๐ Your Badges
๐ Topic Progress
๐ AI Board Exam Simulator
Practice in CBSE exam conditions. Choose your exam type and start โ timer begins immediately. Questions drawn from all 5 chapters and advanced topics.
๐ค AI & ML Study Portal
CBSE Class 9โ10 ยท Artificial Intelligence ยท Code 417 ยท CBSEMOCK
Chapter 1: Introduction to AI
History, definitions, types of AI, Turing Test, AI vs Human, applications.
Chapter 2: AI Project Cycle
Project lifecycle, 4W Canvas, problem scoping, Big Data 5 V's, data acquisition.
Chapter 3: Data Exploration
EDA, preprocessing, cleaning, normalization, feature engineering, visualisation.
Chapter 4: Modelling
ML algorithms, decision trees, KNN, linear regression, model training & testing.
Chapter 5: Evaluation
Confusion matrix, accuracy, precision, recall, F1 score, model improvement.
T1: AI Ethics & Bias
Fairness, accountability, deepfakes, bias types, global frameworks.
T2: Data Literacy
Data types, quality, Big Data, labelling, misleading statistics.
T3: Maths for AI
Probability, Bayes theorem, statistics, matrices, correlation.
T4: Generative AI
LLMs, GANs, diffusion models, hallucinations, applications.
T5: Neural Networks
Perceptrons, activation functions, CNNs, backpropagation.
T6: NLP
Tokenisation, TF-IDF, Word2Vec, Transformers, sentiment analysis.
T7: Model Evaluation
Confusion matrix, precision vs recall, F1, K-fold cross-validation.
T8: Computer Vision
Image classification, CNNs, convolution, object detection, filters.
T9: Deep Learning
CNN architectures, ResNet, transfer learning, RNN, LSTM, Transformers.
T10: Data Science
Python, Pandas, NumPy, Matplotlib, data analysis workflow.
1.1 What is Artificial Intelligence?
Artificial Intelligence (AI) is the simulation of human intelligence processes by machines, especially computer systems. It enables machines to learn from experience, adjust to new inputs, and perform human-like tasks.
Fig 1.1: Human Brain vs AI Machine โ similarities and differences
1.2 History and Evolution of AI
Fig 1.2: Timeline of AI milestones
1.3 Types of AI
Fig 1.3: Three levels of AI โ from Narrow (today) to Super-intelligence (theoretical future)
1.4 Turing Test โ Can Machines Think?
Fig 1.4: The Turing Test โ evaluating machine intelligence by human indistinguishability
1.5 Domains of AI
Natural Language Processing
Understanding & generating human language. e.g. Google Translate, chatbots
Computer Vision
Understanding images & video. e.g. Face recognition, medical imaging
Reinforcement Learning
Learning via reward & punishment. e.g. AlphaGo, self-driving cars
Speech Recognition
Converting spoken words to text. e.g. Siri, Alexa, Google Assistant
Robotics
Physical machines with AI. e.g. Factory bots, surgical robots
Expert Systems
Rule-based AI for expert decisions. e.g. Medical diagnosis, tax software
1.6 AI vs Human Intelligence
| Feature | ๐ง Human Intelligence | ๐ค Artificial Intelligence |
|---|---|---|
| Learning | From experience & education | From data & algorithms |
| Speed | Limited by biology | Millions of ops/second |
| Creativity | True creativity & emotions | Pattern-based imitation |
| Fatigue | Gets tired | No fatigue, 24/7 |
| Common Sense | Natural common sense | Limited, domain-specific |
| Adaptability | Adapts quickly to new tasks | Limited to trained domain |
| Energy | ~20 watts (brain) | Kilowatts (data centres) |
1.7 Applications of AI in Daily Life
Fig 1.5: AI application areas in everyday life
1.8 Howard Gardner's 9 Types of Intelligence
Psychologist Howard Gardner proposed that human intelligence is not a single ability but a collection of distinct intelligences. AI systems are designed to replicate specific types of these intelligences โ and currently, no AI replicates all nine.
Fig 1.6: Howard Gardner's 9 Types of Intelligence and their AI equivalents
1.9 Interactive: Match AI Applications to Their Domain
Test your knowledge! Click each application card and match it to the correct AI domain. Each card reveals its domain when clicked.
1.10 AI Future Opportunities โ Careers & Possibilities
AI is not replacing humans โ it is creating entirely new professions that did not exist before. Students who understand AI today will lead the workforce of tomorrow.
Fig 1.7: AI career opportunities and future sectors โ where your learning leads
โข AI currently replicates Linguistic, Logical, and Spatial intelligences best
โข Intrapersonal and Existential intelligences remain uniquely human
โข AI = machines simulating human intelligence
โข Turing Test (1950) = benchmark for machine intelligence
โข ANI (Narrow) โ AGI (General) โ ASI (Super) โ current AI is ANI
โข AI does NOT have emotions, consciousness, or true creativity
โข AI is creating new careers, not just replacing old ones
CBSE Practice Paper โ Chapter 1: Introduction to AI
Class: 9โ10 | Subject: Artificial Intelligence (Code 417) | Max Marks: 40
Time: 1.5 Hours | All questions are compulsory
Section A โ Multiple Choice Questions (1 mark each ร 10 = 10 marks)
1. The term "Artificial Intelligence" was coined by: 1
- a) Alan Turing
- b) John McCarthy โ
- c) Marvin Minsky
- d) Claude Shannon
2. The Turing Test was proposed in the year: 1
- a) 1943
- b) 1950 โ
- c) 1956
- d) 1969
3. Which type of AI exists today? 1
- a) ASI
- b) AGI
- c) ANI โ
- d) All of these
4. Which of the following is NOT an AI application? 1
- a) Google Maps
- b) Simple calculator โ
- c) Netflix recommendations
- d) Face recognition
5. Deep Blue (1997) was famous for: 1
- a) Playing chess against Garry Kasparov โ
- b) Answering quiz questions
- c) Generating images
- d) Driving a car
6. AGI stands for: 1
- a) Advanced Graphical Intelligence
- b) Artificial General Intelligence โ
- c) Automated Graph Interface
- d) None of these
7. In the Turing Test, the interrogator communicates via: 1
- a) Voice call
- b) Video call
- c) Text interface โ
- d) Physical meeting
8. Which domain of AI deals with understanding images? 1
- a) NLP
- b) Computer Vision โ
- c) Expert Systems
- d) Robotics
9. Which AI assistant uses voice recognition? 1
- a) Deep Blue
- b) Excel
- c) Siri โ
- d) Paint
10. Reinforcement Learning is best described as: 1
- a) Learning from labelled data
- b) Learning by reward and punishment โ
- c) Learning without any data
- d) None of these
Section B โ Short Answer Questions (2 marks each ร 5 = 10 marks)
11. Define Artificial Intelligence and give two examples from daily life. 2
12. What is the Turing Test? Who proposed it and when? 2
13. Differentiate between ANI and AGI with one example each. 2
14. List any four domains of AI with one application for each. 2
15. How does AI differ from traditional computer programs? 2
Section C โ Long Answer Questions (5 marks each ร 2 = 10 marks)
16. Explain the three types of AI (ANI, AGI, ASI) with examples. Why is only ANI achieved today? 5
17. Compare human intelligence and artificial intelligence on any five parameters. Which do you think is better and why? 5
Section D โ HOTS & Case Study (10 marks)
18. (HOTS) If a robot could perfectly mimic human behaviour, should it be treated as a human? Discuss ethically. 5
19. (Case Study) A hospital uses AI to analyse X-rays to detect cancer. The AI was trained on 1 million X-rays. (a) What type of AI is this? (b) What domain of AI is used? (c) What are the advantages of using AI for this task? (d) What are the risks? (e) Should doctors fully trust the AI? 5
2.1 The AI Project Lifecycle
Building an AI solution is a structured process โ like constructing a building. Each phase builds on the previous one. Skipping a step leads to poor AI outcomes.
Fig 2.1: The AI Project Lifecycle โ 5 iterative stages
2.2 The 4W Problem Canvas
Fig 2.2: The 4W Canvas โ structured way to define an AI problem before building
2.3 Big Data โ The 5 V's
Fig 2.3: Big Data 5 V's โ characteristics that define truly "big" data
| V | Instagram Example | Scale |
|---|---|---|
| Volume | Photos, reels, stories uploaded daily | 100M+ posts/day |
| Velocity | Likes, comments, shares happening instantly | Real-time processing |
| Variety | Images, videos, text captions, emojis, audio | Multi-format |
| Veracity | Fake accounts, spam bots, misleading content | Quality challenge |
| Value | Ad targeting, creator analytics, trend detection | Billions $ revenue |
2.4 Interactive โ 4W Problem Scenario Canvas
Apply the 4W framework to a real scenario! Read the scenario below and click each quadrant to reveal how 4W thinking applies. Try building your own AI project using this thinking.
2.5 AI and the 17 UN Sustainable Development Goals (SDGs)
The United Nations has set 17 Sustainable Development Goals (SDGs) to build a better world by 2030. AI is a powerful tool to accelerate progress towards these goals โ especially in developing countries like India.
