AI Basics Course link is in description box

Using AI Is Easy. Understanding Its Technology Is Power

AI Is More Than Tools: It’s Algorithms, Models, and Intelligence
Higher Salaries Come from Deeper AI
Knowledge - Not Just Tools
 What You Will Learn in This Course
LEVEL 1: AI FOUNDATIONS LEVEL 2: MACHINE LEARNING & DEEP LEARNING CONCEPTS
• Tokenization • CNN (Convolutional Neural Network)
• Embeddings • RNN (Recurrent Neural Network)
• Inference • Transformers
• Neural Network • Self-Supervised Learning
• MCP (Model Context Protocol) • Reinforcement Learning (RL)
• AutoML
• Model Compression (Pruning & Quantization)
 What You Will Learn in This Course
LEVEL 3: AI MODELING & GENERATIVE AI LEVEL 4: INFORMATION RETRIEVAL &
• LLM (Large Language Model) KNOWLEDGE SYSTEMS
• Fine-Tuning • Vector Database
• LORA (Low-Rank Adaptation) • RAG (Retrieval Augmented Generation)
• GAN (Generative Adversarial Network) • Knowledge Graphs
• Diffusion Models • Explainable AI (XAI)
• Multimodal AI • Federated Learning
• Vision-Language Models (VLMs)
• Synthetic Data
What You Will Learn in This Course
LEVEL 5: AUTONOMOUS AI & FUTURE TECH (Cutting-Edge)
• AI Agent
• Edge AI
• Zero-Shot Learning
• AGI (Artificial General Intelligence)
• ASI (Artificial Super Intelligence)
Let’s Begin the Advanced AI Journey

LEVEL 1: AI FOUNDATIONS
• Tokenization
• Embeddings
• Inference
• Neural Network
• MCP (Model Context Protocol)
 LEVEL 1: AI FOUNDATIONS
1 Tokenization in AI

Tokenization is the process of breaking a long piece of text into

smaller units called tokens so that an AI model can understand,
analyze, and process the text.
 LEVEL 1: AI FOUNDATIONS
1 Tokenization in AI

AI systems cannot read long sentences the way humans do.

They require text to be converted into small, structured pieces →
tokens → which can later be converted into numbers.

Token kya hota hai?

Token ka simple matlab:
Words ka small piece
Ek pura word
Ek character
Ya word ke parts (subwords)
 LEVEL 1: AI FOUNDATIONS
1 Tokenization in AI
Sentence:

"Artificial Intelligence is powerful."

Step-by-step tokenization:
A. Word-level tokenization

["Artificial", "Intelligence", "is", "powerful"]

B. Character-level tokenization

["A","r","t","i","f","i","c","i","a","l", ...]

C. Subword-level tokenization (LLMs me mostly use hota hai)

["Art", "ificial", "Intelli", "gence", "is", "power", "ful"]

Isme AI complex words ko small meaningful pieces me todta hai.

Kyuki agar kal ko koi naye word aa jaye jo model ne training me
nahi dekha, tab bhi wo word ke sub-parts ko samajh sakta hai.
 LEVEL 1: AI FOUNDATIONS
1 Tokenization in AI
Tokenization kyu zaroori hai?

Tokenization ke bina AI:

Text ka structure nahi samajh payega

Word boundaries nahi pehchan payega
Similar words ka relationship track nahi kar sakta
Efficient processing possible nahi hoti

Tokenization AI ko yeh samajhne me help karta hai:

Sentence me kaunsi cheez important hai

Kaunsa word kya meaning rakhta hai
Word ke similar parts kaise behave karte hain
Long text ko small manageable parts me kaise break kare
 LEVEL 1: AI FOUNDATIONS
1 Tokenization in AI

AI models tokens ko numbers me

kaise convert karte hai?

Har token ka ek unique ID (number)

hota hai.

Example:
"Data" → 1432
"Sci" → 987
"ence" → 654
"." → 29

Phir AI in numbers ko vectors me convert karta hai — ise bolte hain Embedding.
 Image Tokenization
AI model image ko seedha ek photo ki tarah samajh nahi sakta.
AI ko image ko math form (numbers) me convert karna hota hai.
Ye numbers banane ke liye image ko tokens me todna padta hai.

Image Tokenization kaise hota hai?

Step 1: Image → Patches (Chhote squares)

Example:
Ek 256×256 pixel ki image ko 16×16 pixel ke patches me tod diya.

256/16 = 16 patches per row

Total patches: 16 × 16 = 256 patches
Ye chhote patches AI ke liye "words" ki tarah hote hain.

Step 2: Patches → Numbers (Flattening)

Har patch ko flatten karke ek vector banaya jata hai.

Step 3: Vector → Token

Ye vector embedding layer se pass hota hai aur token ban jata hai.

The Result: The image is now a sequence of vectors. To the AI, the image looks just like a
sentence, where every 16 X 16 patch is equivalent to one "word."
Image Tokenization kaise hota hai? Image Tokenization
Step 1: Image → Patches (Chhote squares)

Example:
Ek 256×256 pixel ki image ko 16×16 pixel ke patches
me tod diya.

Step 2: Patches → Numbers (Flattening)

Step 3: Vector → Token

The Result: The image is now a sequence of vectors.

To the AI, the image looks just like a sentence, where
every 16 X 16 patch is equivalent to one "word.“

Advanced Method (VQ-VAE) - VQ-VAE stands for

Vector Quantized Variational AutoEncoder.
For generating images, AI often uses a "Visual
Codebook."9 It compares each patch to a dictionary
of known visual patterns (e.g., "edge," "blue curve")
and assigns it a specific ID number.10 The image
becomes a list of IDs, like [452, 112, 889...]

"DALL-E 1" used this exact VQ-VAE method to learn

how to generate images from text for the first time
 Audio Tokenization
Audio ek continuous wave hoti hai.
AI continuous signal ko directly samajh nahi sakta,
isliye isko small parts (frames) me todta hai — yehi hota hai audio tokenization.

Audio Tokenization kaise hota hai?

Step 1: Audio → Frames (10–25 ms segments)

1 second audio = approx 100–200 frames

Each frame = 10ms ya 20ms ka sound piece

Step 2: Frames → Spectrogram

Har frame ka frequency & amplitude calculate karke spectrogram banaya jata hai.
Ye ek image jaisa hota hai (mel-spectrogram).

Step 3: Spectrogram → Audio Tokens

AI is spectrogram ko small grid patches me todkar tokens banata hai, bilkul image
ki tarah.

Example

“I love data science.”

AI isko token me break karega:

Small audio frames → spectrogram patches → tokens → numbers → model.
 Video Tokenization
Video = Images (frames) + Audio + Motion
Isliye video tokenization 3 parts me hota hai.

Video Tokenization kaise hota hai?

Step 1: Video → Frames

Ek 1-second video me 24–60 frames hoti hain.

Har frame ek image hai → isko image tokenization se patches me
todte hain.
Example:
1 sec, 30 fps → 30 images → each image 256 patches → total
7680 tokens

Step 2: Motion Tokenization

Frames ke beech difference (movement) detect hota hai:
Optical flow
Motion vectors
Temporal patches
Ye motion tokens ban jaate hain.

