AI Class 10 – Comprehensive Bullet-Point Notes

Acknowledgements

Patrons
- Mr. Rahul Singh (IAS), Chairperson, CBSE
Guidance & Support
- Dr. Biswajit Saha – Director (Skill Education & Training), CBSE
- Ms. Shweta Khurana – Senior Director, APJ Government Partnerships & Initiatives, Intel
Education Value-Adders / Coordinators
- Sh. Ravinder Pal Singh (Joint Secretary, Dept. of Skill Education, CBSE)
- Ms. Saloni Singhal (Program Manager, Intel Digital Readiness)
- Ms. Sarita Manuja (Educational Consultant, NHES)
- Ms. Shatarupa Dasgupta (National Program Manager, Intel)
Content Curation Team
- Ms. Ambika Saxena, Ms. Prachi Chandra (Intel AI for Youth Coaches)
- Teachers from APS Meerut, DPS Rourkela, DPS Bangalore South, Shaheed Rajpal DAV Delhi, DAV Gurugram, Indirapuram PS Ghaziabad

About the Book & Curriculum Structure

Context
- AI recognised globally as driver of future digital economy
- India/CBSE–Intel partnership (since 2019) to embed AI readiness in secondary education
Design philosophy
- Provide roadmap to navigate AI concepts while fostering creativity & social good
- Updated tech + social concepts, real-life cases, no-code project guides
Key features
- Enhanced, elaborated content with fresh examples
- Additional real-world scenarios to demystify complex ideas
- Emphasis on AI-enabled social-impact solutions
- Domain-wise use-case walkthroughs
Unit–Session Matrix (Grade X)
- Unit 1 (15 h) – Revisiting AI Project Cycle & Ethical Frameworks
- Unit 2 (25 h) – Advanced Concepts of Modelling
- Unit 3 (25 h) – Evaluating Models
- Unit 4 (28 h) – Statistical Data & No-Code Tools
- Unit 5 (30 h) – Computer Vision
- Unit 6 (27 h) – Natural Language Processing

Unit 1 – AI Project Cycle & Ethical Frameworks

Learning Objectives & Outcomes

Recall 6 stages of AI Project Cycle & three AI domains
Define frameworks ⇢ ethical frameworks ⇢ bioethics
Apply a chosen ethical framework to avoid unintended consequences

1.1 AI Project Cycle (6 Stages)

Problem Scoping → Data Acquisition → Data Exploration → Modelling → Evaluation → Deployment
Analogy: Making a birthday greeting card – sequential steps mirror project stages
Key points per stage
- Problem Scoping: specify goal, parameters, constraints
- Data Acquisition: collect authentic, reliable datasets
- Data Exploration: visualise via graphs/maps to uncover patterns
- Modelling: choose/test candidate algorithms, pick most efficient
- Evaluation: test with fresh data, compute accuracy/error, refine
- Deployment: integrate into real-world to deliver value

1.2 AI Domains (determined by nature of data)

Statistical Data
- Deals with large tabular/quantitative datasets
- Example: Price-comparison websites (PriceGrabber, Shopzilla) analyse multi-vendor price tables
Computer Vision (CV)
- Machines capture → screen → analyse images/videos → generate decisions
- Inputs: photos, drone footage, IR images, etc.
- Examples: Crop monitoring via drones, surveillance, retail, self-driving cars
Natural Language Processing (NLP)
- Algorithmic extraction of information from human language (spoken/written)
- Goal: $\text{Read} \;\to\; \text{Decipher} \;\to\; \text{Understand} \;\to\; \text{Generate Insight}$
- Examples: Spam filters, Google Translate, sentiment analysis

1.3 Frameworks & Ethical Frameworks

Framework = ordered set of steps to solve problems; provides common language & rigour
Ethical framework = ensures choices avoid unintended harm; systematic navigation of moral dilemmas

Why AI Needs Ethical Frameworks

AI acts as decision-making/influencing tool
Documented failures (e.g., gender-biased hiring algorithm) reveal risk of bias, discrimination
Embedding ethics early prevents costly post-deployment fixes

Factors Influencing Human Decisions (exposed via “My Goodness” activity)

Recipient identity & location
Familial bias
Information transparency
Personal religion, intuition, value of humans/non-humans

Classification of Ethical Frameworks

Sector-Based
- Tailored to industry context (healthcare, finance, law enforcement, etc.)
- Example: Bioethics for health/life-sciences
Value-Based
- Rooted in moral philosophy
  - Rights-Based ⇒ protect human dignity/autonomy
  - Utility-Based ⇒ maximise overall good, minimise harm
  - Virtue-Based ⇒ actions align with integrity, honesty, compassion

Bioethics (Sector-Based Example)

Four guiding principles
1. Respect for Autonomy
2. Non-Maleficence (Do no harm)
3. Beneficence (Ensure maximum benefit)
4. Justice (Fair distribution of benefits/burdens)