Fig 2.4: AI applications aligned to UN Sustainable Development Goals โ with focus on Zero Hunger in India
2.6 Data Acquisition โ Collecting the Right Data
Data acquisition is the process of collecting, gathering, and importing raw data for an AI project. The quality of your AI model depends entirely on the quality and quantity of data collected โ "Garbage In, Garbage Out."
Fig 2.5: Data Acquisition โ techniques, categories, dataset types, and risks of poor data
Authentic data means real, accurate, unbiased data from genuine sources. If a hospital AI is trained only on data from wealthy urban patients, it will fail for rural patients with different dietary and environmental factors.
Genuine data sources include: Government databases (data.gov.in), open repositories (Kaggle, UCI Machine Learning Repository), research papers with published datasets, hospital records with consent, sensor networks, and verified social media APIs.
Practice Paper โ Chapter 2: AI Project Cycle
Class: 9โ10 | Max Marks: 30 | Time: 1 Hour
Section A โ MCQ (1 mark ร 8)
1. How many stages are in the AI Project Lifecycle? 1
- a) 3
- b) 4
- c) 5 โ
- d) 7
2. The "W" in 4W Canvas that asks about stakeholders is: 1
- a) What
- b) Who โ
- c) When
- d) Where
3. "Velocity" in Big Data refers to: 1
- a) Data size
- b) Data types
- c) Data speed โ
- d) Data quality
4. Which V of Big Data deals with fake and noisy data? 1
- a) Volume
- b) Velocity
- c) Variety
- d) Veracity โ
5. The first step in AI Project Cycle is: 1
- a) Collect data
- b) Define problem โ
- c) Build model
- d) Evaluate
6. Unstructured data is: 1
- a) Excel spreadsheet
- b) SQL database
- c) Videos and images โ
- d) CSV files
7. The goal of data acquisition is: 1
- a) Clean the data
- b) Gather relevant data for the AI problem โ
- c) Train the model
- d) Deploy AI
8. YouTube uploading 500 hours of video per minute is an example of: 1
- a) Veracity
- b) Value
- c) Volume โ
- d) Variety
Section B โ Short Answer (2 marks ร 4 = 8)
9. What is the purpose of the 4W Canvas in AI project planning? 2
10. Why is the AI project cycle iterative? Give a real-world reason. 2
11. Explain Variety and Value of Big Data with examples. 2
12. What is data acquisition? Name three sources of data. 2
Section C โ Long Answer (7 marks)
13. A school wants to build an AI system to predict student dropouts. (a) Apply the 4W Canvas to this problem. (b) What data would you need? (c) Which stage would be most challenging and why? 7
Section D โ Case Study (7 marks)
14. Zomato processes millions of food orders, reviews, GPS locations, restaurant data, and user preferences every day. (a) Which 5 V's does this represent? Give specific examples. (b) Why is this considered Big Data? (c) How can AI add value to this data? 7
3.1 What is Data Exploration (EDA)?
Exploratory Data Analysis (EDA) is the process of visually and statistically examining a dataset to understand its structure, find patterns, detect anomalies, and check assumptions before modelling.
Fig 3.1: Six-step EDA process before building any AI model
3.2 Data Preprocessing Techniques
Fig 3.2: Data quality issues in raw data and preprocessing solutions
3.3 Normalisation โ Why and How
AI algorithms (like KNN, Neural Networks) are distance-sensitive. If Age ranges from 0โ100 but Income ranges from 0โ1,000,000, the algorithm will unfairly favour Income. Normalisation rescales all features to the same range.
X_normalised = (X - X_min) / (X_max - X_min)
Result: All values scaled between 0 and 1
โข 150 โ (150-150)/(190-150) = 0/40 = 0.00
โข 160 โ (160-150)/(190-150) = 10/40 = 0.25
โข 170 โ (170-150)/(190-150) = 20/40 = 0.50
โข 180 โ (180-150)/(190-150) = 30/40 = 0.75
โข 190 โ (190-150)/(190-150) = 40/40 = 1.00
Insight: The shortest student (150cm) gets 0, tallest (190cm) gets 1.
โข Missing values โ fill with mean/median (imputation) or remove row
โข Outlier โ unusually extreme value (e.g. Age = 150)
โข Normalisation โ rescale features to 0โ1 range
โข Standardisation โ rescale features to mean=0, std=1 (Z-score)
Practice Paper โ Chapter 3: Data Exploration
Max Marks: 30 | Time: 1 Hour
Section A โ MCQ (1 mark ร 8)
1. EDA stands for: 1
- a) Estimated Data Analysis
- b) Exploratory Data Analysis โ
- c) Extended Data Algorithm
- d) None
2. Min-Max Normalisation scales values to: 1
- a) -1 to +1
- b) 0 to 100
- c) 0 to 1 โ
- d) Any range
3. An outlier is best described as: 1
- a) Most common value
- b) Extremely different value โ
- c) The median
- d) A missing value
4. Filling missing values with the average is called: 1
- a) Outlier removal
- b) Normalisation
- c) Mean imputation โ
- d) Encoding
5. Which chart shows frequency distribution? 1
- a) Pie chart
- b) Histogram โ
- c) Line graph
- d) Scatter plot
6. In df.describe(), which stat is NOT shown? 1
- a) Mean
- b) Std
- c) Mode โ
- d) Min/Max
7. Feature Engineering is: 1
- a) Removing features
- b) Creating new features from existing ones โ
- c) Labelling data
- d) Testing the model
8. Converting "Male/Female" to 0/1 is called: 1
- a) Normalisation
- b) Label Encoding โ
- c) Imputation
- d) Scaling
Section B โ Short Answer (3 marks ร 4 = 12)
9. A dataset has ages: [25, 28, 26, 150, 27]. Identify the outlier. How would you handle it? 3
10. Normalise the following heights using Min-Max: [140, 155, 170, 185, 200] 3
11. What are three common data quality issues? Give an example for each. 3
12. Explain the difference between normalisation and standardisation. 3
Section C โ Long Answer (10 marks)
13. A hospital dataset for heart disease prediction has: Missing blood pressure for 15% of patients; Age outlier (patient listed as 205 years); Gender stored as M/F; Blood pressure in mmHg (range 80โ200). (a) Identify all preprocessing steps needed. (b) How would you handle each issue? (c) Why is it critical to clean this data before training? (d) Write pseudocode for two preprocessing steps. 10
4.1 Types of Machine Learning
Fig 4.1: Three paradigms of Machine Learning
4.2 Decision Tree โ Visual Example
Fig 4.2: Decision Tree โ hierarchical if-then-else rules for classification
4.3 ML Algorithm Comparison
| Algorithm | Type | How it Works | Best Used For | Example |
|---|---|---|---|---|
| Linear Regression | Supervised | Fits a line y=mx+b to data | Predicting numbers | Predict house price |
| Decision Tree | Supervised | Series of if-else questions | Classification | Disease diagnosis |
| KNN | Supervised | Find K nearest neighbours | Classification | Recommend movies |
| Naive Bayes | Supervised | Probability with Bayes theorem | Text classification | Spam detection |
| K-Means | Unsupervised | Group data into K clusters | Customer segmentation | Marketing groups |
| Neural Network | Supervised | Layers of connected neurons | Images, speech | Face recognition |
โข Unsupervised: Data has no labels โ AI finds patterns on its own
โข Reinforcement: Agent learns from environment through trial and error
4.4 Supervised Learning โ Deep Dive
Supervised Learning is like studying with a teacher who shows you both the question AND the correct answer. The AI learns to map inputs to outputs using labelled training data.
Fig 4.3: Supervised Learning โ how it works, with Classification vs Regression explained visually
4.5 Unsupervised Learning โ Deep Dive
Unsupervised Learning works without any labels. The AI explores data on its own and discovers hidden patterns, groupings, or structures. Think of it like organising a messy room โ no one tells you how, you figure it out.
Fig 4.4: Unsupervised Learning โ Clustering groups similar data; Dimensionality Reduction simplifies complex data
4.6 Reinforcement Learning โ Deep Dive
Fig 4.5: Reinforcement Learning โ agent takes action, gets reward/penalty, improves strategy
4.7 Deep Learning vs Traditional Machine Learning
Fig 4.6: Deep Learning vs Traditional ML โ when to use which approach
4.8 Interactive โ Match the Scenario to the Learning Technique
Click each scenario card to reveal which ML technique is the best fit and why. This is great exam practice!