Step 3: Audio Tokenization

Video ke audio part ka tokenization bhi hota hai using spectrograms
 LEVEL 1: AI FOUNDATIONS
2 Embedding in AI

Agar simple shabdon mein

kahein, toh Embedding kisi bhi
data (text, image, audio) ko
numbers ki ek list (vector)
mein convert karne ka tareeka
hai, taaki computer uske
meaning ko samajh sake.
 LEVEL 1: AI FOUNDATIONS
2 Embedding in AI
Embedding Kya Hai?

Imagine karo aapko computer ko samjhana hai ki "Apple" aur

"Banana" similar hain (dono fruit hain), lekin "Apple" aur "Car" alag
hain.

Agar aap sirf ID doge (e.g., Apple=1, Banana=2, Car=3), toh

computer ko lagega 2 aur 3 (Banana aur Car) zyada close hain, jo ki
galat hai.

Embedding har word ko ek lambi list of numbers (coordinates) mein

convert kar deta hai, jise Vector kehte [Link]:

Apple: [0.9, 0.1, 0.05, ...]

Banana: [0.88, 0.12, 0.04, ...] (Apple ke numbers ke bahut paas)
Car: [0.1, 0.8, 0.9, ...] (Bilkul alag numbers)

Isse machine ko pata chalta hai ki Apple aur Banana ka "meaning"

paas-paas hai.
 2 Embedding in AI
Yeh Kaam Kaise Karta Hai?

Embedding ko aap ek Map (Naksha) ki tarah samajh sakte ho.

Is map par har word ka ek address (coordinates) hota hai.

Jo words meaning mein same hote hain, wo map par paas-paas hote hain
(Jaise: Happy aur Joy).

Jo words alag hote hain, wo door hote hain (Jaise: Happy aur Sad).

Vector: Numbers ki wo list jo data ko represent karti hai. Example: [0.12, -0.45,
0.88].

Dimensions: Vector kitna lamba hai. OpenAI ke models mein aksar 1536
dimensions (numbers) hote hain ek word/sentence ke liye. Jitne zyada
dimensions, utni zyada details (meaning) capture hogi.

Cosine Similarity: Yeh ek formula hai jisse hum measure karte hain ki do
vectors (embeddings) kitne close hain. Agar score 1 hai toh bilkul same, 0 hai
toh koi connection nahi.
 2 Embedding in AI
Embeddings ke Types
Mainly do tarah ki situations hoti hain:

A. Word Embeddings (Purane Models) Isme har word ka ek

fixed vector hota tha.

Problem: Word "Bank" ka matlab river bank bhi ho sakta hai

aur money bank bhi. Purane models (jaise Word2Vec )
confuse ho jate the kyunki unke paas "Bank" ke liye ek hi
vector tha.

B. Contextual Embeddings (Modern Models - LLMs) Aajkal ke

AI (jaise GPT-4, BERT) Contextual Embedding use karte hain.

Yeh poore sentence ko padhte hain. Agar sentence hai "I went
to the bank to deposit money", toh yahan "bank" ka
embedding financial institution wala banega.

Isliye aaj kal ke AI human-like samajh rakhte hain.

 2 Embedding in AI
Real World Use Cases

Semantic Search (Google Search): Agar aap search karo "shoe for
running", toh result mein "jogging sneakers" bhi aayega. Bhale hi word
match nahi hua, par embedding ne pakad liya ki running aur jogging
ka meaning same hai.

Recommendation Systems (Netflix/YouTube): Agar aapne ek sci-fi

movie dekhi, toh system us movie ka embedding dekhega aur vector
space mein uske paas wali doosri movies recommend karega.

RAG (Retrieval-Augmented Generation): Aajkal Custom Chatbots

banane mein iska bahut use hota hai. Aap apne business ka data
embeddings mein convert karke database (Vector Database) mein
daal dete ho. Jab user sawal puchta hai, AI relevant data dhundh kar
answer deta hai.

Sentiment Analysis: Review positive hai ya negative, yeh samajhne ke

liye text ko numbers mein badla jata hai.
 LEVEL 1: AI FOUNDATIONS
3 Inference in AI
AI me do main phases hote hain:

Training Phase
Model ko data + correct answers (labels) diye jaate hain
Model patterns seekhta hai

Example:
10,000 images → “ye dog hai”, “ye cat hai”
Model sikhta hai: dog kaisa dikhta hai

Inference Phase
Ab model seekhna band karta hai
Naye data par answer deta hai
Example:
New image → model bolta hai: “95% probability ye dog hai”
Real-world me hum mostly inference hi use karte hain.
LEVEL 1: AI FOUNDATIONS
3 Inference in AI
Inference ka simple real-life example

Doctor Example
Doctor ne years tak padhai ki → Training
Patient aaya → symptoms dekhe → diagnosis bataya
→ Inference

AI Example
Model ne lakhon examples dekhe → Training
User ne input diya → model ne output diya → Inference
 LEVEL 1: AI FOUNDATIONS
4 Neural Network in AI
The Biological Neuron (Human Brain)
This is the physical cell in your head.

Dendrites (The Receivers): These are tree-like

branches that receive chemical signals from other
neurons. They are the "Input" collectors.

Cell Body / Nucleus (The Processor): This is the

center of the cell. It collects all the signals coming
from the dendrites and decides if the signal is
strong enough to pass forward.

Axon (The Cable): If the signal is strong, it travels

down this long tail. It is the transmission line.

Synapse (The Connection): This is the gap

between the axon of one neuron and the dendrite
of the next. The strength of this connection
determines how much signal gets through.
 LEVEL 1: AI FOUNDATIONS
4 Neural Network in AI
The Artificial Neuron (AI Model)
This is the mathematical code in the computer.

Inputs : These replace the Dendrites. This is the

data entering the system (like pixels of an image).

Weights : These replace the Synapse. Every

input has a "weight" (importance). If a weight is
high, that input is treated as very important.

Summation & Activation Function: These replace

the Cell Body.

Summation : The computer adds up all the inputs

multiplied by their weights.

Activation Function: This is a mathematical rule

(like a threshold) that decides: "Is this number big
enough to fire?" If yes, it passes the result
forward.

Output : This replaces the Axon. It is the final

result sent to the next layer of the network.
 LEVEL 1: AI FOUNDATIONS
4 Neural Network in AI
Human Brain (Biology) AI Neural Network (Math) Function (Kaam kya hai?)

Dendrites Inputs ($x$) Data receive karna (Receive signals).

Connection ki strength (How

Synapse Weights ($w$)
important is the signal?).

Saare signals ko jodna (Calculate

Cell Body (Soma) Summation ($\sum$)
total input).

Decide karna ki signal aage bhejna

Threshold / Firing Activation Function
hai ya nahi (Decision making).

Result ko aage transfer karna

Axon Output ($y$)
(Transmit result).
 LEVEL 1: AI FOUNDATIONS
5 MCP (Model Context Protocol) in AI
Model Context Protocol (MCP) AI ki duniya mein ek naya aur revolutionary open standard hai. Ise simple bhaasha mein
samjhein toh yeh AI models ke liye "USB port" ki tarah hai.

