AI Curriculum Notes for Class 10 Facilitator Handbook (CBSE + Intel)

Unit 1: Revisiting AI Project Cycle & Ethical Frameworks for AI

• Context: AI Curriculum for Class 10 Facilitator Handbook curated with support from Intel; CBSE collaboration since 2019.
• Purpose: Equip educators and students with AI concepts, project workflow, ethical reasoning, and real‑world relevance for an AI‑driven world.
• Acknowledgements: Patrons and content curation team listed (CBSE, Intel, government officials, educators).

1.1 AI Project Cycle

• The AI Project Cycle is a cyclical process to complete an AI project with 6 stages:
  • 1) Problem Scoping: define the goal and scope of the AI project.
  • 2) Data Acquisition: collect data from reliable sources; data are the base for the project.
  • 3) Data Exploration/Visualization: interpret patterns using graphs, databases, maps, etc. to understand relations.
  • 4) Modelling: decide on a suitable model(s) after researching online and evaluating alternatives.
  • 5) Testing/Evaluation: test models on newly fetched data to assess performance.
  • 6) Deployment: integrate the model into real-world environments to deliver value.
• Practical example: Greeting card creation to illustrate planning; parallels to AI project planning (idea → materials → execution → iteration → delivery).
• Key takeaway: Planning and iterative refinement are essential; data quality and model selection determine success.

1.2 Introduction to AI Domains

• Three broad AI domains introduced:
  • Statistical Data: AI domain focused on data systems, data collection, storage, and extraction of insights; examples include price comparison websites (PriceGrabber, PriceRunner, Junglee, Shopzilla, DealTime).
  • Computer Vision (CV): Enables machines to acquire, analyze, and act on visual data (images, videos, infrared sensors); aims to translate visual data into computer‑readable descriptions; examples include agricultural monitoring, surveillance, etc.
  • Natural Language Processing (NLP): Interaction between computers and humans using natural language; aims to read, decipher, understand, and extract meaning from human language; examples include email filters and machine translation.
• CV applications: Facial recognition, face filters, image search, retail analytics, self‑driving cars, medical imaging, Google Translate app.
• NLP applications: Email filters, machine translation, sentiment analysis, text classification, keyword extraction.
• Importance of domain understanding: Contextualizing AI concepts to real‑world domains enhances learning and social impact.

1.3 Ethical Frameworks for AI

• What is a framework? A framework provides a step‑by‑step, organized approach to solving problems; common language for collaboration; helps ensure consistency.
• What is an ethical framework? A systematic approach to navigate moral dilemmas by considering principles and stakeholder impacts; helps ensure choices do not cause unintended harm.
• Why do we need Ethical Frameworks for AI?
  • To mitigate bias and ensure morally acceptable recommendations from AI systems.
  • To anticipate and avoid unintended outcomes before deployment.
• Types of ethical frameworks:
  • Sector-based frameworks: Tailored to specific sectors (e.g., Bioethics in healthcare). Other sectors include finance, education, governance, etc.
  • Value-based frameworks: Grounded in fundamental ethical principles (e.g., rights-based, utility-based, virtue-based).
• Value-based frameworks (3 families):
  • Rights-based: Emphasize protection of human rights and dignity; autonomy and freedoms; avoid discrimination.
  • Utility-based (Utilitarian): Maximize overall utility/benefit; weigh benefits vs. risks to maximize societal good.
  • Virtue-based: Focus on character and intentions of individuals/organizations; alignment with virtuous traits (honesty, compassion, integrity).
  • Note: Value-based frameworks assess moral worth of actions and guide ethical behavior across contexts.
• Bioethics (a prominent sector‑based framework in healthcare):
  • Core principles: Respect for Autonomy; Do No Harm (non‑maleficence); Beneficence (maximize benefit); Justice (fair distribution of benefits/burdens).
  • Non‑maleficence: Avoid causing harm; minimize harm where possible.
  • Maleficence: Avoid intentional harm.
  • Beneficence: Promote well‑being and positive outcomes for stakeholders.
• Case study: Applying bioethics to AI in healthcare
  • Example problem: An AI algorithm trained on healthcare expense data (US‑centric) may misrepresent illness severity across regions, leading to biased resource allocation.
  • Application of bioethics:
  • Respect for Autonomy: Provide transparency, reproducibility, and accessible model information to patients.
  • Do No Harm: Distribute benefits and harms justly; minimize regional biases in data and outcomes.
  • Maximum Benefit: Use unbiased, clinically relevant data; ensure medical practice standards guide AI deployment.
  • Justice: Consider social determinants of health; address structural biases affecting outcomes across populations.
• Activities to reveal ethical biases:
  • My Goodness game: interactive exercise exploring donation decisions to surface personal biases and decision‑making factors.
  • Factors that can influence decisions: identity/location of recipients, bias toward relatives, information disclosure, cultural and religious values, intuition.
• Ethical frameworks in practice:
  • Students learn to outline the AI project cycle, identify AI domains, classify ethical frameworks, explore bioethics principles, and practice applying frameworks to AI solutions.