โข Labelled data + predict a number โ Supervised + Regression (price, temperature, marks)
โข No labels + find groups โ Unsupervised + Clustering (customer segments, news topics)
โข No labels + reduce complexity โ Unsupervised + Dimensionality Reduction (PCA)
โข Agent learning in environment โ Reinforcement Learning (games, robots, trading)
Practice Paper โ Chapter 4: Modelling
Max Marks: 30 | Time: 1 Hour
Section A โ MCQ (1 mark ร 8)
1. In supervised learning, training data is: 1
- a) Unlabelled
- b) Labelled โ
- c) Randomly generated
- d) Always numeric
2. KNN stands for: 1
- a) Kernel Neural Network
- b) K Nearest Neighbours โ
- c) K Node Network
- d) None
3. K-Means is an example of: 1
- a) Supervised learning
- b) Reinforcement learning
- c) Unsupervised learning โ
- d) Deep learning
4. A Decision Tree uses: 1
- a) Mathematical equations
- b) If-else rules โ
- c) Probability tables
- d) Distance calculations
5. Linear Regression predicts: 1
- a) A category
- b) A label
- c) A continuous number โ
- d) A cluster
6. AlphaGo (Google DeepMind) uses: 1
- a) Supervised Learning
- b) Unsupervised Learning
- c) Reinforcement Learning โ
- d) Linear Regression
7. Overfitting means the model: 1
- a) Performs well on new data
- b) Memorises training data but fails on new data โ
- c) Has too few parameters
- d) None of above
8. The leaf nodes of a Decision Tree represent: 1
- a) Questions
- b) Features
- c) Final decisions/predictions โ
- d) Training data
Section B โ Short Answer (3 marks ร 4 = 12)
9. Draw and explain a simple Decision Tree for "Should I buy an umbrella?" 3
10. Explain the difference between classification and regression with examples. 3
11. What is overfitting? How can it be prevented? 3
12. Explain how KNN works with a step-by-step example for K=3. 3
Section C โ Case Study + HOTS (10 marks)
13. A bank wants to detect fraudulent transactions. (a) Which type of ML is best and why? (b) What features/data would you use? (c) Draw a simple decision tree for fraud detection. (d) What challenges might this model face? (e) How would you evaluate if the model is working well? 10
5.1 The Confusion Matrix โ Your AI Report Card
After training an AI model, we must test how well it performs. The confusion matrix is a 2ร2 table that shows where your AI got confused โ predicting correctly, predicting wrong in different ways.
Fig 5.1: Confusion Matrix for Cancer AI (200 patients)
5.1b Understanding TP, FP, TN, FN โ With Everyday Examples
The four cells of the confusion matrix each represent a different type of outcome. Understanding them deeply is the key to choosing the right metric for the right problem.
Fig 5.1b: TP, FP, TN, FN โ explained with medical and spam examples for both types of outcomes
The SECOND word (Positive/Negative) tells you what the AI predicted.
โข True Positive = AI said YES and it WAS yes โ โ Good!
โข True Negative = AI said NO and it WAS no โ โ Good!
โข False Positive = AI said YES but it was actually NO โ โ False alarm!
โข False Negative = AI said NO but it was actually YES โ โ Missed case! (most dangerous in medical AI)
5.1c Why Evaluation Matters โ Real Consequences
Fig 5.1c: Why AI evaluation is critical โ real consequences of unevaluated AI systems
5.2 Four Evaluation Metrics โ Calculated
โ Accuracy = (TP + TN) / Total = (85 + 80) / 200 = 165/200 = 82.5%
โ "82.5% of all predictions were correct"
โก Precision = TP / (TP + FP) = 85 / (85 + 20) = 85/105 = 80.95%
โ "When AI said cancer, it was right 81% of the time"
โข Recall (Sensitivity) = TP / (TP + FN) = 85 / (85 + 15) = 85/100 = 85%
โ "AI found 85% of all actual cancer cases"
โฃ F1 Score = 2 ร (Precision ร Recall) / (Precision + Recall) = 2 ร (0.8095 ร 0.85) / (0.8095 + 0.85) = 82.92%
โ "Balanced measure combining precision and recall"
Fig 5.2: Precision vs Recall โ the trade-off and when each matters
Accuracy = (TP + TN) / (TP + TN + FP + FN)
Precision = TP / (TP + FP)
Recall = TP / (TP + FN)
F1 Score = 2 ร (Precision ร Recall) / (Precision + Recall)
โข TN = Correct negative prediction (best!)
โข FP = False alarm โ predicted YES but was NO (Type I error)
โข FN = Missed case โ predicted NO but was YES (Type II error) โ most dangerous in medical AI!
โข Accuracy paradox: If 99% of emails are ham, a model that predicts EVERYTHING as ham gets 99% accuracy but is useless!
Practice Paper โ Chapter 5: Evaluation
Max Marks: 30 | Time: 1 Hour
Section A โ MCQ (1 mark ร 8)
1. TP + TN + FP + FN equals: 1
- a) Precision
- b) Recall
- c) Total samples โ
- d) F1 Score
2. Which error is most dangerous in cancer detection AI? 1
- a) False Positive
- b) False Negative โ
- c) True Positive
- d) True Negative
3. F1 Score is the harmonic mean of: 1
- a) Accuracy and Recall
- b) Precision and Recall โ
- c) TP and TN
- d) FP and FN
4. An AI spam filter should prioritise: 1
- a) Recall
- b) Precision โ
- c) F1 Score
- d) Accuracy
5. If TP=80, FP=20, what is Precision? 1
- a) 0.6
- b) 0.7
- c) 0.8 โ
- d) 0.9
6. The Accuracy Paradox occurs when: 1
- a) Accuracy = 100%
- b) High accuracy but poor performance on minority class โ
- c) Recall is high
- d) F1 > Accuracy
7. Recall is also known as: 1
- a) Specificity
- b) Sensitivity โ
- c) Precision
- d) Accuracy
8. Which metric is best when both FP and FN are important? 1
- a) Accuracy
- b) Precision
- c) Recall
- d) F1 Score โ
Section B โ Calculation (12 marks)
9. An AI tested on 500 patients gave: TP=140, FN=60, FP=30, TN=270. Calculate Accuracy, Precision, Recall and F1 Score. Show all working. 6
10. For a security screening AI, should you prioritise Precision or Recall? Justify with reasoning. 3
11. Explain the Accuracy Paradox with a numerical example. 3
Section C โ Case Study (10 marks)
12. A COVID-19 detection AI gave these results on 1000 patients: TP=850, FN=150, FP=200, TN=800. (a) Draw and fill the confusion matrix. (b) Calculate all 4 metrics. (c) Is this a good model? (d) What type of error is more dangerous here? (e) How would you improve recall without harming precision too much? 10
1.1 Why Ethics in AI Matters
AI systems make decisions that affect real human lives โ hiring, lending, medical diagnosis, criminal justice. If these systems are biased or unethical, they can cause serious harm at massive scale.
1.2 The 7 Principles of Ethical AI
Fig T1.1: The 7 interconnected principles of ethical AI development
1.3 Types of AI Bias โ With Real Examples
| Bias Type | What It Means | Real-World Example | Harm Caused |
|---|---|---|---|
| Historical Bias | Data reflects past discrimination | AI trained on old hiring data โ men dominated tech โ AI prefers men | Women unfairly rejected (Amazon, 2018) |
| Representation Bias | Some groups under-represented in training data | Face recognition trained mostly on light-skinned faces | Dark skin faces misidentified 35% more (MIT study) |
| Measurement Bias | Wrong/noisy data collected | Health AI trained on hospital data โ uninsured people under-represented | Wrong predictions for lower-income groups |
| Aggregation Bias | One model for all groups when separate models needed | Diabetes AI trained on average data โ different across ethnicities | Poor outcomes for minorities |
| Evaluation Bias | Tested on non-representative benchmark | NLP tested only on English โ bad performance on Hindi, Bengali | Excludes non-English speakers |
| Deployment Bias | Used differently from how it was designed | Crime prediction AI designed for resource allocation, used for sentencing | Unjust imprisonment |
1.4 How GANs Create Deepfakes
Fig T1.2: GAN architecture showing how deepfakes are generated through adversarial training
1.5 Security, Privacy, Accountability & Transparency โ How They Connect to Ethics
Fig T1.3: Four pillars interconnected โ each directly supports AI Ethics
| Pillar | What it means in AI | If violated โ real harm | Indian Context |
|---|---|---|---|
| ๐ Security | Protect AI from hackers, adversarial attacks, data poisoning | Self-driving cars hacked; medical AI manipulated to mis-diagnose | CERT-In governs cybersecurity; AI in banking must be secured |
| ๐ Privacy | Users must consent to how their data is collected and used | Cambridge Analytica: 87M Facebook profiles used without consent | India's DPDP Act 2023 protects digital personal data |
| ๐ Accountability | Someone must be responsible when AI causes harm | An AI loan system denies thousands unfairly โ who is responsible? | RBI guidelines: banks responsible for AI-based credit decisions |
| ๐ Transparency | AI must explain decisions in understandable terms (XAI) | A "black box" AI rejects job applications โ no reason given | EU AI Act (2024) mandates explanation rights for AI decisions |
1.6 Interactive Ethics Scenario Game โ What Would You Do?
Click each scenario card to reveal the ethical analysis. Think first โ then check!
โข GAN (Generative Adversarial Network) creates deepfakes: Generator vs Discriminator
โข XAI (Explainable AI) โ making AI decisions understandable to humans
โข GDPR (EU), DPDP Act 2023 (India), EU AI Act 2024 โ laws governing AI data use
โข Security = protect AI from attack | Privacy = protect user data | Accountability = own the consequences | Transparency = explain decisions
โข Always ask: Who benefits? Who is harmed? Is this fair? Who is responsible?