Jaise USB cable se aap mouse, keyboard, ya printer kisi bhi computer mein laga sakte hain, waise hi MCP ke through AI
models kisi bhi data source ya tool se connect ho sakte hain.
 LEVEL 1: AI FOUNDATIONS
5 MCP (Model Context Protocol) in AI
MCP (Model Context Protocol) ek open standard hai jise Anthropic ne introduce kiya hai.
Iska main kaam hai AI assistants (jaise Claude, ChatGPT) ko system ke data (files,
servers, databases) ke saath secure aur standardized tarike se connect karna.

Example:Pehle agar aapko Claude ko Google Drive, Slack, aur GitHub se connect karna
hota, toh developers ko har tool ke liye alag-alag integration likhna padta tha.

MCP se pehle: Har AI model X Har Tool = Bahut saare connections (Messy).

MCP ke baad: Ek standard protocol jo sab jagah chalega.

The USB Analogy:Sochiye MCP "USB-C" hai.

AI Model aapka Laptop hai.

External Tool (Google Drive/Slack) aapka Printer ya Hard Drive hai.
MCP wo cable hai jo bina kisi naye driver ke inhe connect kar deti hai.
 LEVEL 1: AI FOUNDATIONS
5 MCP (Model Context Protocol) in AI
MCP Kaise Kaam Karta Hai?
MCP architecture mein mainly teen parts hote hain. Isko samajhna zaroori hai:

MCP Host (The App): Yeh wo application hai jisme AI model chalta hai. Example
ke liye Claude Desktop App ya Cursor IDE. Yeh host decide karta hai ki user kis
tool ko access kar sakta hai.

MCP Client (The Brain): Yeh Host ke andar hota hai. Yeh AI model aur MCP
Server ke beech baat karwata hai.

MCP Server (The Data Provider): Yeh sabse important part hai. Yeh ek
lightweight program hota hai jo aapke data (jaise local files, PostgreSQL
database, ya Google Drive) ko expose karta hai taaki AI usse padh sake.

Process Flow: User Request → MCP Host → MCP Client → MCP Server →
Data Fetch → AI Response
 LEVEL 2 :Advanced Machine Learning & Deep Learning
6 Convolutional Neural Networks in AI
CNN (Convolutional Neural Networks) AI aur Deep Learning ka ek bahut hi important part hai. Agar simple
shabdon mein kahein, toh CNN computer ki "aankhein" (eyes) hain.

Jab bhi computer ko koi photo, video, ya visual data dekh kar samajhna hota hai (jaise face unlock, self-
driving cars), toh wahan CNN ka use hota hai.
 LEVEL 2 :Advanced Machine Learning & Deep Learning
6 Convolutional Neural Networks in AI
CNN Kya Hai?

CNN ka full form hai Convolutional Neural Network. Yeh ek tarah ka Deep Learning
algorithm hai jo images ko input ke taur par leta hai, usme patterns (features) ko
pehchanta hai, aur batata hai ki wo image kis cheez ki hai.

Analogy: Sochiye aap ek Gulab (Rose) ko kaise pehchante hain?

Pehle aap dekhte hain ki petals ka shape kaisa hai (curved lines).

Fir aap color dekhte hain (Red/Pink).

Fir aap stem (tana) aur kaante dekhte hain.

Aakhir mein, aapka dimaag decision leta hai: "Yeh ek Gulab hai."

CNN bhi bilkul aise hi kaam karta hai. Wo puri image ko ek baar mein nahi ratta
maarta, balki use tukdon (layers) mein todkar patterns dhundta hai.
 LEVEL 2 :Advanced Machine Learning & Deep Learning
6 Convolutional Neural Networks in AI
Convolution: Features Dhundna. Image scan
karke lines/shapes nikalta hai.

ReLU: Activation. Data ko clean karta hai

(Negative values hataata hai).

Pooling: Size Chhota Karna. Image ko

compress karta hai taaki fast chale.

Flattening: Seedha Karna. 2D image data ko

1D list (line) banata hai.

Fully Connected: Decision Lena. Patterns

jodkar sochta hai ki yeh kya hai.

Output: Result Dena

LEVEL 2 :Advanced Machine Learning & Deep Learning
7 Recurrent Neural Network in AI
RNN (Recurrent Neural Networks) AI ka wo hissa hai jo unn cheezon ko samajhne ke liye bana hai jahan
Order (Sequence/Kram) aur Time (Waqt) matter karta hai.

Agar CNN computer ki "Aankhein" (Eyes) hain, toh RNN computer ki "Yaadasht" (Short-term Memory) hai.

RNN Kya Hai? (What is RNN?)

RNN (Recurrent Neural Network) ek aisa Deep Learning network hai
jo purani information ko yaad rakhta hai taaki aage aane wali
information ko samajh sake

Best Example:

Normal Network: Har shabd ko alag dekhega: "I", "Love", "India".

RNN: "I" dekha → Memory mein rakha. "Love" dekha → "I" ke saath
connect kiya. "India" dekha → Teeno ko connect karke samjha ki
baat desh-bhakti ki ho rahi hai.
 LEVEL 2 :Advanced Machine Learning & Deep Learning
7 Recurrent Neural Network in AI
RNN Kaise Kaam Karta Hai? (The Mechanism)

RNN ke kaam karne ka tarika ek Loop jaisa hota hai.

Iske structure mein ek khaas cheez hoti hai jise "Hidden State" kehte hain.

Input : Naya data jo abhi aaya hai (Current word/input).

Hidden State : Yeh RNN ki Memory hai. Yeh pichle step ka output hota hai
jo current step mein wapas feed kiya jata hai.

The Loop: RNN naye input ko purani memory ke saath combine karta hai
aur ek nayi memory create karta hai.

Equation ka Logic (Simple Bhaasha mein):

Aaj Ka Faisla = (Abhi Ki Khabar) + (Pichla Experience)

Current Output = Current Input + Previous Hidden State

LEVEL 2 :Advanced Machine Learning & Deep Learning
7 Recurrent Neural Network in AI

RNN Kahan Use Hota Hai? (Applications)

Jahan bhi data ek sequence mein ho, wahan RNN use hota hai:

Google Translate (Machine Translation): English sentence pura sunkar use Hindi mein
convert karna. (Grammar ka order change hota hai, isliye memory chahiye).

Speech Recognition (Siri/Alexa): Audio waves lagatar aati hain. "Hey" ke baad "Siri"
aane ki probability RNN hi calculate karta hai.

Predicting Next Word: Jab aap mobile par type karte hain aur keyboard suggest karta hai
agla word, wo RNN (ya uske advanced versions) kar rahe hote hain.

Stock Market Prediction: Aaj ka share price pichle 10 din ke trend par depend karta hai.
Sequence matter karta hai.
LEVEL 2 :Advanced Machine Learning & Deep Learning
8 Transformers in AI
Transformers AI ki duniya ka sabse bada "Game
Changer" hai. Aajkal aap jo bhi powerful AI dekh rahe
hain (jaise ChatGPT, Gemini, Claude), wo sab
Transformers architecture par hi bane hain.

Agar RNN "Short-term Memory" tha, toh Transformer

ek "Super Computer Brain" hai jo puri kitaab ek saath
padh sakta hai aur sab yaad rakh sakta hai.