Unit 2: Advanced Concepts of Modeling in AI

• Goal: Introduce modeling concepts, distinguishing rule‑based vs learning‑based AI, and the major learning paradigms.

2.1 Revisiting AI, ML, DL

• Definitions and relationships:
  • AI (Artificial Intelligence): Umbrella term for techniques enabling computers to mimic human intelligence.
  • ML (Machine Learning): Subset of AI; systems improve with experience/data; learns from data to improve performance.
  • DL (Deep Learning): Subset of ML; uses large volumes of data; multiple layers of neural networks to learn representations.
• Relationship: DL is a subset of ML, which is a subset of AI (AI ⊇ ML ⊇ DL).
• Conceptual visualization: AI includes a broad set of techniques; ML is about learning from data; DL is a specific family of learning methods using deep neural networks.

2.2 Modelling: Rule‑Based vs Learning‑Based; ML Model Categories

• Two broad modelling approaches:
  • Rule‑Based: Developer defines rules and data interactions; model behavior is fixed after training; limited ability to adapt to new data without reprogramming.
  • Learning‑Based: Machine learns rules from data; adapts to changing data patterns; examples include spam filtering where rules evolve with data.
• Categories of learning‑based models (ML/DL):
  • Supervised Learning: Trained on labeled data; tasks include classification and regression.
  • Unsupervised Learning: Learns structure from unlabeled data; tasks include clustering and association.
  • Reinforcement Learning: Learns via rewards/punishments to maximize cumulative reward; interacts with environment.
• Neural networks (NN): Core mechanism in many learning systems; comprised of layers and nodes; automatic feature extraction; capable of handling large datasets (e.g., images).
• Subtopics to know:
  • Supervised Learning: Classification vs Regression; labeled data; examples like coin weight classification (weights as features; currency as labels).
  • Unsupervised Learning: Clustering and Association; no labels; discover structure/patterns; examples include customer segmentation and market basket analysis.
  • Reinforcement Learning: Trial‑and‑error with reward signal; used for tasks like parking a car or humanoid walking.
• Practical note: Real‑world AI often combines approaches to handle data gaps and changing environments.

2.3 Neural Networks (NNs)

• What is a neural network? A system inspired by the brain, consisting of layers of nodes (neurons) arranged as input, hidden, and output layers.
• How it works:
  • Data flows from input to output through hidden layers where each node applies a function to weighted inputs plus a bias.
  • Activation functions introduce nonlinearity to enable complex mappings.
  • Training adjusts weights (via error signal) to minimize loss; learning continues through backpropagation across multiple epochs.
• Depth and capacity: More hidden layers (deep networks) can model more complex functions; the number of nodes/layers depends on problem complexity.
• Real‑world NN applications: Facial recognition, chatbots, price prediction, etc.
• Quick perceptron note (illustrative): A simple decision unit with weights, bias, and threshold; output often uses a step function to map to a binary decision.

Unit 3: Evaluating Models

• Purpose: Understand model performance using standard metrics; establish a feedback loop to improve accuracy and reliability.

3.1 Importance of Model Evaluation

• Evaluation is the process of using metrics to quantify model performance.
• Why evaluation matters: prevents deployment of flawed models; provides a basis for comparison and improvement.