Practice Paper โ T1: AI Ethics & Bias
Max Marks: 40 | Time: 1.5 Hours
Section A โ MCQ (1 mark ร 10)
1. Which principle means AI results must be explainable? 1
- a) Privacy
- b) Reliability
- c) Transparency โ
- d) Inclusivity
2. Amazon's AI hiring tool was shut down because it: 1
- a) Was too expensive
- b) Discriminated against women โ
- c) Didn't work at all
- d) Was too slow
3. A deepfake is created using: 1
- a) Decision Trees
- b) Linear Regression
- c) GAN โ
- d) KNN
4. XAI stands for: 1
- a) Extra AI
- b) Explainable AI โ
- c) Extreme AI
- d) Extended AI
5. GDPR is a data protection law from: 1
- a) USA
- b) China
- c) India
- d) European Union โ
6. Representation bias occurs when: 1
- a) Data is too large
- b) Some groups are under-represented in training data โ
- c) AI is too accurate
- d) Model is too simple
7. In a GAN, the Discriminator's role is to: 1
- a) Generate fake images
- b) Tell real from fake โ
- c) Train the generator
- d) Store data
8. Which AI principle ensures it works for ALL users? 1
- a) Privacy
- b) Inclusivity โ
- c) Transparency
- d) Accountability
9. Algorithmic accountability means: 1
- a) AI should make profit
- b) AI developers are responsible for their system's impact โ
- c) AI must be free
- d) AI must be open source
10. Historical bias in AI comes from: 1
- a) Future predictions
- b) Programming errors
- c) Past discriminatory data โ
- d) Hardware failures
Section B โ Short Answer (3 marks ร 5 = 15)
11. Explain "Fairness" in AI with a real example where AI was unfair. 3
12. What is a deepfake? How can it be harmful? Give two examples. 3
13. Explain the difference between Historical Bias and Representation Bias. 3
14. What is XAI? Why is it important for medical AI? 3
15. Name and explain any three global AI ethics frameworks. 3
Section C โ HOTS + Case Study (15 marks)
16. (HOTS) A criminal justice AI predicts a person is 85% likely to reoffend. The judge uses this to deny bail. (a) What type of AI bias might this system have? (b) Which ethical principles are violated? (c) Should judges use AI for sentencing? (d) How would you make this system more ethical? 8
17. (Case Study) A social media company's content moderation AI removes posts about political protests. It works well in English but removes legitimate posts in Hindi and Bengali. (a) What type of bias does this show? (b) How was this bias introduced? (c) How should the company fix this? (d) What laws might apply? 7
2.1 Types of Data
Fig T2.1: Complete data type taxonomy โ understanding your data type is the first step
2.2 Structured vs Unstructured Data
Fig T2.2: Structured vs Unstructured data โ 80% of world's data is unstructured!
2.3 Data Quality โ 6 Dimensions
Accuracy
Data reflects reality. Error: Age listed as 250
Timeliness
Data is current. Error: 10-year-old census for AI
Consistency
Same format throughout. Error: Date as DD/MM vs MM/DD
Completeness
No missing values. Error: Empty blood pressure column
Relevance
Data fits the problem. Error: Hair colour for cancer diagnosis
Uniqueness
No duplicates. Error: Same record saved twice
2.4 Big Data โ The 5 V's
Fig T2.3: Big Data 5 V's โ each dimension creates unique challenges for AI systems
2.5 Why Data Literacy is Important
Fig T2.4: Data literacy = being able to read, work with, analyse, and communicate data effectively
In the AI age, data literacy is as important as reading and writing. A citizen who cannot critically evaluate data can be misled by fake news, manipulated statistics, and biased AI. Every student, voter, and employee needs these skills.
2.6 Sufficient vs Insufficient Data โ Why It Matters
Fig T2.5: The cascade from insufficient data to dangerous AI; Overfitting vs Underfitting patterns
| Overfitting | Underfitting | |
|---|---|---|
| Definition | Model memorises training data โ too specific | Model too simple โ misses patterns |
| Training Accuracy | Very high (95%+) | Very low (<60%) |
| Test/New Data Accuracy | Very low โ fails on unseen data | Also very low โ misses everything |
| Cause | Too little data OR too complex model | Too simple model OR too few features |
| Fix | More data, regularisation, cross-validation | Better features, more complex model |
| Analogy | Student who memorises answers but can't solve new problems | Student who didn't study โ guesses randomly |
2.7 Why Train-Test Split is Important
Fig T2.6: Train-test split โ the only honest way to measure AI performance
2.8 Genuine Sources of Data for AI Projects
Government Portals
data.gov.in (India), data.gov (USA), WHO, World Bank โ official, reliable, free
Research Repositories
Kaggle, UCI ML Repository, Google Dataset Search โ curated ML datasets
Open Data Projects
OpenStreetMap, Wikipedia, Common Crawl โ web-scale open datasets
Domain-Specific Sets
NIH (health), ImageNet (vision), LibriSpeech (audio), COCO (objects)
What to AVOID
Scraped social media without consent, purchased datasets from unknown vendors, data with no provenance or documentation
Sufficient Data Checklist
Diverse? Representative? Large enough? Labelled correctly? Legally collected? Updated? Balanced across classes?
โข 80% of world data is unstructured โ photos, videos, text
โข Big Data 5 Vs: Volume, Velocity, Variety, Veracity, Value
โข Overfitting = memorises training data, fails on new data (like a student who only memorises)
โข Underfitting = too simple, misses patterns in both training and test data
โข Train:Test split = typically 80:20 โ never test on training data (cheating!)
โข Sufficient data = diverse, representative, correctly labelled, legally obtained
โข Poor data quality = poor AI. "Garbage in, Garbage out" (GIGO)
Practice Paper โ T2: Data Literacy
Max Marks: 40 | Time: 1.5 Hours
Section A โ MCQ (1 mark ร 10)
1. Which is an example of discrete data? 1
- a) Temperature
- b) Number of students โ
- c) Height
- d) Weight
2. "Blood group (A, B, AB, O)" is what type of data? 1
- a) Continuous
- b) Discrete
- c) Ordinal
- d) Nominal โ
3. Veracity in Big Data means: 1
- a) Data speed
- b) Data size
- c) Data quality/trustworthiness โ
- d) Data types
4. "1st, 2nd, 3rd place" is what type of data? 1
- a) Nominal
- b) Ordinal โ
- c) Continuous
- d) Discrete
5. What percentage of world's data is unstructured? 1
- a) 20%
- b) 40%
- c) 60%
- d) 80% โ
6. Which of these is structured data? 1
- a) Twitter posts
- b) Videos
- c) SQL database โ
- d) MRI scans
7. GIGO stands for: 1
- a) Get It Go Out
- b) Garbage In, Garbage Out โ
- c) Get Input, Get Output
- d) None
8. Data labelling is important because: 1
- a) It makes data colourful
- b) It provides answers for supervised learning โ
- c) It removes outliers
- d) It speeds up training
9. A bar chart that looks dramatic with a cropped Y-axis starting at 90 is an example of: 1
- a) Good data visualisation
- b) Misleading statistics โ
- c) Normalisation
- d) Clustering
10. Which quality dimension ensures no duplicate records? 1
- a) Accuracy
- b) Completeness
- c) Uniqueness โ
- d) Timeliness
Section B โ Short Answer (3 marks ร 5 = 15)
11. Classify these as Structured/Unstructured: (a) School report card (b) Doctor's voice note (c) Bank transaction record (d) Instagram reel (e) Weather CSV file 3
12. Explain three ways a chart or graph can be misleading. Give examples. 3
13. What is data labelling? Who labels data and how? Give an example. 3
14. Apply all 5 V's of Big Data to WhatsApp. 3
15. Explain Relevance as a data quality dimension. Give an example where irrelevant data harmed an AI model. 3
Section C โ Case Study + HOTS (15 marks)
16. (Case Study) A health insurance company uses AI to set premium prices. The AI was trained on patient data from 2010. It includes: age, gender, BMI, past claims, PIN code, and social media activity. (a) What data quality issues might this dataset have? (b) Is social media activity relevant? Why/why not? (c) What bias could the PIN code introduce? (d) How should the data be improved? (e) What are the ethical concerns? 8
17. (HOTS) A student says: "AI trained on 10 million data points must be accurate!" Critically evaluate this statement. Is more data always better? 7
3.1 Probability โ The Language of Uncertainty
3.2 Bayes' Theorem โ The Core of AI Probability
P(A|B) = P(B|A) ร P(A) / P(B)
Where: P(A|B) = Probability of A given B has occurred
โข P(Spam) = 0.40, P(Not Spam) = 0.60
โข P(Lottery|Spam) = 0.80, P(Lottery|Not Spam) = 0.05
โข P(Lottery) = (0.80 ร 0.40) + (0.05 ร 0.60) = 0.32 + 0.03 = 0.35
โข P(Spam|Lottery) = (0.80 ร 0.40) / 0.35 = 0.32 / 0.35 = 91.4%
โ A 91.4% chance the email is spam! The filter blocks it.