Ise 2017 mein Google ke researchers ne introduce

kiya tha ek famous paper mein jiska naam tha:
"Attention Is All You Need".
 LEVEL 2 :Advanced Machine Learning & Deep Learning
8 Transformers in AI
Transformer Kya Hai? (What is a Transformer?)
Transformer ek Deep Learning model hai jo Sequence-to-Sequence tasks (jaise Translation, Text Summarization)
ke liye banaya gaya tha.

Lekin isne purane models (RNN/LSTM) ki sabse badi kami door kar di: Sequential Processing.

RNN: Ek-ek shabd padhta tha (One by one). Isliye slow tha aur lambi baatein bhool jata tha.

Transformer: Pura sentence ek saath (Parallel) dekhta hai. Isliye yeh bahut fast hai aur door-door ke words ka
rishta bhi samajh leta hai.

Iska Secret Sauce: "Self-Attention" Mechanism

Transformer ki sabse khaas cheez hai "Self-Attention". Iska matlab hai sentence ke har word ka dusre words se
connection check karna.

Example se samjhein: Sentence: "The animal didn't cross the street because it was too tired."

Jab insaan isko padhta hai, wo jaanta hai ki "it" ka matlab "animal" hai, "street" nahi.

Purane models (RNN) confuse ho jate the kyunki "it" aur "animal" ke beech doori hai.

Transformer (Attention) har word ko highlight karta hai. Jab wo "it" par pahunchta hai, toh wo piche mudkar
dekhta hai aur "animal" par sabse zyada attention (dhyan) deta hai.
 LEVEL 2 :Advanced Machine Learning & Deep Learning
8 Transformers in AI
Transformer Ka Structure (Architecture)
Transformer mainly do hisson mein banta hota hai: Encoder aur Decoder.

1. Encoder (The Reader / Samajhne wala)

Iska kaam hai input data (English sentence) ko lena aur use samajhna. Yeh data ke features
aur meaning nikalta hai.

Yeh input words ko numbers (vectors) mein convert karta hai.

Phir "Self-Attention" use karke context samajhta hai.

2. Decoder (The Writer / Banane wala)

Iska kaam hai Encoder ki samajh ko lekar output generate karna (Hindi sentence ya Answer).

Yeh predictions karta hai ki agla word kya hona chahiye.

Note:

BERT (Google Search mein use hota hai) sirf Encoder use karta hai (sirf samajhne ke liye).

GPT (ChatGPT) mainly Decoder use karta hai (text generate karne ke liye).
 LEVEL 2 :Advanced Machine Learning & Deep Learning
9 Self Supervised Learning in AI
Self-Supervised Learning (SSL) AI ki duniya mein ek
revolutionary concept hai. Agar simple shabdon mein
kahein, toh ye wo technique hai jahan AI model bina
kisi human help (labels) ke khud hi seekhta hai.

Isse pehle ki hum deep dive karein, ek simple analogy

samajhte hain:

Analogy: Imagine karo ek student hai jo library mein

baitha hai. Uske paas karodon books hain, lekin koi
teacher nahi hai jo bataye ki "Ye history hai" ya "Ye
science hai".

Wo student kya karta hai? Wo sentences ke aadhe

words ko dhak leta hai aur guess karta hai ki wahan
kya likha hoga. Agar wo sahi guess karta hai, matlab
use language samajh aa rahi hai. Yehi Self-Supervised
Learning hai.
LEVEL 2 :Advanced Machine Learning & Deep Learning
9 Self Supervised Learning in AI
Self-Supervised Learning Kya Hai?
Traditionally, AI do tarah se seekhta tha:

Supervised Learning: Jahan hum model ko batate hain (Label dete hain), e.g., "Ye
photo billi ki hai."

Unsupervised Learning: Jahan hum data de dete hain aur model patterns dhundhta hai
(Clustering).

Self-Supervised Learning in dono ke beech ka "Sweet Spot" hai. Isme hum Unlabeled
Data (jiska koi answer key nahi hai) use karte hain, lekin model data ke andar se hi
labels generate kar leta hai.

Core Mantra: The Data is its own Label. (Data khud hi apna answer hai).
 LEVEL 2 :Advanced Machine Learning & Deep Learning
9 Self Supervised Learning in AI
Ye Kaam Kaise Karta Hai? (Mechanism)
SSL main do steps mein kaam karta hai. Is process ko aksar "Pre-training" aur "Fine-tuning" kaha jata hai.
Step 1: Pretext Task (The Learning Phase)
Model ko ek fake problem di jati hai solve karne ke liye, taaki wo data ke structure ko samajh sake. Isse "Pretext
Task" kehte hain.
Examples of Pretext Tasks:
•Fill in the blanks (NLP):
• Input: "The sky is [MASK]."
• Model ka kaam: "[MASK]" ki jagah "blue" predict karna.
• Result: Is process se model grammar, context aur world knowledge seekh leta hai.
•Jigsaw Puzzle (Computer Vision):
• Input: Ek image ke 9 tukde kar diye aur unhe mix kar diya.
• Model ka kaam: Un tukdon ko wapas sahi order mein lagana.
• Result: Model ko samajh aata hai ki object kaisa dikhta hai (e.g., pair niche hote hain, sar upar).
Step 2: Downstream Task (The Application Phase)
Jab model "Pretext Task" se data ko samajh leta hai (Pre-trained ho jata hai), tab hum use kisi specific kaam
(Real task) ke liye use karte hain.
•Ab hum thoda sa labeled data use karke use Fine-tune karte hain.
•Kyunki model pehle se intelligent ho chuka hai, wo bohot kam data mein bhi nayi cheez seekh leta hai.
LEVEL 2 :Advanced Machine Learning & Deep Learning
10 Reinforcement Learning in AI
Reinforcement Learning (RL) AI sikhne ka wo
tareeka hai jo bilkul waise hi kaam karta hai
jaise ek chhota bacha seekhta hai—Galtiyan
karke aur unse seekh kar (Trial and Error).

Ye wo learning hai jahan koi teacher nahi hota jo

answer bataye. Balki, agar aap sahi kaam karte
ho toh Inaam (Reward) milta hai, aur galat karte
ho toh Saza (Penalty) milti hai. Bas isi reward ko
maximize karne ke chakkar mein AI intelligent
ban jata hai.
 LEVEL 2 :Advanced Machine Learning & Deep Learning
10 Reinforcement Learning in AI
RL ko samajhne ke liye ye 5 terms samajhna zaroori hai:

Agent (The Learner): Wo AI model jo seekh raha hai (e.g., Mario game
khelne wala bot).

Environment (The World): Wo jagah jahan Agent kaam kar raha hai
(e.g., Video game ka level ya Chess board).

Action (The Move): Agent kya kar sakta hai (e.g., Left jana, Right jana,
Jump karna).

Reward (The Feedback): Action ka result.

Positive (+10 points): Accha kaam kiya.

Negative (-5 points): Galti ki (e.g., aag mein gir gaya).

State (The Situation): Current situation kya hai (e.g., Mario abhi kahan
khada hai).
LEVEL 2 :Advanced Machine Learning & Deep Learning
10 Reinforcement Learning in AI
Ye Kaam Kaise Karta hai? (The Loop)
Ye ek continuous cycle hai jo chalti rehti hai:
[Link]: Agent dekhta hai ki wo abhi kahan hai (State).
[Link]: Wo decide karta hai ki kya karna hai (Action).
[Link]: Environment use batata hai ki ye action kaisa tha.