3.2 Train–Test Split

• Concept: Divide data into training and testing subsets to estimate performance on unseen data.
• Rationale: Training data are used to learn; testing data evaluate predictive ability on new inputs; avoid overfitting.
• General rule: Use train‑test split when there is a sufficiently large dataset.

3.3 Accuracy and Error

• Accuracy: Proportion of correct predictions across all predictions.
• Error: Proportion (or rate) of incorrect predictions; used to quantify miss rate.
• Relationship: Accuracy = 1 − Error.
• Real‑world nuance: In imbalanced datasets, accuracy alone can be misleading; other metrics may be preferred.

3.4 Evaluation Metrics for Classification

• Classification: Assigning items to discrete classes; metrics include:
  • Confusion matrix: 2x2 table for binary cases; entries include True Positive (TP), True Negative (TN), False Positive (FP), False Negative (FN).
  • Accuracy = (TP + TN) / (TP + TN + FP + FN)
  • Precision = TP / (TP + FP)
  • Recall (Sensitivity) = TP / (TP + FN)
  • F1 Score = 2 * (Precision * Recall) / (Precision + Recall)
• When to use which metric:
  • Precision emphasizes reducing FP (e.g., high‑risk medical tests, satellite launch decisions).
  • Recall emphasizes reducing FN (e.g., disease detection to avoid missing cases).
  • F1 Score balances precision and recall, especially in imbalanced datasets.

3.5 Ethical Concerns in Model Evaluation

• Bias: Evaluation data can reflect societal biases; must ensure fair assessment across groups.
• Transparency: Clear reporting of model behavior, limitations, and risk disclosures.
• Accountability: Define responsibility for outcomes and potential harms.

Unit 4: Statistical Data & No‑Code AI for Statistics

• No‑Code AI for Statistical Data: No‑code vs low‑code approaches; aim to enable non‑programmers to build AI models without writing code.
• No‑Code vs Low‑Code: No‑code tools emphasize drag‑and‑drop GUIs; low‑code offers some scripting; both aim to reduce traditional coding requirements.
• No‑Code tools highlighted in this unit include Orange Data Mining, Lobe, Teachable Machine, Google Cloud AutoML, Azure ML, etc.
• Orange Data Mining (ODMap): A no‑code data mining tool with a GUI for loading data, exploring, preprocessing, modeling, evaluating, and deploying models; supports both supervised and unsupervised workflows via widgets.
• Palmer Penguins case study (OD Map): Demonstrates data exploration, modeling, and evaluation using a real dataset to predict penguin species; illustrates mapping of the AI project cycle to a concrete data science problem.
• Statistical concepts covered:
  • Descriptive statistics: Mean, Median, Mode.
  • Distributions: Normal distribution and other distributions; understanding skew and central tendency.
  • Probability and variance; standard deviation.
  • Outliers and data quality considerations.
• Excel for statistical analysis: Introduction to using Excel’s Analysis ToolPak for simple linear regression (speed vs distance) with steps to load add‑ins, run regression, and interpret outputs including regression equation and R².

Unit 5: Computer Vision

• 5.1 Introduction to CV: CV enables machines to see and interpret visual data; CV is a branch of AI focusing on image/video understanding.
• 5.2 CV Applications: Face recognition, face filters, image search, CV in retail, self‑driving cars, medical imaging, Translate by image.
• 5.3 CV Tasks:
  • Classification: Assign an image to one of several categories.
  • Classification + Localisation: Locate and classify a single object in an image.
  • Object Detection: Identify and locate multiple instances of objects within an image.
  • Image Segmentation / Instance Segmentation: Pixel‑level labeling of objects.
  • Image basics: Pixels, resolution, grayscale vs RGB, image channels.
• 5.4 No‑Code AI tools for CV:
  • Lobe: Auto‑ML for image classification; drag‑and‑drop, no coding required.
  • Teachable Machine: Web tool by Google for image, audio, and pose classification; runs on TensorFlow.js.
  • Orange Data Mining: Includes CV‑oriented widgets and use case walkthroughs (e.g., coral bleaching, smart sorters).
• 5.5 Image Features: Features such as corners, edges, and blobs; corners are generally robust and good features to extract.
• 5.6 Convolution: Core operator in CV; convolution multiplies image with a kernel to produce a feature map; used for edge detection, blurring, sharpening, etc.
• 5.7 Convolutional Neural Network (CNN): A deep learning model that uses convolutional layers, ReLU activation, pooling layers, and fully connected layers to perform image understanding.
  • 5.7.1 Convolution Layer: Extracts high‑level features (edges, textures) and produces a feature map.
  • 5.7.2 ReLU: Introduces non‑linearity by setting negative values to zero; enhances feature detection.
  • 5.7.3 Pooling Layer: Reduces spatial dimensions; max pooling vs average pooling; increases invariance to small shifts.
  • 5.7.4 Fully Connected Layer: Maps features to class probabilities and outputs final prediction.
• 5.8 Python libraries: TensorFlow, Keras, OpenCV; used for implementing CV workflows in code.