Case Study – Biased Healthcare Risk Algorithm

Aim: optimise patient-care resource allocation
Flaw
- Trained on “health-care spend” $\neq$ actual illness severity
- Historical under-spend for western-region patients ⇒ algorithm under-estimated their risk
Ethical correction via Bioethics
- Autonomy ⇒ dataset & model logic transparent to patients
- Non-Maleficence ⇒ re-train on equitable data to avoid harm
- Beneficence ⇒ maximise positive outcomes across all regions
- Justice ⇒ address systemic biases & social determinants of health

Unit 2 – Advanced Concepts of Modelling in AI

2.1 Revisiting AI vs ML vs DL

Artificial Intelligence: umbrella term – machines mimic human intellect
Machine Learning: subset enabling improvement via experience
- Broad pipeline: $Input \;\xrightarrow{Model}\; Output$ (learns from labeled or unlabeled data)
Deep Learning: subset of ML using multiple algorithms (layers of neural networks) + vast data for self-training

Common Data Terminology

Feature = column/attribute (e.g., colour, price)
Label = target tag (depends on task context)
Labeled vs Unlabeled Data
Training Set: examples used for learning
Testing Set: unseen data to assess accuracy

2.2 Modelling Approaches

Rule-Based

Developer explicitly encodes rules; learning static
Example: FAQ chatbot using decision tree – no adaptation beyond scripted paths

Learning-Based (Machine Learning / Deep Learning)

Model derives rules from data; adapts when data distribution shifts
Example: Spam filter that refines with each new labelled email

Categories of ML (Learning-Based) Models

Supervised Learning (labelled data)
- Sub-types
  - Classification (discrete labels) – e.g., spam/not-spam, hot/cold weather
  - Regression (continuous output) – e.g., house price, temperature forecast
Unsupervised Learning (unlabelled data)
- Sub-types
  - Clustering – form natural groups (e.g., user segments, song preferences)
  - Association – discover co-occurrence rules (e.g., bread → butter in supermarket)
Reinforcement Learning
- Agent interacts with environment; learns policy $\pi$ to maximise cumulative reward
- Examples: self-parking car, walking humanoid in DeepMind demo

Rule vs Learning – Quick Contrast

Rule Based → fixed, brittle, limited generalisation
Learning Based → adaptive, handles exceptions, but requires data & compute

2.3 Neural Networks (NN)

Inspired by human neurons; organised in layers
- Input Layer – receives raw features
- Hidden Layers – perform weighted computations $Net = \sum w<em>i x</em>i + b$ , pass through activation
- Output Layer – produces final prediction/probability
Training adjusts weights $w_i$ & bias $b$ to minimise error (gradient descent/optimisation)
Capable of automatic feature extraction, especially with large complex datasets (images, text)

Perceptron Illustration – “Go to the Park?”

Inputs (binary): jacket?, umbrella?, sunny now?, forecast later?
Weights represent importance; bias tunes decision threshold
Decision rule y = \begin{cases}1, & \text{if } \sum wi xi - b > 0\0, & \text{otherwise}\end{cases}
- Scenario 1 yielded $y = 0.5$ ⇒ go out
- Scenario 2 yielded $y = -0.5$ ⇒ stay in

Classroom Activity – Human Neural Network

Roles: 7 Input nodes → 6 Hidden L1 → 6 Hidden L2 → 1 Output node
Process
- Each input student writes 6 key words about a secret image and passes one to each L1 node
- L1 condenses to 4 words, passes to L2
- L2 condenses to 2 words, passes all to output
- Output summarises in ≤ 5 lines; accuracy reveals effectiveness of collective “network”
Rules: no talking, one word per chit, fair distribution of chits; demonstrates feature abstraction across layers

Summary Diagram – Family of Models

Supervised → Classification, Regression
Unsupervised → Clustering, Association
Reinforcement → Reward maximisation
Deep Learning implements above via ANN / CNN architectures

Conceptual & Practical Connections

Ethical considerations (Unit 1) remain relevant across all modelling choices (Unit 2) → bias, transparency, accountability during data selection, training & deployment
No-Code tools (Unit 4 onward) allow rapid prototyping of concepts learned in Units 1-3, making AI accessible to non-programmers while maintaining ethical rigor

Philosophical / Social Implications Highlighted

Bias amplification in AI can reinforce systemic inequities (gender, race, region) if unchecked
Sector-based ethics safeguard domain-specific stakes (e.g., patient lives in healthcare)
Value-based ethics promote universal moral grounding transcending industries

Numerical / Statistical References & Equations Mentioned

AI decision threshold example: 0.5 > 0 \Rightarrow \text{Go Out}; -0.5 < 0 \Rightarrow \text{Stay In}
Generic neural computation: $Net = \sum<em>{i=1}^n w</em>i x_i + b$

Quick Self-Check Prompts (Derived)

Name six stages of AI Project Cycle.
Distinguish Statistical Data vs CV vs NLP with one real-world use-case each.
List four bioethics principles.
Map “loan-default prediction” to its learning & model category.
Provide one advantage & one drawback of rule-based systems.
Explain why reinforcement learning is fit for dynamic, uncertain environments.