1. Agar accha tha -> Reward.

2. Agar bura tha -> Penalty.
[Link]: Agent apne dimaag (Policy) mein note kar leta hai ki "Is
situation mein ye action lene se fayda hua ya nuksan".
[Link]: Ye process lakhon baar repeat hoti hai jab tak Agent perfect
nahi ho jata.
 LEVEL 2 :Advanced Machine Learning & Deep Learning
11 Auto ML in AI
AutoML (Automated Machine Learning) ek aisi technology hai jo Machine Learning (ML)
models banane ke process ko automate karti hai. Iska main goal ye hai ki non-experts bhi
easily powerful AI models bana sakein aur experts ka time repetitive tasks se bache.

Analogy: Maan lo aapko ek cake banana hai.

Traditional ML: Aapko khud decide karna padega ki kitna aata lena hai, kitni cheeni, kitne
degree par bake karna hai, aur kaunsa oven use karna hai. Agar galti hui, toh dobara try
karna padega. (Ye experts karte hain).

AutoML: Aap bas ingredients (Data) machine mein daal dete ho, aur machine khud
decide karti hai ki best recipe kya hogi, temperature kya hoga, aur perfect cake (Model)
bana kar de deti hai.
LEVEL 2 :Advanced Machine Learning & Deep Learning
11 Auto ML in AI
AutoML Kya Hai?
Traditionally, ek ML model banane ke liye ek Data Scientist ko bohot saare steps
manually karne padte hain:
[Link] ko saaf karna (Cleaning).
[Link] choose karna (Feature Engineering).
[Link] algorithm select karna (e.g., Random Forest vs Neural Network).
[Link] (Hyperparameters) ko tune karna.
Ye kaam bohot time-consuming aur mushkil hota hai. AutoML in sabhi steps ko (ya
inmein se kuch ko) automate kar deta hai. Aap bas apna data upload karte ho, aur
AutoML background mein hazaron combinations try karke aapko Best Model de deta
hai.
LEVEL 2 :Advanced Machine Learning & Deep Learning
11 Model Compression (Pruning & Quantization) in AI

Model Compression AI models ko "Diet" par bhejne jaisa hai.

Aajkal ke AI models (jaise GPT, BERT) bohot "bhaari" (bulky)
hote hain. Inke parameters billions mein hote hain aur size
GBs mein hota hai. Inhe chalane ke liye mehenge GPUs aur
servers lagte hain.

Agar humein in models ko Mobile Phones, Smart Watches,

ya IoT devices par chalana hai (jahan memory aur battery
kam hoti hai), toh humein model ka size kam karna padta hai
bina uski intelligence (accuracy) khoye. Isi process ko Model
Compression kehte hain.

Iske do sabse main tareeke hain: Pruning aur Quantization.

LEVEL 2 :Advanced Machine Learning & Deep Learning
11 Model Compression (Pruning & Quantization) in AI

Pruning (Kaat-Chaat Karna)

Concept: Unn connections ko kaat dena jo zaroori nahi
hain.

Neural Networks mein karodon connections (weights)

hote hain. Training ke baad dekha gaya hai ki bohot
saare connections ki value 0 ke paas hoti hai, yaani wo
final result mein koi khaas yogdaan nahi de rahe.
Pruning ka matlab hai un bekar connections ko hata
dena.

Analogy: Imagine karo ek bada ghana ped (tree) hai.

Agar aap uski sookhi aur bekar tahniyan (branches)
kaat doge, toh ped ka shape wahi rahega, lekin wo
halka aur saaf ho jayega.
LEVEL 2 :Advanced Machine Learning & Deep Learning
12 Model Compression (Pruning & Quantization) in AI
Quantization (Precision Kam Karna)
Concept: Numbers ki detail (resolution) ko kam karke memory bachana.

Computers mein normally AI models ke weights 32-bit Floating Point (FP32) mein store
hote hain.

Example: 3.14159265 (Ye bohot jagah leta hai memory mein).

Quantization mein hum in numbers ko 8-bit Integers (INT8) mein convert kar dete hain.

Example: 3.14 ya sirf 3.

Analogy: Imagine karo aap ek 4K (Ultra HD) movie dekh rahe ho. File size 10GB hai.
Agar aap use 720p (HD) mein convert kar do, toh file size 1GB reh jayega. Picture
quality thodi kam hogi, lekin movie samajh mein aa jayegi. FP32 = 4K Movie | INT8 =
720p Movie
What You Will Learn in This Course
LEVEL 3: AI MODELING & GENERATIVE AI
• LLM (Large Language Model)
• Fine-Tuning
• LORA (Low-Rank Adaptation)
• GAN (Generative Adversarial Network)
• Diffusion Models
• Multimodal AI
• Vision-Language Models (VLMs)
• Synthetic Data
 LEVEL 2 : AI Modelling & Generative AI Advanced
13 Large Language Model in AI

LLM (Large Language Model) aaj ki AI revolution ka

sabse bada hero hai. Agar aapne ChatGPT, Gemini, ya
Claude use kiya hai, toh aap technically ek LLM se hi
baat kar rahe the.
Simple shabdon mein kahein toh:

Analogy: Maan lijiye ek aisa insaan hai jisne internet ki

saari kitabein, articles, aur websites padh li hain.
Use har sawal ka jawab pata hai aur wo kisi bhi bhasha
mein baat kar sakta hai. Bas yahi LLM hai.
 LEVEL 2 : AI Modelling & Generative AI Advanced
13 Large Language Model in AI
LLM koi jaadu nahi hai, ye Statistics aur Probability par chalta hai. Iska main
kaam bas ek hi hai: "Agla shabd (Next Word) guess karna.“

Sochiye: Agar main kahun: "Aaj mausam bohot _______" Aapka dimaag turant
bolega "suhana hai" ya "garam hai". LLM bhi yahi karta hai. Wo pichle words ko
dekh kar calculate karta hai ki agla shabd kya hona chahiye jo sabse sahi fit ho.

Lekin ye itna simple "Autocorrect" nahi hai. Ye poore paragraphs aur concepts ko
connect kar sakta hai kyunki ye Transformer Architecture par bana hota hai.

Transformer Architecture (The Brain): 2017 mein Google ne 'Transformer'

concept nikala tha. Isse pehle AI bhool jata tha ki paragraph ke shuru mein kya
likha tha. Transformer mein "Attention Mechanism" hota hai, jo poore sentence
mein important words par focus karta hai, chahe sentence kitna bhi lamba kyun na
ho.
 LEVEL 2 : AI Modelling & Generative AI Advanced
14 Fine Tuning in AI

Fine-Tuning AI ki duniya mein wahi process hai jo

ek Student ko Specialist doctor banane ka hota hai.
Agar simple shabdon mein kahein:

"Ek aise AI model ko jo pehle se kaafi kuch jaanta

hai (Pre-trained), use kisi ek khaas kaam (specific
task) ke liye expert banana hi Fine-Tuning hai."
 LEVEL 2 : AI Modelling & Generative AI Advanced
14 Fine Tuning in AI

The Best Analogy: MBBS Student vs. Heart Surgeon

Fine-Tuning ko samajhne ke liye ye example sabse best hai:
[Link]-Trained Model (MBBS Student):
1. Ek student ne MBBS kiya. Use dawaiyon, anatomy, aur bimariyon ka
general knowledge hai. Wo thoda bohot sab kuch janta hai.
2. AI Context: Ye GPT-4 ya BERT jaisa bada model hai jisne internet ka
saara data padh rakha hai.
[Link]-Tuning (Specialization):
1. Ab wo student "Heart Surgery" ki special training leta hai. Wo ab history
ya geography nahi padhta, sirf Heart (Dil) ke baare mein gehrai se
padhta hai.
2. AI Context: Hum us AI model ko medical data dete hain taaki wo poetry
likhna bhale hi bhool jaye, par bimari pakadne mein expert ho jaye.
 LEVEL 3 : AI Modelling & Generative AI Advanced
15 LORA – Low Rank Adaption in AI

LoRA (Low-Rank Adaptation) AI fine-tuning ki duniya

mein ek "Smart Hack" (Jugaad) hai.