Unit 6: Natural Language Processing (NLP)

• 6.1 Introduction to NLP: NLP focuses on enabling computers to understand and process human language; features of natural language include ambiguity, context dependence, and dynamic evolution.
• 6.2 Applications of NLP: Autogenerated captions, voice assistants, language translation, sentiment analysis, text classification, keyword extraction.
• 6.3 Stages of NLP:
  • Lexical Analysis: Tokenization and lexicon handling; breaking text into words/tokens.
  • Syntactic Analysis (Parsing): Grammar checking and relationships among words; syntactic structure.
  • Semantic Analysis: Meaning and validity of statements; semantic plausibility checks.
  • Discourse Integration: Coherence across sentences; narrative flow.
  • Pragmatic Analysis: Practical usage and real‑world intent; contextual relevance.
• 6.4 Chatbots: Scripted vs smart bots; chatbots simulate conversation; examples of chatbots and experiences with them.
• 6.5 Text Processing: Text normalization, Bag of Words, TFIDF; importance of preprocessing and feature extraction for ML models.
• 6.6 NLP: Use case walkthroughs; no‑code tools in NLP (Orange Data Mining) highlighted; emphasis on sentiment analysis and other NLP tasks.
• No‑Code NLP tools: Orange Data Mining, NLP workflows; focus on sentiment analysis and keyword extraction.

Practical Concepts and Formulas

• TFIDF in text processing (concept and formula):
  • Let N be the total number of documents; DF(W) is the document frequency (how many documents contain word W).
  • IDF(W) = \log_{10}\left( \frac{N}{DF(W)} \right).
  • For a document d, TF(W, d) is the term frequency of word W in document d (how many times W occurs in d).
  • TFIDF(W, d) = TF(W, d) * IDF(W).
  • TFIDF values are used to identify important words in a document relative to the corpus; high TFIDF indicates a word that is frequent in a document but not common across all documents.
• Perceptron (illustrative): Output ≈ step(\sumi wi x_i + b); a simple binary classifier; foundational concept for neural nets.
• Convolutional Neural Network (CNN) architecture (high‑level):
  • Convolution Layer → ReLU → Pooling Layer → Fully Connected Layer → Output.
  • Feature maps (activation maps) produced by convolution; ReLU introduces nonlinearity; pooling reduces dimensionality and increases translation invariance; fully connected layers perform final classification.

Connections to Real World and Foundational Principles

• Ethical considerations: Throughout Unit 1, ethical frameworks underpin design decisions, data collection, model evaluation, and deployment, ensuring fairness, transparency, and accountability.
• Cross‑unit connections:
  • Unit 1 sets the stage for applying ethical frameworks across all AI projects.
  • Unit 2 clarifies the modeling choices that drive Unit 3’s evaluation metrics.
  • Unit 4 introduces statistics and no‑code tools that support data exploration used in CV and NLP tasks (Unit 5 and Unit 6).
  • Unit 5 and Unit 6 provide domain‑specific methods (CV and NLP) that rely on data representations learned in earlier units (features, vectors, embeddings, etc.).