3.3 Statistics โ Five Essential Measures
3.4 Images as Matrices โ How AI Sees Images
Fig T3.1: Images as numerical matrices โ the foundation of Computer Vision AI
3.5 Measures of Central Tendency โ With Real-World Examples
Fig T3.2: Three measures of central tendency โ each suited to different types of AI problems
3.6 Probability with Real Examples โ Used Everywhere in AI
Fig T3.3: Probability is the mathematical language AI uses to express confidence in its predictions
3.7 Why Mathematics is Essential for AI โ Core ML & Deep Learning
Fig T3.4: Mathematics is the engine of AI โ from basic ML to the most advanced deep learning
โข Mean = Sum of all values รท Number of values (sensitive to outliers)
โข Median = Middle value after sorting (use when data is skewed, e.g. salaries)
โข Mode = Most frequently occurring value (use for categories, e.g. shoe sizes)
โข Std Dev = Measure of spread from mean
โข Correlation (r): +1 = perfect positive, โ1 = perfect negative, 0 = no relation
โข Math powers AI: Linear Algebra โ data storage | Calculus โ training | Probability โ predictions
Practice Paper โ T3: Maths for AI
Max Marks: 40 | Time: 1.5 Hours
Section A โ MCQ (1 mark ร 10)
1. In grayscale images, 255 represents: 1
- a) Black
- b) Red
- c) White โ
- d) Blue
2. What does a correlation of -0.9 indicate? 1
- a) Strong positive correlation
- b) No correlation
- c) Strong negative correlation โ
- d) Weak positive
3. Bayes' theorem is used in: 1
- a) Image processing only
- b) Spam filtering and medical diagnosis โ
- c) Physical robots
- d) None
4. For the dataset {3,3,3,5,7}, which statistical measure is DIFFERENT from median? 1
- a) Mode (3)
- b) Mean (4.2)
- c) Median (3)
- d) None โ (Mode=Median=3, Mean differs)
5. A 200ร200 pixel RGB image has how many numbers? 1
- a) 40,000
- b) 80,000
- c) 120,000 โ
- d) 200,000
6. P(A|B) in Bayes' theorem means: 1
- a) P(A) and P(B) multiplied
- b) Probability of A given B โ
- c) P(A) divided by P(B)
- d) P(B) given A
7. Standard deviation measures: 1
- a) The centre of data
- b) The spread of data โ
- c) The most common value
- d) The maximum value
8. If 60% of spam emails say "Win" and 2% of genuine emails say "Win", and spam is 30% of all email, P(Win|Spam)รP(Spam) = 1
- a) 0.18 โ
- b) 0.60
- c) 0.02
- d) 0.30
9. The "average" is the same as the: 1
- a) Median
- b) Mode
- c) Mean โ
- d) Range
10. In matrix multiplication, a 3ร4 matrix times a 4ร2 matrix gives: 1
- a) 3ร4 matrix
- b) 4ร2 matrix
- c) 3ร2 matrix โ
- d) 4ร4 matrix
Section B โ Calculation (15 marks)
11. Dataset: {12, 15, 15, 18, 20, 22, 25, 30}. Find: (a) Mean (b) Median (c) Mode (d) Range (e) Which average best describes this data and why? 5
12. A disease test is 90% accurate. 1% of the population has the disease. If a person tests positive, what is the probability they actually have the disease? Show full Bayes' calculation. 5
13. A 5ร5 greyscale image has pixel values as follows: Row1=[10,20,30,20,10]. Describe what this might look like visually. If we apply a filter [[1,0,-1],[1,0,-1],[1,0,-1]] what operation is being done? 5
Section C โ HOTS (10 marks)
14. A school tests a COVID screening AI. Out of 1000 students: 10 actually have COVID. The AI correctly identifies 8 of them (TP=8). It also flags 50 healthy students as COVID-positive (FP=50). (a) Calculate P(Actually COVID | AI says COVID). (b) Is this AI useful for school screening? (c) What does the "base rate" problem teach us about AI in rare-disease screening? 10
4.1 What is Generative AI?
Generative AI creates new content โ text, images, audio, video, code โ that didn't exist before. Unlike traditional AI (which classifies or predicts), GenAI generates.
4.2 How Large Language Models Work
Fig T4.1: How Large Language Models process and generate text
4.3 GAN โ Forger vs Detective
Fig T4.2: GAN adversarial training โ both networks improve each other through competition
4.4 AI Hallucinations
4.5 Generative AI vs Traditional AI โ Complete Comparison
Fig T4.3: Generative AI creates new content; Traditional AI classifies or predicts from existing data
4.6 Recent Developments in Generative AI (2022โ2025)
Fig T4.4: The rapid evolution of Generative AI โ 2022 to 2025
4.7 Ethical Concerns of Generative AI
Deepfakes & Misinformation
GANs create fake videos of politicians, celebrities. Used to spread false news, manipulate elections. 96% of deepfakes are non-consensual.
Academic Dishonesty
Students use ChatGPT for homework, essays. AI-written assignments indistinguishable from human ones. Schools banning AI tools globally.
Copyright & Plagiarism
AI trained on artists' work without consent. Midjourney/DALL-E replicate styles. Artists suing OpenAI and Stability AI.
Job Displacement
Writers, graphic designers, call centre agents at risk. Goldman Sachs: 300M jobs may be impacted by generative AI by 2030.
Environmental Cost
Training GPT-4 used estimated 1,000+ MWh of electricity โ same as 100+ households for a year. AI's carbon footprint is growing rapidly.
Hallucination Risk
AI confidently states false "facts." Dangerous in medical, legal, scientific contexts. Users may not verify AI-generated information.
โข GAN = Generative Adversarial Network โ Generator vs Discriminator
โข Diffusion Model โ Stable Diffusion, DALL-E โ removes noise to create images
โข Hallucination โ AI confidently states false information
โข Prompt Engineering โ crafting good inputs to get good AI outputs
โข Generative AI creates NEW content; Traditional AI predicts/classifies from existing data
โข Key ethical concerns: Deepfakes, copyright, academic dishonesty, job loss, hallucinations
Practice Paper โ T4: Generative AI
Max Marks: 40 | Time: 1.5 Hours
Section A โ MCQ (1 mark ร 10)
1. ChatGPT is an example of: 1
- a) Computer Vision AI
- b) Large Language Model โ
- c) Reinforcement Learning
- d) Decision Tree
2. In a GAN, which network creates fake content? 1
- a) Discriminator
- b) Generator โ
- c) Transformer
- d) Encoder
3. AI Hallucination means: 1
- a) AI dreaming
- b) AI generating images
- c) AI confidently stating false facts โ
- d) AI being slow
4. Tokenisation in LLMs means: 1
- a) Removing stop words
- b) Breaking text into numerical units โ
- c) Translating text
- d) Summarising text
5. Diffusion models create images by: 1
- a) Combining existing images
- b) Starting with noise and removing it step by step โ
- c) Using decision trees
- d) Copying from the internet
6. DALL-E creates: 1
- a) Text responses
- b) Music
- c) Images from text prompts โ
- d) Video
7. Prompt Engineering is: 1
- a) Building AI hardware
- b) Crafting effective inputs to get good AI outputs โ
- c) Training AI models
- d) Cleaning data
8. LLM stands for: 1
- a) Learning Language Module
- b) Large Language Model โ
- c) Linear Learning Machine
- d) Logical Logic Machine
9. A GAN is considered "converged" when: 1
- a) Generator always wins
- b) Discriminator can't tell real from fake โ
- c) Training stops
- d) Images become blurry
10. Which application uses Generative AI? 1
- a) Spam detection
- b) AI-generated music โ
- c) Price prediction
- d) Customer clustering
Section B โ Short Answer (3 marks ร 5 = 15)
11. Explain the difference between a GAN Generator and Discriminator with the "forger and detective" analogy. 3
12. What is an AI hallucination? Give two real examples and explain why it's dangerous. 3
13. Explain how LLMs generate text token by token. 3
14. Compare Generative AI (ChatGPT) and Traditional AI (spam filter) in terms of purpose, input, output and training. 3
15. What is a diffusion model? How is it different from a GAN? 3
Section C โ Case Study + HOTS (15 marks)
16. (Case Study) A news website uses AI to generate articles. The AI writes convincing fake stories about celebrity deaths and political events that never happened. (a) What AI technology is being misused? (b) What is this phenomenon called? (c) How can readers identify AI-generated fake news? (d) What are the societal harms? (e) How should this be regulated? 8
17. (HOTS) If AI can generate code, essays, music, and art better than humans, what happens to human creativity? Is this a threat or an opportunity? Argue both sides. 7
5.1 The Biological Inspiration
Fig T5.1: Biological neuron mapped to its artificial counterpart โ the perceptron
5.2 Activation Functions โ Visual Comparison
Fig T5.2: Three key activation functions โ their shapes and use cases
5.3 Convolutional Neural Network (CNN) โ Layer by Layer
Fig T5.3: CNN architecture โ Conv โ ReLU โ MaxPool โ Flatten โ Dense โ Softmax pipeline
Weights: wโ = 0.7, wโ = 0.4, wโ = 0.3, bias b = -0.6
Step 1: Weighted sum z = (0.7ร0.8) + (0.4ร0.6) + (0.3ร0.5) + (-0.6) = 0.56 + 0.24 + 0.15 - 0.6 = 0.35
Step 2: Apply Sigmoid: ลท = 1/(1+e^(-0.35)) = 1/(1+0.705) = 0.587
Step 3: If ลท > 0.5 โ predict HIGH RISK โ Patient flagged for diabetes screening โ ๏ธ
5.4 ANN Architecture โ Full Deep Network Diagram
Fig T5.4: Full ANN โ information flows left to right through interconnected layers of neurons
5.5 Forward Propagation & Backpropagation โ The Training Cycle
Fig T5.5: Forward propagation predicts; backpropagation corrects โ together they teach the network
Without backpropagation, a neural network makes random predictions forever and never learns. Backprop is what makes deep learning possible โ it's the algorithm that tells each individual weight in a network of millions of weights exactly how much it contributed to the error and by exactly how much it should change. Geoffrey Hinton (Nobel Prize 2024) pioneered its use in neural networks in the 1980s. Modern GPT-4 has 1.76 trillion parameters โ all trained by backpropagation.