Isse pehle humne "Fine-Tuning" samjha tha. Lekin

traditional Fine-Tuning mein ek bohot badi problem hai:

Cost aur Hardware. Agar aapko GPT-3 ya Llama-3

jaise huge models ko train karna hai, toh aapko
supercomputers chahiye. LoRA ne is problem ko
solve kar diya.

LoRA ki wajah se aaj ek normal gaming laptop par bhi

bade AI models ko train karna possible ho gaya hai.
 LEVEL 3 : AI Modelling & Generative AI Advanced
15 LORA – Low Rank Adaption in AI
LoRA = Low-Rank Adaptation.
Iska main logic ye hai: Bade models mein jab hum nayi
cheez sikhate hain, toh humein saare billion parameters
ko change karne ki zaroorat nahi hoti. Actual changes
bohot thode (Low Rank) hote hain.

Analogy (The Textbook Example):

•Pre-trained Model: Ek moti College ki Textbook hai.
•Full Fine-Tuning: Aap puri kitaab ke har page ko mita
kar dobara likh rahe ho. (Very heavy).

•LoRA: Aapne kitaab ko touch bhi nahi kiya. Aapne bas

important pages par Sticky Notes chipka diye jisme
naye notes likhe hain.

•Jab aap padhoge (Inference), toh aap Textbook +

Sticky Notes dono ko mila kar padhoge.
 LEVEL 3 : AI Modelling & Generative AI Advanced
16 GAN – Generative Adversarial Network

GAN (Generative Adversarial Network) AI ki duniya

ka sabse creative aur fascinating concept hai. Isne AI
ko sirf "Data Analyze" karne se "Data Create" karne ki
power di hai.

Agar ek line mein samjhein:

"Ye do AI networks ki ladai (fight) hai, jisme ladte-
ladte dono expert ban jate hain aur naya content
(Images, Audio, Video) banana seekh lete hain.“

Isse 2014 mein Ian Goodfellow ne banaya tha.

 LEVEL 3 : AI Modelling & Generative AI Advanced
16 GAN – Generative Adversarial Network

Simple Analogy: Chor aur Police (Thief vs. Police)

GAN ko samajhne ke liye sabse famous example ye hai:
[Link] Counterfeiter (Nakli Note Banane Wala Chor):
1. Iska kaam hai nakli note banana aur koshish karna ki wo asli lage.
2. Shuru mein wo bekar note banata hai.
[Link] Police (Inspector):
1. Iska kaam hai note ko check karna aur batana ki ye "Asli" hai ya
"Nakli".
Process:
•Chor note banata hai -> Police pakad leti hai ("Ye nakli hai!").
•Chor apni galti sudharta hai aur behtar note banata hai.
•Police bhi apni checking strict karti hai.
•Ye cycle chalti rehti hai.
•End Result: Ek time aisa aata hai jab Chor itna perfect nakli note banata
hai ki Police bhi fark nahi bata paati. Bas yahi GAN hai.
 LEVEL 3 : AI Modelling & Generative AI Advanced
16 GAN – Generative Adversarial Network

GAN ke 2 Main Parts (Architecture)

Technically, GAN mein do alag-alag Neural Networks
hote hain jo ek dusre ke khilaf kaam karte hain
(Adversarial matlab Virodhi):
A. Generator (The Artist/Faker)
•Kaam: Naya nakli data create karna.
•Input: Ye random numbers (Noise) leta hai.
•Goal: Aisi image banana jo Discriminator ko bewakoof
bana sake.
B. Discriminator (The Critic/Judge)
•Kaam: Pehchanna ki data asli hai ya nakli.
•Input: Iske paas Real Images (Training Data) aur
Generator ki Fake Images dono aati hain.
•Goal: Sahi classify karna (Real = 1, Fake = 0).
 LEVEL 3 : AI Modelling & Generative AI Advanced
17 Diffusion Models

Diffusion Models aajkal ke Generative AI (Image

generation) ka sabse powerful engine hai.

Agar aapne MidJourney, DALL-E 3, ya Stable

Diffusion ka naam suna hai ya use kiya hai, toh
samajh lijiye wo sab Diffusion Models par hi chalte
hain.

Simple shabdon mein:

"Ye wo AI hai jo 'Shor' (Noise) mein se 'Tasveer'
(Image) dhundh nikalta hai."
LEVEL 3 : AI Modelling & Generative AI Advanced
17 Diffusion Models

Simple Analogy: Dhundh se Saaf Tasveer (From

Fog to Clarity)
Imagine kariye aapke paas ek saaf photo hai.
[Link] uspar dhire-dhire dhool (dust) dalna shuru karte
hain.
[Link] der baad wo photo dhundhli ho jayegi.
[Link] mein, wo photo dikhna band ho jayegi, sirf
random noise (TV ki jhilmil) bachegi.
Ab imagine kariye ek aisa Artist (AI) jise humne ye
sikhaya hai ki "Is dhool ko wapas kaise hatana hai?"
Agar wo artist us random dhool (Noise) ko step-by-step
saaf kare, toh wapas ek nayi sundar tasveer ban jayegi.
Bas yahi Diffusion Model hai.
 LEVEL 3 : AI Modelling & Generative AI Advanced
17 Diffusion Models
Ye Kaam Kaise Karta Hai? (The 2-Step Process)
Diffusion models do main steps mein kaam karte hain: Forward aur
Reverse.
Step 1: Forward Diffusion (Destruction - Barbad Karna)
•Hum training data (e.g., Cat ki photo) lete hain.
•Hum uspe thoda-thoda karke Noise (Random pixels) add karte jate hain.
•Result: Ek point par photo puri tarah se Pure Noise ban jati hai. Isme ab
kuch nahi dikh raha.
•Kyun karte hain? Taaki hum record kar sakein ki image kaise bigadti hai.
Step 2: Reverse Diffusion (Creation - Wapas Banana)
•Yahi asli jaadu hai. Hum ek Neural Network (aksar U-Net) ko train karte
hain.
•Hum use wo "Pure Noise" dikhate hain aur kehte hain: "Isme se thoda sa
noise hatao aur batao piche kya image chupi hai."
•AI guess karta hai aur thoda noise hatata hai.
•Ye process 50 se 100 baar (steps) repeat hota hai.
•Har step par image thodi saaf hoti jati hai, aur ant mein ek HD Image ban
jati hai.
 LEVEL 3 : AI Modelling & Generative AI Advanced
18 Multimodal AI

