UNIT 2: ADVANCED CONCEPTS OF MODELING IN AI
Class 10 CBSE | Artificial Intelligence (Subject Code 417)
Target: Periodic Test (PT) & Board Exams Comprehensive Study & Revision Notes
1. Basics of Data & AI Modeling
What is AI Modeling?
AI Modeling is the process of building mathematical algorithms and computational models that allow machines to
analyze raw data, detect underlying patterns, learn autonomously, and make predictions or decisions without explicit
step-by-step programming.
Essential Data Terminology
• Data: Unprocessed raw facts, numbers, text, images, or measurements collected for training and evaluation.
• Features (Inputs / Attributes): Individual measurable properties or attributes of the data used as inputs for prediction. In
tables, features correspond to columns.
Example: Color, weight, size, and texture of a fruit.
• Labels (Outputs / Ground Truth): The final target variable, answer, or category that the AI model is expected to predict.
Example: Identifying whether the fruit is an Apple or an Orange.
• Labeled Data: Dataset containing both input features ($X$) and correct target labels ($Y$). Used in Supervised Learning.
• Unlabeled Data: Dataset containing input features ($X$) only, with no target labels provided. Used in Unsupervised
Learning.
Training Set vs. Testing Set Split
To ensure an AI model learns effectively and doesn't just memorize answers, data is partitioned into two sets:
Dataset Split Ratio Primary Purpose Real-World Analogy
Training Data 70% – 80% Fed directly into the algorithm during Textbook chapters and worked-out
training to discover patterns and build practice problems with solution
internal logic/rules. keys.
Testing Data 20% – 30% Kept completely hidden during training; The final surprise exam paper
used to test accuracy on new, unseen real- administered to evaluate true
world scenarios. mastery.
2. The AI Hierarchy: AI vs. ML vs. DL
Artificial Intelligence, Machine Learning, and Deep Learning are nested subsets of computer science:
Artificial Intelligence (AI) — Broadest Field
Machines executing tasks that mimic human intelligence (e.g., natural language, decision making).
Machine Learning (ML) — Subset of AI
Algorithms that learn and improve performance directly from data without explicit hard-coding.
Deep Learning (DL) — Subset of ML
Multi-layered Artificial Neural Networks (ANN) handling complex, unstructured data (e.g., images, audio).
CBSE Class 10 AI (417) | Quick Revision Notes Page 1 of 4
Relatable Analogy for the AI Hierarchy
• AI = Vehicles: Any machine designed to transport people (broad overall category).
• ML = Cars: A specific type of motorized vehicle that operates automatically on roads (subset).
• DL = Electric Autonomous Cars: A highly advanced, specialized subset of cars using cutting-edge battery and
sensor networks.
3. Approaches to AI Modeling
There are two fundamental approaches to creating AI models based on how rules and intelligence are established:
1. Rule-Based Approach Static 2. Learning-Based Approach Dynamic
• How it works: Developers manually code exact logic, • How it works: Raw data and desired outcomes are
IF-THEN rules, and constraints into the computer provided. The model analyzes patterns to formulate its
alongside data to get results. own internal rules.
• Nature: Static . Does not learn or alter its rules • Nature: Dynamic . Continuously learns, adapts, and
independently. improves with new experience.
• Limitation: Fails when faced with unexpected inputs • Advantage: Handles vast, complex, changing real-world
outside programmed rules. scenarios.
• Example: A basic pocket calculator or automated tax • Example: Email spam filter, YouTube recommendation
calculator. engine.
Rule-Based vs. Learning-Based (High-Yield Comparison Table)
Feature / Aspect Rule-Based Approach Learning-Based Approach
Source of Rules Explicitly written by human programmers. Learned automatically by the machine from
data.
Adaptability Static: Cannot adjust to unforeseen changes. Dynamic: Adapts automatically as new data
arrives.
Data Required Minimal to no dataset needed. Requires large datasets for effective training.