5.6 Neural Networks vs CPU Processing โ Why GPUs & Why Deep Learning?
Fig T5.6: CPU handles sequential tasks; GPU enables parallel deep learning; Neural Networks learn complex hierarchical features
โข Sigmoid: f(z) = 1/(1+e^(-z)) โ output 0 to 1
โข ReLU: f(z) = max(0, z)
โข Forward Propagation: Input โ compute z โ apply f(z) โ get prediction ลท
โข Backpropagation: Measure error (Loss) โ compute gradients โ update weights (w = w โ ฮท ร โL/โw)
โข Error = Loss = how wrong AI is. Backprop minimises it over thousands of iterations
โข GPU is essential for deep learning โ parallel processing of millions of matrix operations
โข CNN: Convolution โ ReLU โ MaxPool โ Flatten โ Dense โ Softmax
Practice Paper โ T5: Neural Networks
Max Marks: 40 | Time: 1.5 Hours
Section A โ MCQ (1 mark ร 10)
1. The biological equivalent of an artificial neuron's "weights" is: 1
- a) Dendrites
- b) Axon
- c) Synaptic strength โ
- d) Cell body
2. Which activation function outputs values between 0 and 1? 1
- a) ReLU
- b) Tanh
- c) Sigmoid โ
- d) Softmax
3. Max Pooling in CNN is used to: 1
- a) Add more layers
- b) Reduce spatial dimensions โ
- c) Increase resolution
- d) Apply activation
4. Backpropagation is used to: 1
- a) Add more data
- b) Update weights using error signal โ
- c) Build the network
- d) Normalise inputs
5. In a CNN, convolution filters detect: 1
- a) Data types
- b) Edges, textures and patterns โ
- c) Output labels
- d) Network depth
6. The "Flatten" layer in CNN converts: 1
- a) 1D to 2D
- b) 2D feature maps to 1D vector โ
- c) RGB to grayscale
- d) Input to output
7. ReLU outputs: 1
- a) -1 to +1
- b) 0 to 1
- c) max(0, x) โ
- d) Always 0 or 1
8. A perceptron that outputs 0.73 with threshold 0.5 predicts: 1
- a) Class 0 (negative)
- b) Class 1 (positive) โ
- c) Uncertain
- d) Error
9. Softmax is used in the _____ layer of a classification CNN: 1
- a) First
- b) Middle
- c) Last/Output โ
- d) Pooling
10. "Deep" in Deep Learning refers to: 1
- a) Large datasets
- b) Many hidden layers โ
- c) Complex maths
- d) High accuracy
Section B โ Calculation + Short Answer (15 marks)
11. A perceptron has inputs xโ=0.5, xโ=0.8 with weights wโ=0.6, wโ=0.4 and bias b=-0.3. Calculate: (a) Weighted sum z (b) Sigmoid output (c) ReLU output (d) Prediction if threshold=0.5 5
12. Draw and label the 6 layers of a CNN architecture. What does each layer do? 5
13. Explain backpropagation in simple terms. Why is it important? 5
Section C โ HOTS + Case Study (15 marks)
14. A CNN is trained to detect tumours in X-rays. (a) What would the Input Layer look like? (b) What patterns would early Convolution layers detect? (c) What would late layers detect? (d) Why is Recall more important than Precision here? (e) What happens if the model is overtrained (overfitting)? 8
15. Compare a biological brain and a deep neural network on: number of neurons, energy consumption, learning speed, creativity, and reliability. Which is better for which tasks? 7
Topic 6: Natural Language Processing (NLP)
NLP is the branch of AI that helps computers understand, interpret, and generate human language. From chatbots to translation apps โ NLP is everywhere.
Humans understand "I saw the man with the telescope" instantly โ but a computer struggles because the sentence has TWO meanings (did you use the telescope, or did the man have it?). Language is ambiguous, contextual, and cultural.
1. The 7-Step NLP Pipeline
Every NLP system processes text through a series of steps. Think of it like a factory conveyor belt โ raw text enters one end, structured understanding comes out the other.
Fig 6.1 โ The complete NLP pipeline: raw text is progressively cleaned, tagged, and converted to numbers for ML models to process.
2. Tokenisation โ Splitting Text into Units
"I love AI" โ ["I", "love", "AI"] โ splits on spaces
"Hello. How are you?" โ ["Hello.", "How are you?"]
"AI" โ ["A", "I"] โ used in some deep learning models
"unhappiness" โ ["un", "happi", "ness"] โ used in BERT, GPT
3. Stop Words โ The "Noise" Words
Stop words are common words that carry little meaning and are removed to reduce noise. Think of them as the "filler" words in language.
Fig 6.2 โ Stop word removal reduces noise and helps models focus on meaningful content words.
4. Stemming vs Lemmatisation
Fig 6.3 โ Stemming chops word endings; lemmatisation looks up the true root form. Lemmatisation is more accurate but computationally expensive.
5. TF-IDF โ How Important Is a Word?
TF-IDF (Term FrequencyโInverse Document Frequency) measures how important a word is to a specific document compared to all documents.
TF(word) = (count of word in doc) รท (total words in doc)IDF(word) = log(total docs รท docs containing word)TF-IDF = TF ร IDF
Fig 6.4 โ TF-IDF rewards words that appear frequently in one document but rarely across all documents, identifying truly distinctive terms.
6. Word Embeddings โ Words as Vectors
Word2Vec converts words to numerical vectors so that similar words cluster together in a mathematical space. The famous example: King โ Man + Woman = Queen
Fig 6.5 โ In Word2Vec space, "King" and "Queen" are close (both royalty) but separated along the gender dimension. Words with similar meanings cluster together.
7. Transformers โ The Power Behind ChatGPT
Transformers revolutionised NLP with self-attention โ the ability to consider every word's relationship to every other word simultaneously.
Fig 6.6 โ Transformers use self-attention to understand which words relate to each other. "it" in a sentence about a tired cat correctly refers to the cat, not the mat.
8. Real-World NLP Applications
Amazon reviews: "Amazing product!" โ POSITIVE (0.94). Used in brand monitoring and customer feedback.
Google Translate: Hindi โ English โ French using seq2seq transformer models.
Siri, Alexa, Google Assistant use NLU (understanding) + NLG (generation).
Converting a 10-page report into a 5-sentence summary automatically.
Google uses BERT to understand search intent beyond exact keyword matching.
Gmail classifies millions of emails as spam/not-spam using NLP + ML.
Corpus โ Large collection of text used for training | Token โ Individual unit (word/character) | Vocabulary โ All unique tokens in corpus | Embedding โ Numerical vector representation of a word | Attention โ Mechanism to focus on relevant parts of input | BERT โ Bidirectional Encoder Representations from Transformers | GPT โ Generative Pre-trained Transformer
Practice Paper โ Topic 6: Natural Language Processing
CBSE Class 9โ10 | Code 417 | Full Marks: 40 | Time: 1.5 hrs
Section A โ Multiple Choice Questions (1 mark each)
- Which NLP step splits text into individual words or sentences?
(a) Stemming (b) Tokenisation (c) Lemmatisation (d) NER - What does TF stand for in TF-IDF?
(a) Text Feature (b) Total Frequency (c) Term Frequency (d) Token Form - Which of the following is a stop word?
(a) quantum (b) algorithm (c) the (d) neural - Word2Vec converts words into:
(a) Images (b) Numerical vectors (c) Sentences (d) Audio signals - Transformers were introduced in the paper:
(a) "ImageNet" (b) "Attention is All You Need" (c) "AlphaGo" (d) "Deep Blue"
Section B โ Short Answer Questions (2 marks each)
- Differentiate between stemming and lemmatisation with one example each. [2]
- What is a "corpus" in NLP? Give one example of a well-known NLP corpus. [2]
- What is Named Entity Recognition (NER)? Give two examples of entity types. [2]
- Why are stop words removed during text preprocessing? [2]
- State the full form of BERT and explain in one sentence how it differs from traditional language models. [2]
Section C โ Long Answer Questions (5 marks each)
- Draw and explain the 7-step NLP pipeline. Apply each step to the sentence: "She quickly ran to the beautiful garden." [5]
- Explain TF-IDF with a numerical example using the following 2 documents:
Doc1: "cat sat mat cat" Doc2: "dog sat mat" Calculate TF-IDF for "cat" in Doc1. [5]
Section D โ Case Study / HOTS (5 marks)
- Case Study: Riya's school is building an AI system that reads student feedback forms and automatically identifies whether feedback about each subject is positive, negative, or neutral. The system must also identify the subjects mentioned by name.