Ab tak humne dekha ki AI ya toh Text samajhta tha

(LLMs), ya Images banata tha (Diffusion), ya Audio
sunta tha. Lekin Multimodal AI in sabka All-in-One
Combo hai.
Simple shabdon mein:
"Multimodal AI wo hai jo dekh sakta hai (Vision),
sun sakta hai (Audio), aur padh sakta hai (Text) —
sab ek saath."
Jaise hum insaan baat karte waqt sirf shabdon par
dhyan nahi dete, balki saamne wale ke chehre ke
expressions (Image) aur aawaz ke tone (Audio) ko bhi
samajhte hain, Multimodal AI bhi wahi karta hai.
LEVEL 3 : AI Modelling & Generative AI Advanced
19 Vision Language Model

VLM Kya Hai? (The Core Concept)

VLM do badi technologies ka mishran (mixture)
hai:
[Link] Encoder (Aankhein): Ye image ko
scan karta hai aur uske features nikalta hai (e.g.,
"Yahan ek billi hai", "Rang safed hai").
[Link] Model (Dimaag/Zubaan): Ye un
features ko samajhta hai aur insaani bhasha
mein convert karta hai.
Result: Aap ek photo upload karte hain aur
puchte hain "Is photo mein kya gadbad hai?" aur
VLM uska jawab text mein deta hai.
 LEVEL 3 : AI Modelling & Generative AI Advanced
20 Synthetic Data
Synthetic Data ka matlab hai wo data jo asli duniya
(real world) se collect nahi kiya gaya, balki computer
programs aur AI algorithms dwara banaya gaya hai.
Agar simple shabdon mein kahein:
"Ye nakli data hai jo bilkul asli data jaisa behave
karta hai, taaki AI ko train kiya ja sake bina kisi ki
privacy khatre mein daale."
 LEVEL 3: AI Modelling & Generative AI Advanced
20 Synthetic Data

Simple Analogy: Video Game vs. Real Driving

Imagine karo aapko gaadi chalana seekhna hai.
[Link] Data: Aap ek asli car lekar road par nikal jao.
(Risk: Accident ho sakta hai, petrol ka kharcha, traffic ki
problem).
[Link] Data: Aap ek High-Quality Video Game
(simulator) mein gaadi chalao.
1. Wahan sadak, gaadi, aur traffic sab computer
ne banaya hai (Synthetic).
2. Lekin rules wahi hain (break lagane par gaadi
rukti hai).
3. Aap bina kisi risk ke 100 baar crash kar sakte
ho aur seekh sakte ho.
AI ki training mein ye "Video Game wala data" hi
Synthetic Data kehlata hai.
What You Will Learn in This Course
LEVEL 4: INFORMATION RETRIEVAL &
KNOWLEDGE SYSTEMS
• Vector Database
• RAG (Retrieval Augmented Generation)
• Knowledge Graphs
• Explainable AI (XAI)
• Federated Learning
LEVEL 4 :Information Retrieval & Knowledge system
Advanced +
21 Vector Database
Vector Database wo jagah hai jahan AI apni "Yaadein"
(Memories) store karta hai.

Traditional databases (jaise SQL/Excel) sirf Exact

Match samajhte hain. Lekin AI ko Meaning (Matlab)
samajhna hota hai. Vector Database wahi technology
hai jo AI ko "Context" aur "Meaning" ke basis par data
dhundhna sikhati hai.

Simple shabdon mein:

"Agar SQL Database ek Dictionary hai (jahan exact
word dhundhna padta hai), toh Vector Database ek
librarian hai (jise aap story samjhaoge toh wo waisi
hi kitab nikal kar de dega)
 LEVEL 4 :Information Retrieval & Knowledge system
Advanced +
22 RAG Retrieval Augmented Generation
RAG (Retrieval-Augmented Generation) wo
technique hai jo ChatGPT jaise models ko "Open Book
Exam" dena sikhata hai.
Agar simple shabdon mein kahein:

"LLM (dimaag) ke paas apni yaaddasht hai, lekin

RAG use ek 'Kitaab' (External Data) pakda deta hai
taaki wo padh kar sahi jawab de sake, na ki sirf
yaad karke.“

Bina RAG ke AI "Ratta maar ke" exam deta hai. RAG

ke saath wo "Kitaab khol kar" answer dhundhta hai.
LEVEL 4 :Information Retrieval & Knowledge system
Advanced +
22 RAG Retrieval Augmented Generation
Problem: LLMs ki Kamiyan
RAG kyun chahiye? Kyunki GPT-4 ya Llama jaise
models mein 2 badi problems hain:
[Link] Cutoff (Purana Data):
1. Agar model ki training 2023 mein khatam hui
hai, toh use aaj ki news nahi pata.
[Link] (Fenkna):
1. Agar use answer nahi pata, toh wo confidence
ke saath jhoot bol deta hai.
[Link] Data:
1. ChatGPT ko aapki company ki internal
policies ya emails ke baare mein kuch nahi
pata.
Solution: RAG in problems ko solve karta hai AI ko
Live Data ya Private Data ka access dekar.
LEVEL 4 :Information Retrieval & Knowledge system
Advanced +
22 RAG Retrieval Augmented Generation
RAG Kaam Kaise Karta Hai? (The 3 Steps)
RAG ka full form hi iska process batata hai:
1.R - Retrieval (Dhundhna):
1. Jab user sawal puchta hai, system sabse
pehle aapke database (Vector DB) mein
search karta hai aur relevant information
nikalta hai.
2.A - Augmentation (Jodna):
1. Us nikali hui information ko user ke sawal ke
saath jod diya jata hai (Context banana).
3.G - Generation (Jawab Banana):
1. Ab ye poora masala (Sawal + Information)
LLM ko diya jata hai. LLM ise padhta hai aur
answer generate karta hai.
 LEVEL 4 :Information Retrieval & Knowledge system
Advanced +
23 Knowledge Graphs
Knowledge Graph AI ka wo "Dimaag" hai jo
data ko sirf store nahi karta, balki connect karta
hai aur uska matlab (meaning) samajhta hai.
Agar simple shabdon mein kahein:

"Agar SQL Database ek Excel sheet hai

(Rows & Columns), toh Knowledge Graph ek
Detective ka Board hai jahan photos (data)
ek dusre se laal dhaage (relationships) se
judi hoti hain."

Ye AI ko insaanon ki tarah "Rishte"

(Relationships) aur "Context" samajhne mein
madad karta hai.
LEVEL 4 :Information Retrieval & Knowledge system
Advanced +
24 Explainable ai (XAI)
Explainable AI (XAI) wo technique hai jo humein ye
samajhne mein madad karti hai ki AI ne koi faisla
kyun liya.

Aajkal AI models (jaise Deep Learning) itne complex ho

gaye hain ki wo sahi jawab toh de dete hain, par kaise
diya ye batana mushkil hota hai. Isi problem ko solve
karne ke liye XAI aaya hai.

