len g th input
size, suitable for grid-I ike sequences, ma intnining co nte x t
structures. across time steps.
Feature Extraction Learns temporal dependencies and
Automatically learns spatial relationships in sequences.
hierarchies and patterns through
Memory convolutions. Maintains hidden s tates to
Lacks memory of previous inputs remember previous inputs ,
once the input is processed. allowing for context.
Training Time Slower training due to sequential
Generally faster to train due to nature; training must occur step-
parallel processing capabilities. by-step.
Applications Natural language processing, time
Image classification, object series forecasting, and speech
detectiun, and video analysis. recognition.

Aspect GNNs CNNs RNNs

Input type Graphs (nodes, edges, Grid-like data (e.g. Sequential data (e .g .
features) images) time series)
Information Flow Propagates information Local receptive fields in Information passed
across nodes convolution sequentially
Architecture Message passing, node Hierarchical layers of Sequential layers of
update convolutions neurons
Memory of past Incorporates global Captures local patterns Captures temporal
data graph structure in the grid dependencies
Applications Social networks , Image recognition , Natural language
molecular structures computer vision processing, speech
Training Moderate complexity Complex, numerous Complex due to
complexity due to graphs layers, large data sequential dependencies
Parallel Limited due to graph High due to paral lei L imited due to
Processing structure convolutions sequential nature
Data size Sensitive to graph size Less sensitive, scales Sensitive to sequence
tolerance and structure with data length
H8 Fundamcnt.il or Al

4.4 Deep Learning Applicati ons : no ise. leadin g to appl icat ions like art
generation. deepJ'akcs, and vi rtual n:tili ty
Deer learning has tran sformed many
environments.
industries by enabling sophisticated data
analysis and decision-making. 2. Natural Language Processing (NLP)
I. Computer Vision • Sentimen t Analysis :
2. Natural Language Processin g (NLP) Analyzing text data to determine sentiment
3. Healthcare (positive, negative, neutral). Businesses use
thi s to gauge publi c opinion about
4. Autonomous Vehicles products or services.
5. Finance
• Machine Translation :
6. Gaming and Entertainment Deep learning models have significant ly
7. Manufacturing and Industry improved translation services. as seen in
8. Agriculture tools like Google Translate , enabling
1. Compute r Visio~ accurate translatio ns between multiple
languages.
• I magc Recognition:
• Chatbots and Conversa tional Agents :
- Deep learning models, particula rly
Convolu tional Neural Network s NLP models ' power chatbots that can
(CNNs), excel in identifying objects understan d and respond to customer
within images. Applications include : queries in real-time, enhancing customer
service across various platforms.
- Facial Recogniti on : Used in security
• Text Generatio n :
systems and social media platforms for
tagging and identification. Models like GPT (Generative Pre-trained
Transform er) can create coherent text,
- Medical Imaging : Analyzing X-rays,
useful for content creation, summarization,
MRis, and CT scans to detect diseases
and even coding assistance.
such as cancer, enabling earlier and
more accurate diagnoses . 3. Healthcar e :
• Object Detection : • Medical Diagnosis
Detecting and classifying multiple objects Deep learning algorithms analyze medical
within an image. This technology is vital images and patient data to identify
111 : conditions such as diabetic retinopathy or
- Autonom ous Vehicles: Identifyi ng pneumon ia. They assist doctors by
pedestria ns, traflic signs, and other providin g second op1111ons and
vehicles. highlighting areas of concern.
- SurveiJla nce Systems: Monitori ng • Drug Discovery :
public spaces for safety and security. Deep learning models predict ho\','
• Image Generation : different compoun ds might interact.
speeding up the process of finding ne,,
Generative Adversarial Networks (GANs)
drugs and treatments.
can create realistic images from random
 11 _ _ _ _ _ __ __ _89
_
t)t•rP 1' ,, r,,1-..i:_n:_.:_ _ _ _ _ _ _ _ _ _ _ _ _ __ _ _ _ _ _ _ __
-
- -prt·sonu • .. Mt•dicine :
1·z,.,t
6. Gnming nnd Entcrtninmcnt
• 8 , unnl y1 ing geneti c informntion nnd • Gnmc Al :
h~n lth records. d ee p lenrning cnn help
trcntmcnts to individual putienls. De ep lea rnin g e nh a n ces n o n- pl aye r
Ill ·1or
I
improving outcomes. character (NPC) behav iours. making them
more realisti c and responsive lo playe r
Autonomous Vehicles
nclions.
J. Deep lenrnin g is cruc in l for the
development o r sdf-driving cors. enabling • Content Creation:
them to: Dee p learning al go rithms arc used lo
Perceive the Environment : Us ing generate music. orl, and stories, providing
sen sors and cam e ra s to identify n ew w ays for c re a tors to ex pre ss
objects. road conditions. and obstacles. themselves.
Path Planning : Determining the best 7. Manufacturing and Industry
route while avoiding collisions and • Predictive Maintenance:
optimizing for time or distance.
By analyzing data from machinery, deep
Renl-Timc Decision Making : Making learning predicts when equipment is likely
split-s econd decisions based on to fail, allowing for timely ma intenance
dynamic data from the environment. and reducing downtime.
s. Finance • Quality Control :
Fraud Detection
Deep learning systems can inspect
Deep learning models analyze transaction products on production lines to identify
patterns to detect anomalies indicative of defects, ensuring high-quality standards are
fraudulent activity. helping banks and maintained.
financial institutions protect customers.
8. Agriculture
Algorithmic Trading:
• Crop Monitoring
Traders use deep learning to predict stock
Using drone imagery and deep learning
prices and execute trades based on vast
algorithms, farmers can monitor crop
amounts of financial data, identifying
health, detect diseases , and optimize
trends and making decisions in real-time.
yields.
Credit Scoring:
• Precision Agriculture :
By evaluating a wide array of financial
Deep learning models analy ze environ-
data, deep learning helps assess credit risk
mental data to inform decisions on
more accurately, improving lending
planting, watering, and harvesting. leading
processes.
to more efficient resource use.

F1ind1mcntal or Al / 2025 / 12