(a) Which NLP task is being performed to identify positive/negative/neutral? (1)
(b) Which NLP step identifies the subject names (Maths, Science etc.)? (1)
(c) Why should common words like "the", "is", "a" be removed before analysis? (1)
(d) Explain how TF-IDF could help identify which subject is most discussed. (2) [5]
Topic 7: Model Evaluation & Improvement
Building an AI model is only half the job. Knowing how well it works โ and why it fails โ is equally important. This topic teaches you to measure, diagnose, and improve ML models.
A model that is 99% accurate sounds impressive โ until you learn it's diagnosing cancer. If it misses 1% of positive cases, those are real patients who go untreated. Choosing the right metric is as important as building the model itself.
1. The Confusion Matrix โ Visualising Predictions
A confusion matrix is a table that shows how many predictions were correct vs incorrect for each class. It is the foundation of classification evaluation.
Fig 7.1 โ Confusion matrix for a COVID test model. The 10 false negatives (missed COVID cases) are more serious than the 15 false positives (healthy patients incorrectly flagged).
2. Evaluation Metrics โ Precision, Recall, F1
Accuracy = (TP + TN) / Total = (80+95)/200 = 87.5%
Precision = TP / (TP + FP) = 80/(80+15) = 84.2% โ "Of all COVID+ predictions, how many were right?"
Recall (Sensitivity) = TP / (TP + FN) = 80/(80+10) = 88.9% โ "Of all real COVID+ cases, how many did we catch?"
F1-Score = 2 ร (Precision ร Recall) / (Precision + Recall) = 86.5%
Fig 7.2 โ The choice between precision and recall depends on which type of error is more costly in the real-world context.
3. The PrecisionโRecall Trade-off
Fig 7.3 โ The precision-recall trade-off: increasing one typically decreases the other. F1-score finds the optimal balance point.
4. Overfitting vs Underfitting
Fig 7.4 โ Underfitting models are too simple; overfitting models memorise training data but fail on new data. The goal is the right fit in the middle.
5. K-Fold Cross Validation
Instead of using a single train-test split, K-Fold uses the data K times, rotating which portion is used for testing each time. This gives a more reliable performance estimate.
Fig 7.5 โ In 5-Fold Cross Validation, each of the 5 equal portions takes a turn as the test set. The final accuracy is the average of all 5 test scores โ more reliable than a single split.
6. How to Improve a Model
Fig 7.6 โ Six strategies to improve model performance, categorised by whether they address data quality, model architecture, or training procedure.
7. Learning Curves โ Diagnosing Model Problems
Fig 7.7 โ Learning curves plot accuracy vs training size. Underfitting: both curves low. Good fit: curves converge. Overfitting: large gap between train and test accuracy.
Accuracy = (TP+TN)/Total โ Overall correctness; misleading with imbalanced data
Precision = TP/(TP+FP) โ Quality of positive predictions; use when FP is costly
Recall = TP/(TP+FN) โ Coverage of real positives; use when FN is dangerous
F1-Score = Harmonic mean of Precision & Recall โ Best for imbalanced datasets
Confusion Matrix โ Visual table of all prediction outcomes
Cross-Validation โ Technique to get robust performance estimate
Practice Paper โ Topic 7: Model Evaluation & Improvement
CBSE Class 9โ10 | Code 417 | Full Marks: 40 | Time: 1.5 hrs
Section A โ Multiple Choice Questions (1 mark each)
- A model correctly identifies 80 sick patients out of 90 actual sick patients. What is the Recall?
(a) 80% (b) 88.9% (c) 84.2% (d) 90% - Which cell in a confusion matrix represents a healthy person incorrectly labelled as sick?
(a) True Positive (b) True Negative (c) False Positive (d) False Negative - Overfitting occurs when:
(a) Model is too simple (b) Model memorises training data (c) Training accuracy is low (d) Test data is too large - K-Fold Cross Validation uses the data:
(a) Once (b) K times (c) Kยฒ times (d) Only for testing - The F1-Score is the harmonic mean of:
(a) Accuracy and Recall (b) Precision and Accuracy (c) Precision and Recall (d) Recall and Specificity
Section B โ Short Answer Questions (2 marks each)
- What is underfitting? Give one way to fix it. [2]
- Define Precision and state one real-world scenario where it is the most important metric. [2]
- Why is a confusion matrix more informative than accuracy alone? [2]
- What is regularisation? Name the two common types. [2]
- Explain with a diagram what a learning curve shows. [2]
Section C โ Long Answer Questions (5 marks each)
- A spam detection model was tested on 500 emails:
TP = 140, FP = 20, FN = 10, TN = 330
Calculate: (a) Accuracy (b) Precision (c) Recall (d) F1-Score
Which metric is most important for spam detection? Justify. [5] - Explain 5-fold cross validation with a neat diagram. Why is it preferred over a single train-test split? [5]
Section D โ Case Study / HOTS (5 marks)
- Case Study: A hospital built an AI model to detect diabetic retinopathy (eye disease) from retina scans. The model achieves 97% accuracy โ but the hospital finds it is almost never predicting the disease.
(a) Why can a model with 97% accuracy still be bad? (1)
(b) What problem does this indicate โ overfitting or class imbalance? (1)
(c) Which metric should the hospital prioritise: Precision or Recall? Why? (2)
(d) Suggest one technique to improve the model's performance on rare positive cases. (1) [5]
Topic 8: Computer Vision (CV)
Computer Vision teaches machines to see and interpret images and videos โ enabling self-driving cars, face recognition, medical imaging, and much more.
A digital image is simply a 2D grid of numbers. Each number (0โ255) represents the brightness of one pixel. A 640ร480 colour image = 640 ร 480 ร 3 = 921,600 numbers. Computer Vision is the art of finding patterns in these numbers.
1. Images as Numbers โ Pixels and Matrices
Fig 8.1 โ Every digital image is a matrix of numbers. Dark pixels have low values (near 0); bright pixels have high values (near 255). Colour images use three separate matrices.
2. Convolution โ How CNNs Detect Features
A convolutional filter (kernel) slides across an image, multiplying its values with the pixels it covers and summing the result. Different filters detect different features โ edges, curves, textures.
Fig 8.2 โ A Sobel edge filter multiplies pixel values and sums them. Where dark and bright pixels are adjacent (an edge), the output is large; in flat regions, it is near zero.
3. CNN Architecture โ Layer by Layer
Fig 8.3 โ A full CNN pipeline: convolution detects features, pooling reduces dimensions, dense layers classify. Backpropagation trains the filters automatically.
4. Computer Vision Applications
Fig 8.4 โ Computer Vision has transformed industries from healthcare to agriculture. The same CNN architecture underlies all these applications.
5. RGB Channels โ How Colour Images Work
Fig 8.5: RGB = 3 stacked matrices. A colour image is a 3D tensor (H ร W ร 3) โ this is the input to every CNN
6. Object Classification vs Detection vs Segmentation
Fig 8.6: Three levels of CV โ Classification answers "what", Detection answers "what and where", Segmentation answers "exactly which pixels"
7. Computer Vision in Healthcare & Agriculture
Fig 8.7: Computer Vision in Healthcare and Agriculture โ two sectors where CV is saving lives and transforming food security
Pixel โ Smallest unit of a digital image | RGB โ 3 channels (Red, Green, Blue) forming a 3D tensor (H ร W ร 3) | Convolution โ Feature extraction operation using a sliding filter | Filter/Kernel โ Small matrix that detects specific patterns | Feature Map โ Output of a convolution layer showing where features were detected | MaxPooling โ Downsampling by keeping the maximum value in each region
Classification โ "What is this?" (one label per image) | Detection โ "What and WHERE?" (bounding boxes, YOLO) | Segmentation โ "Which exact pixels?" (pixel-level masks, U-Net)
Practice Paper โ Topic 8: Computer Vision
CBSE Class 9โ10 | Code 417 | Full Marks: 40 | Time: 1.5 hrs
Section A โ MCQ (1 mark each)
- A 100ร100 grayscale image contains how many pixels?
(a) 100 (b) 1,000 (c) 10,000 (d) 30,000 - In a CNN, which layer reduces the spatial dimensions of a feature map?
(a) Convolutional (b) MaxPooling (c) Flatten (d) Softmax - What does a Sobel filter detect?
(a) Colour (b) Brightness (c) Edges (d) Faces - The final layer of an image classifier typically uses which activation function?