Simple shabdon mein:

"Normal AI ek Student ki tarah hai jo bas 'Answer'
likh deta hai. Explainable AI wo Student hai jo
Answer ke saath 'Steps aur Logic' bhi dikhata hai."
 LEVEL 4 :Information Retrieval & Knowledge system
Advanced +
25 Federated Learning
Federated Learning (FL) AI ki duniya mein Privacy ka
sabse bada rakshak (guardian) hai.

Ab tak humne jitne bhi AI models dekhe (LLMs, Vision,

etc.), wo sab "Centralized" tareeke se kaam karte
hain: "Saara data user se lo aur ek bade server par
jama karo, phir wahan model train karo.“

Lekin Federated Learning ka concept ulta hai:

"Data ko Model ke paas mat laao, balki Model ko

Data ke paas le jao."
Simple shabdon mein: Aapka personal data aapke
phone se bahar kabhi nahi jata, phir bhi AI seekh
leta hai.
What You Will Learn in This Course
LEVEL 5: AUTONOMOUS AI & FUTURE TECH (Cutting-Edge)
• AI Agent
• Edge AI
• Zero-Shot Learning
• AGI (Artificial General Intelligence)
• ASI (Artificial Super Intelligence)
LEVEL 5 :Autonomous AI & Future Tech – Cutting Edge
26 AI Agent

AI Agent Artificial Intelligence ka wo "Action Mode" hai

jo sirf baatein nahi karta, balki kaam (tasks) karke deta
hai.

Ab tak humne LLMs (ChatGPT) dekha, jo ek "Smart

Advisor" ki tarah hai. Aap sawal puchte ho, wo jawab
deta hai.

Lekin AI Agent ek "Smart Employee/Intern" ki tarah

hai. Aap use goal dete ho, aur wo khud tools use karke
us goal ko pura karta hai.
LEVEL 5 :Autonomous AI & Future Tech – Cutting Edge
26 AI Agent
AI Agent Kya Hai? (The Core Concept)
AI Agent ek aisa system hai jiske paas 3 cheezein hoti
hain:
[Link] (LLM): Sochne ke liye (jaise GPT-4).
[Link] (Haath-Pair): Kaam karne ke liye (Web
Search, Calculator, Code Editor, APIs).
[Link] (Autonomy): Khud decision lene ki taqat.
Difference Example:
•Normal AI: Aapne pucha "Mumbai ka weather kaisa
hai?" -> Wo apne purane data se jawab dega (shayad
galat ho).
•AI Agent: Aapne pucha "Mumbai ka weather kaisa
hai?" -> Ye turant Google Search Tool use karega, aaj
ka weather check karega, aur sahi jawab dega.
 LEVEL 5 :Autonomous AI & Future Tech – Cutting Edge
26 AI Agent
AI Agent ek loop mein kaam karta hai jise aksar "Perception-
Action Loop" kehte hain.
[Link] (Maalik ka hukum): User ne bola: "Mere liye ek website
banao aur host kar do."
[Link] (Sochna):
1. Agent ka dimaag (LLM) sochega: "Pehle code likhna
padega, phir HTML file banani padegi, phir server par
upload karna padega."
[Link] Use (Action):
1. Wo Code Interpreter tool use karke code likhega.
2. Agar error aaya, toh wo khud debug karega.
[Link] (Feedback):
1. Wo check karega ki code chala ya nahi.
[Link]:
1. Jab kaam ho jayega, wo aapko bolega: "Sir, website
live hai, ye raha link."
Is process mein insaan ki zarurat nahi padi, agent ne khud steps
liye.
LEVEL 5 :Autonomous AI & Future Tech – Cutting Edge
27 Edge AI

Edge AI ka simple matlab hai: "AI jo Internet (Cloud)

par nahi, balki aapke Device (Phone, Watch, Car)
par chalta hai.“

Ab tak humne dekha ki zyadatar AI (jaise ChatGPT)

Cloud par chalta hai. Aap sawal bhejte ho, wo server
par jata hai, process hota hai, aur jawab wapas aata
hai. Lekin Edge AI mein ye processing servers par
nahi, balki "Edge" (kinare) par hoti hai—yaani jahan
data paida ho raha hai (aapke haath mein).

Simple shabdon mein:

"Agar Cloud AI 'Restaurant se khana mangwana'
hai (Time lagta hai), toh Edge AI 'Ghar par khana
banana' hai (Turant mil jata hai)."
LEVEL 5 :Autonomous AI & Future Tech – Cutting Edge
28 Zero Shot Learning
Zero-Shot Learning (ZSL) AI ki wo capability hai jahan
model un cheezon ko pehchan leta hai ya un tasks ko
kar leta hai jo usne training ke dauran kabhi dekhe
hi nahi.

Normal AI (Supervised Learning) ko sikhana padta hai:

"Ye billi hai, ye kutta hai." Tab wo inhe pehchanta hai.
Lekin Zero-Shot Learning mein AI bina kisi example
ke, sirf Description ya Logic ke dum par nayi cheezon
ko pehchan leta hai.

Simple shabdon mein:

"Ye waisa hi hai jaise ek bachche ko 'Zebra' ki

photo na dikhayi jaye, bas bataya jaye ki 'Ye ghode
(Horse) jaisa dikhta hai par uspe kaali-safed
dhariyan (stripes) hoti hain.' Jab bachcha pehli baar
Zebra dekhega, wo use pehchan lega."
 LEVEL 5 :Autonomous AI & Future Tech – Cutting Edge
29 Artificial General Intelligence AGI
Artificial General Intelligence (AGI) AI ki wo future stage hai jahan machine ka
dimaag bilkul insaan ke dimaag jaisa ho jayega.
Isse simple points mein samjhein:
[Link] hai AGI?
1. Abhi ke AI (Narrow AI) sirf ek kaam mein expert hote hain (jaise ChatGPT
likhne mein, AlphaGo chess mein).
2. AGI wo AI hoga jo koi bhi kaam kar sakega—coding, painting, gaadi
chalana, nayi bhasha seekhna, aur common sense lagana—bina kisi insaan
ki madad ke.
[Link]:
1. Narrow AI: Ek calculator jaisa hai (sirf hisaab kar sakta hai).
2. AGI: Ek student jaisa hoga (jo maths, science, sports sab seekh sakta hai).
[Link] Status:
1. Abhi tak AGI bana nahi hai. Ye sirf theory aur movies (jaise Iron Man ka
JARVIS) mein hai. Scientists maante hain ki ye aane wale kuch dashakon
(decades) mein aa sakta hai.
Short mein: AGI matlab ek aisa computer jo sochne, samajhne aur seekhne mein
insaan se kam na ho.
LEVEL 5 :Autonomous AI & Future Tech – Cutting Edge
30 Artificial Super Intelligence ASI
Artificial Super Intelligence (ASI) AI ki wo
"God-like" stage hai jahan machine ki
intelligence insaani dimaag se kahin zyada
aage nikal jati hai.

Agar AGI (General Intelligence) insaan ka

Barabari (Equal) hai, toh ASI insaan ka Baap
(Superior) hai.
Simple shabdon mein:

"ASI ek aisa dimaag hai jo duniya ke sabse

intelligent scientists (jaise Einstein) se bhi
karodon guna tez, creative, aur samajhdaar
hoga. Ye wo problems chutki mein solve kar
dega jo insaanon ko solve karne mein
hazaron saal lagenge."
Contact - 7880113112