Best Application Predictable, fully defined logical problems. Complex, real-world problems with hidden
patterns.
4. Types of Machine Learning Paradigms
Learning-based models are categorized into three primary paradigms based on the type of feedback and data:
A. Supervised Learning Labeled Data
The model is trained using labeled data (inputs paired with known target outputs).
Analogy:
Studying with a teacher nearby who checks every answer against a solution key and corrects your mistakes immediately.
Two Core Types of Supervised Learning:
CBSE Class 10 AI (417) | Quick Revision Notes Page 2 of 4
Sub-Type Definition & Output Type Real-World Examples
1. Classification Predicts discrete, categorical values or distinct class • Spam vs. Not Spam
labels (e.g., Yes/No, A/B/C). • Pass vs. Fail
• Disease Diagnosis (Positive/
Negative)
2. Regression Predicts continuous, numerical values along a • Predicting House Prices
numerical scale. • Weather Temperature Forecast
• Predicting Exam Score Percentage
B. Unsupervised Learning Unlabeled Data
The model processes unlabeled data with no guidance or answer key, identifying natural patterns, groupings, and structures
on its own.
Analogy:
Sorting a huge tray of mixed coins or buttons into distinct piles purely by color, shape, or weight without anyone telling
you the categories.
Three Key Techniques in Unsupervised Learning:
• 1. Clustering: Grouping similar data points together based on natural similarities without prior labels.
Example: Segmenting e-commerce shoppers into groups based on buying habits.
• 2. Association: Discovering interesting relationships or correlations between items in datasets.
Example: Market Basket Analysis — finding that customers who buy bread also buy butter 80% of the time.
• 3. Dimensionality Reduction: Reducing the number of input variables/columns in heavy datasets while retaining
essential features.
C. Reinforcement Learning Reward & Penalty
An autonomous agent learns to make decisions by interacting with an environment through trial and error, receiving
feedback in the form of Rewards (+) for good decisions and Penalties (-) for bad decisions.
Analogy:
Training a puppy. Giving a treat (Reward) when it obeys a command, and saying "No" (Penalty) when it misbehaves.
Examples: Self-driving car navigation, AI game bots (e.g., Chess engines, AlphaGo), robot movement training.
5. Model Fitting Errors: Underfitting vs. Overfitting
When an AI model is trained, it can fail to accurately learn patterns due to improper design or bad data split.
Scenario Training Performance Testing Performance Primary Cause
Underfitting Poor / Low Poor / Low High Bias (Model is too
Too Simple simple to capture
underlying trend)
Overfitting Excellent / Near 100% Poor / Low High Variance (Model
Too Complex memorized training data
+ noise word-for-word)
Optimal Fit Ideal High / Accurate High / Accurate Balanced Bias &
Variance (Model learned
true general trends)
CBSE Class 10 AI (417) | Quick Revision Notes Page 3 of 4
Exam Analogy for Overfitting vs. Underfitting
• Underfitting (High Bias): Skimming only 2 pages of a textbook before the exam. You fail both the practice questions
and the actual test.
• Overfitting (High Variance): Rote-memorizing exact answer key sentences word-for-word without understanding.
You get 100% on practice questions, but fail when the teacher changes the numbers on the real exam.
6. Quick Exam Cheat Sheet & Distinctions
Top 4 Concepts Frequently Tested in Class 10 Board / PT Exams:
1. Classification vs. Clustering Difference:
Classification uses predefined categories on labeled data (Supervised).
Clustering creates spontaneous groups on unlabeled data (Unsupervised).
2. Features vs. Labels Difference:
Features are input attributes ($X$), Labels are target outputs ($Y$).
3. Convolutional Neural Network (CNN):
A specialized Deep Learning algorithm designed primarily for Computer Vision tasks (image classification, object
detection, face recognition).
4. Dynamic vs. Static:
Rule-based models are Static; Learning-based models are Dynamic.
CBSE Class 10 AI (417) | Quick Revision Notes Page 4 of 4