(a) ReLU (b) Sigmoid (c) Softmax (d) Tanh - RGB stands for:
(a) Red Green Blue (b) Random Gradient Boost (c) Recurrent Gradient Backprop (d) Radiant Graded Brightness
Section B โ Short Answer (2 marks each)
- What is a pixel? How is a grayscale image different from a colour image in terms of data representation? [2]
- Explain what happens in the MaxPooling layer of a CNN. [2]
- What is the difference between image classification and object detection? [2]
- Why do CNNs use multiple convolutional layers? What does each layer detect? [2]
- Name two real-world applications of computer vision in healthcare. [2]
Section C โ Long Answer (5 marks each)
- Draw and explain the complete CNN architecture for classifying images into cat/dog/bird. Include at least 5 distinct layers. [5]
- A 4ร4 image patch has values: [[1,2,3,4],[5,6,7,8],[9,10,11,12],[13,14,15,16]]. Apply a 2ร2 MaxPooling with stride 2. Show the resulting output matrix. [5]
Section D โ Case Study (5 marks)
- A hospital uses AI to detect diabetic retinopathy from eye scan images. The system analyses 256ร256 pixel images.
(a) How many pixel values does one grayscale 256ร256 image contain? (1)
(b) What type of neural network is best for this image analysis task? (1)
(c) What features might the early convolutional layers detect in the eye scan? (2)
(d) Why is it important the training dataset includes images from diverse patients? (1) [5]
Topic 9: Deep Learning
Deep Learning is a subfield of ML that uses multi-layered neural networks โ "deep" refers to the many layers these networks have โ enabling breakthrough results in vision, language, and beyond.
Fig 9.1 โ Deep Learning is a specialised subset of ML, which is itself a subset of AI. All deep learning is ML, but not all ML is deep learning.
1. Deep Learning Architecture Timeline
Fig 9.2 โ The progression of deep learning architectures from simple CNNs to modern billion-parameter language models.
2. Recurrent Neural Networks (RNNs) โ For Sequential Data
Fig 9.3 โ RNNs pass a "hidden state" (memory) between time steps, allowing the network to consider context from earlier words when making predictions.
3. ResNet โ Skip Connections Solve the Vanishing Gradient
Fig 9.4 โ ResNet's skip connections add the original input x directly to the layer output F(x), creating a "shortcut" for gradients to flow backward, enabling training of 100+ layer networks.
4. Deep Learning vs Traditional ML
Traditional ML
Feature engineering required by humans โ expert must choose which features matter
Works well with small-medium datasets (hundreds to thousands of examples)
Interpretable โ decision tree shows exactly why a decision was made
Algorithms: Decision Trees, SVM, k-NN, Naรฏve Bayes, Linear Regression
Fast to train on standard hardware (CPU)
Deep Learning
Automatic feature extraction โ network learns what to look for from raw data
Needs large datasets (millions of examples) to outperform traditional ML
Black box โ hard to explain why the network made a decision (XAI research)
Architectures: CNN, RNN, LSTM, Transformer, GAN, Diffusion
Requires GPU โ training ResNet-50 takes hours even on powerful GPU
GPT-3 has 175 billion parameters | GPT-4 estimated ~1.8 trillion | ResNet-50 has 50 layers | AlexNet (2012) won ImageNet with 15.3% error โ humans were at ~5% | Today's models achieve <2% error | ImageNet has 14 million images across 1,000 categories
Practice Paper โ Topic 9: Deep Learning
CBSE Class 9โ10 | Code 417 | Full Marks: 40 | Time: 1.5 hrs
Section A โ MCQ (1 mark each)
- What does "deep" in Deep Learning refer to?
(a) The depth of data (b) The number of hidden layers (c) The difficulty of tasks (d) The depth of the hardware - RNNs are most suitable for:
(a) Image classification (b) Sequential data like text and speech (c) Tabular data (d) Object detection - What is the main innovation in ResNet?
(a) Dropout (b) Skip connections (c) Attention mechanism (d) Batch normalisation - LSTM stands for:
(a) Large Scale Training Model (b) Layered Sequential Training Machine (c) Long Short-Term Memory (d) Linear Static Temporal Model - Which architecture is the foundation of ChatGPT?
(a) CNN (b) RNN (c) Transformer (d) LSTM
Section B โ Short Answer (2 marks each)
- What is the vanishing gradient problem? Which architecture solves it and how? [2]
- What is the difference between an RNN and an LSTM? Why is LSTM preferred for long sequences? [2]
- List three advantages of deep learning over traditional machine learning. [2]
- What is transfer learning? Give one example of its use. [2]
- Why does deep learning require large datasets and GPUs? [2]
Section C โ Long Answer (5 marks each)
- Draw and explain the general architecture of a deep neural network. Include input layer, hidden layers, and output layer. Explain the role of activation functions at each stage. [5]
- Compare Deep Learning and Traditional ML under 5 different criteria using a table. [5]
Section D โ Case Study (5 marks)
- A tech company wants to build a voice assistant that understands spoken Hindi and responds in English. The system must understand full sentences and maintain context across a conversation.
(a) Which type of neural network handles sequential audio input? (1)
(b) Why is LSTM preferred over a basic RNN for long conversations? (1)
(c) Which architecture would handle the translation from Hindi to English? (1)
(d) What is "fine-tuning" and how could a pre-trained model help reduce training time? (2) [5]
Topic 10: Introduction to Data Science
Data Science is the discipline of extracting insights from data using statistics, programming, and domain expertise. It is the fuel that powers all of AI and ML.
Fig 10.1 โ Data Science sits at the intersection of three skills: mathematical thinking, programming ability, and knowledge of the application domain (healthcare, finance, education, etc.).
1. The Data Science Workflow
Fig 10.2 โ The CRISP-DM (Cross-Industry Standard Process for Data Mining) framework is the industry-standard data science workflow. It is iterative โ new insights often require revisiting earlier steps.
2. Python Data Science Tools
Fig 10.3 โ The Python data science stack: Pandas for data manipulation, NumPy for maths, Matplotlib for charts, Scikit-learn for classical ML, and TensorFlow/PyTorch for deep learning.
3. Choosing the Right Visualisation
Fig 10.4 โ Choosing the right chart type is crucial. Bar charts compare categories, line charts show trends, scatter plots show correlations, and pie charts show proportions.
4. Data Science Career Roles
Collects and analyses data, creates dashboards and reports. Tools: Excel, SQL, Power BI, Tableau.
Builds predictive models, applies ML. Tools: Python, R, Scikit-learn, Jupyter.
Deploys models to production, builds ML pipelines. Tools: TensorFlow, Docker, Kubernetes, MLflow.
Builds data pipelines, manages warehouses. Tools: Spark, Hadoop, AWS, Airflow.
Develops new algorithms and architectures. Background: PhD in CS/Math.
Business Intelligence โ translates data insights into business decisions. Tools: SQL, Power BI, Tableau.
5. Data Science Real-World Impact
Fig 10.5 โ Data Science is one of the fastest-growing and highest-paid fields globally, with applications across every industry.
Data Science = Statistics + Programming + Domain Knowledge | Core Python tools: Pandas, NumPy, Matplotlib, Scikit-learn | CRISP-DM has 6 phases: Business โ Data โ Clean โ Explore โ Model โ Deploy | EDA (Exploratory Data Analysis) always comes before modelling | Good data scientist = 80% cleaning data, 20% modelling | "Data is the new oil" โ but raw data needs refining to be useful
Practice Paper โ Topic 10: Introduction to Data Science
CBSE Class 9โ10 | Code 417 | Full Marks: 40 | Time: 1.5 hrs
Section A โ MCQ (1 mark each)
- Which Python library is primarily used for data manipulation and analysis?
(a) NumPy (b) Matplotlib (c) Pandas (d) TensorFlow - What does EDA stand for?
(a) Estimated Data Analysis (b) Exploratory Data Analysis (c) Extended Data Attributes (d) External Data Analytics - The CRISP-DM framework has how many phases?
(a) 3 (b) 4 (c) 6 (d) 8 - Which chart type is best for showing trends over time?
(a) Bar chart (b) Pie chart (c) Line chart (d) Scatter plot - Data Science sits at the intersection of Statistics, Programming, and:
(a) Ethics (b) Hardware Knowledge (c) Domain Expertise (d) Networking
Section B โ Short Answer (2 marks each)
- What is EDA? Why is it performed before building an ML model? [2]
- Differentiate between a Data Analyst and a Data Scientist. [2]
- What is the purpose of data cleaning? Name two common data quality issues. [2]
- What does the scatter plot reveal about the relationship between two variables? [2]
- Name two Python libraries used for machine learning and state one function of each. [2]
Section C โ Long Answer (5 marks each)
- Draw and explain all 6 phases of the CRISP-DM data science workflow. Apply the framework to a scenario: "A school wants to predict which students are likely to fail their final exam." [5]
- A dataset of student exam scores contains the following issues: (a) 3 missing values in the "Marks" column, (b) One student listed as age "250", (c) Two duplicate rows. Describe how you would clean this data. Which pandas functions would you use? [5]
Section D โ Case Study (5 marks)
- Zomato (a food delivery app) collects data from: millions of restaurant orders, customer ratings, delivery times, and location data. They want to use Data Science to improve delivery time predictions.
(a) What type of ML problem is "predicting delivery time"? (1)
(b) List two features (input variables) that would help predict delivery time. (1)
(c) What chart would best show delivery time distribution across different time slots? (1)
(d) After deploying the model, delivery time predictions start becoming inaccurate after 3 months. What might cause this? What should Zomato do? (2) [5]