Test Yourself (Representative Questions and Concepts)

• Example question types from the transcript focus on problem scoping, domain categorization, evaluation metrics, CV/NLP concepts, and ethical frameworks. Sample categories include:
  • Problem statement purpose in AI Project Cycle.
  • Data domain categorization based on data fed to models.
  • Statistical data handling and uses of no‑code tools.
  • CV tasks (classification, localization, object detection, segmentation).
  • NLP stages (lexical, syntactic, semantic, discourse integration, pragmatic).
  • Ethical frameworks (rights, utility, virtue, bioethics) and their application to AI projects.

Case Studies and Scenarios

• Healthcare AI case study (bioethics): address bias and fairness; apply four bioethics principles to ensure equity in patient care and resource allocation.
• Palmer Penguins case (Orange ODMap): a practical example of mapping the AI project cycle to a real dataset and evaluating model performance.
• My Goodness bias activity: a practical activity to reflect on bias in charitable decisions and how personal values influence AI design.

Key Takeaways for Exam Preparation

• Know the 6 stages of the AI Project Cycle and their purposes.
• Be able to distinguish AI domains: Statistical Data, Computer Vision, NLP; know typical examples and applications.
• Understand rule‑based vs learning‑based modelling, and the three ML paradigms: supervised, unsupervised, reinforcement learning.
• Be able to explain basic NN concepts (layers, activations, backpropagation) and the core CNN components (convolution, ReLU, pooling, fully connected).
• Master evaluation metrics for classification (accuracy, precision, recall, F1, confusion matrix) and know when to use each in different data contexts (balanced vs imbalanced datasets).
• Understand no‑code vs low‑code AI tools, and their role in democratizing AI without heavy coding, with examples like Orange Data Mining, Lobe, Teachable Machine, and AutoML platforms.
• Grasp TFIDF for NLP, including the definitions of TF, IDF, and TFIDF; understand why high TFIDF words are informative for document classification and information retrieval.
• Appreciate ethical frameworks as guiding principles in AI design and deployment, especially piecewise application to sector‑specific problems (healthcare, finance, governance).

Notation and Formulas (LaTeX)

• TFIDF in document d for word W:
  ext{TFIDF}(W, d) = ext{TF}(W, d) imes ext{IDF}(W)

• Inverse Document Frequency:
  ext{IDF}(W) =
  \log_{10}iggl( \frac{N}{DF(W)} \biggr)
  where N is the total number of documents and DF(W) is the number of documents containing W.

• Confusion Matrix (binary): TP, TN, FP, FN; Accuracy:
  ext{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}

• Precision, Recall, F1:
  ext{Precision} = \frac{TP}{TP + FP}, \quad ext{Recall} = \frac{TP}{TP + FN}, \quad ext{F1} = 2 \times rac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}

• Conceptual equation for a simple perceptron (binary decision):
  y = \begin{cases}1, & \sumi wi x_i + b > 0 \ 0, & \text{otherwise} \end{cases}

• CNN workflow (high level):
  Input Image → Convolution Layer (feature maps) → ReLU → Pooling → Fully Connected Layer → Output (class probabilities)

Connections to Foundational Principles and Real‑World Relevance

• Ethical framing informs all stages of AI development, from data collection to deployment.
• No‑Code tools democratize access to AI, enabling non‑technical stakeholders to prototype and deploy solutions, but require awareness of limitations (flexibility, potential biases, security concerns).
• CV and NLP are critical for visual and language information processing in modern applications (autonomous vehicles, healthcare imaging, sentiment analysis, translation).
• Statistical thinking (descriptive stats, distributions, regression) underpins data exploration and model evaluation across domains.

Quick Reference Glossary

• AI Project Cycle: Problem Scoping, Data Acquisition, Data Exploration, Modelling, Testing, Deployment.

• Ethical Frameworks: Sector‑based (Bioethics) and Value‑based (Rights, Utility, Virtue).

• No‑Code / Low‑Code: Platforms enabling AI development with minimal or no coding.

• TFIDF: Term Frequency–Inverse Document Frequency; a weighting scheme for text representation.

• Convolutional Neural Network (CNN): Deep learning model for image data using convolutional layers.

• Lexical Analysis / Syntactic Analysis / Semantic Analysis / Discourse / Pragmatics: Stages of NLP processing.

• For each unit, be prepared to discuss definitions, examples, ethical considerations, and how to apply these concepts to real situations (classroom activities and case studies).