Unit 2: Advanced Concepts of Modeling in AI
This unit explains how AI models are built and how learning-based models can learn from data. It also connects Artificial Intelligence, Machine Learning, Deep Learning, and Neural Networks in a simple way.
AI, ML and DL: How They Differ and Relate
Artificial Intelligence, Machine Learning and Deep Learning are related, but they are not the same.
- Artificial Intelligence, or AI, is the broad term. It includes any technique that helps computers copy human intelligence.
- An AI machine works using algorithms and data given to it. It then gives the required output.
- Machine Learning, or ML, is a part of AI. It helps machines become better at a task through experience.
- In ML, the machine learns from new data, past mistakes and exceptions. It uses this learning in the next execution.
- Deep Learning, or DL, is a part of ML. It helps software train itself to do tasks using very large amounts of data.
- In DL, many machine learning algorithms can work together for a specific task.
- AI is the umbrella term. ML comes under AI. DL comes under ML.
Machine Learning uses past or historical data as input. The ML model studies this input and produces output by learning from it. For example, if labelled images of apples and strawberries are given, the model can learn to classify a new image as apple or strawberry.
Deep Learning uses large amounts of data. Input is given to an Artificial Neural Network, also called ANN. After processing, the Deep Learning block gives the output. For example, pixels of a bird image can be given to a DL model, and the model can predict that the image is of a bird.
Examples of Machine Learning given in this unit include:
- Object Classification: It finds and labels objects in an image or data point, and tells the category of the object.
- Anomaly Detection: It finds unexpected things in data, such as a sudden spike in heart rate.
Examples of Deep Learning given in this unit include:
- Object Identification: It identifies and labels objects inside an image.
- Digit Recognition: It trains computers to identify handwritten digits from 0 to 9 in images.
Everyday example: A student first follows instructions, then improves by practising many solved examples, and later becomes better at solving similar tasks independently.
Common Terminologies Used with Data
Data is the base for AI models. This unit uses the following terms with data.
-
Data: Data means information in any form. A table containing information about fruits is data. Each row may contain information about a different fruit. Each fruit is described through features.
-
Features: Features are the columns of a table. In a fruit dataset, features may include name, color and size. Some features are special and are called labels.
-
Labels: Data Labeling means adding meaning to data. The label depends on the problem being solved. If the task is to predict the fruit name using the color of the fruit, then color is the feature and fruit name is the label. Data can be of two types: Labeled and Unlabeled.
-
Labeled Data: Labeled data has a tag or label attached to it. Examples of labels include name, type and number.
-
Unlabeled Data: Unlabeled data is raw data. It has no tag attached to it.
-
Training data set: A training data set is a collection of examples given to a model so that it can study and learn. A teacher teaches a topic using many examples and illustrations. In a similar way, a set of labeled data is used to train an AI model.
-
Testing data set: A testing data set is used to test how accurate the model is. A teacher may take a class test to check students’ understanding. In a model, the test is done without labeled data first, and then the results are checked with the labels.
Everyday example: A table of fruits with columns such as color and size can help a learner identify which fruit is being described.
Types of AI Models: Rule Based Approach and Learning Based Approach
AI Modelling means making algorithms, also called models, that can be trained to give intelligent outputs. It means writing code to make a machine artificially intelligent.
AI models are generally classified into two types:
- Rule Based Approach
- Learning Based Approach
Rule Based Approach:
- In this approach, the developer defines the relationships or patterns in the data.
- The machine follows the rules or instructions given by the developer.
- The machine performs the task according to those rules.
- A rule-based chatbot can answer frequently asked questions or give basic customer support.
- Such a chatbot may use predefined questions, predefined answers and a decision tree type of method.
- If a user’s message matches a defined rule, the chatbot gives the related response.
- If the message does not match any rule, it may give a message that it cannot help with that.
A drawback or feature of the rule-based approach is that learning is static. After training, the machine does not consider changes made in the original training dataset. If the test data is different from the rules and data used during training, the machine may fail. It does not learn from that mistake. It also cannot improve itself using feedback once trained.
Learning Based Approach:
- In this approach, the computer learns from examples or feedback.
- It is similar to how humans learn from experience.
- The computer is not directly programmed for every task.
- It studies data and finds patterns or rules on its own.
- The model is trained on the data given to it.
- It can design a model that adapts to changes in data.
- It can handle exceptions by changing according to the data.
In a learning-based model, if random images of stray dogs are given, the machine can find patterns in the images. It may group the data using features such as colour, size or fur style. It may also find an unusual pattern that the user had not expected.
A learning-based spam email filter learns from labeled emails during training. It studies emails labeled as spam or legitimate. It looks at words, sender information and attachments. After training, it classifies new emails as spam or not spam and improves as it sees new examples.
Everyday example: A child who is first given rules follows only those rules, but a child who learns from many examples can adjust when a new situation appears.
Categories of Machine Learning: Supervised, Unsupervised and Reinforcement Learning
Learning-based approaches include machine learning and deep learning. Machine Learning is divided into three parts:
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
Supervised Learning:
- In supervised learning, the dataset given to the machine is labelled.
- The person training the machine knows the dataset and can label it.
- A label is information used as a tag for data.
- Students getting grades according to marks is an example of labels being used to categorize students.
- Supervised Learning means learning with a supervisor like a teacher.
- The machine is trained using labeled data.
A supervised learning example uses currency coins. The problem is to build a model that predicts the coin using its weight. Weight is the feature. Currency is the label. If the model learns the connection between coin weights and currencies, it can identify a coin from its weight.
Unsupervised Learning:
- In unsupervised learning, the model works on an unlabelled dataset.
- The data given to the machine is random.
- The person training the model may not have information about the data.
- The model is used to find relationships, patterns and trends in the data.
- It helps the user understand what the data is about.
- It also helps find major features identified by the machine.
- It is learning without guidance.
- The machine discovers patterns, similarities and differences on its own using the unlabeled dataset.
An unsupervised learning example is a supermarket customer database. If there are records of products bought over time, the model may find groups such as grocery shoppers and non-grocery shoppers without those labels being given beforehand.
Reinforcement Learning:
- Reinforcement Learning allows a computer to make a series of decisions to maximize a reward metric.
- It works without human intervention and without being explicitly programmed to complete the task.
- It is learning through repeated trial and error.
- The machine receives feedback and learns from it.
In the fruit example, the machine is shown an apple. If it predicts cherry, it receives negative feedback and learns that the image is not cherry. If it later predicts apple and receives positive feedback, it learns that the image is apple.
Reinforcement Learning is different because some situations have large and complex problem spaces. The system may need to respond to environments it has not seen before. The environment may change, so the system should be adaptive. It does not need a lot of pre-existing knowledge or data to give useful solutions.
Examples of Reinforcement Learning include:
- Parking a car
- Humanoid walking
Summary of ML models:
- Supervised learning models are used when we want to determine relationships through training.
- Unsupervised learning models are used when we want to discover new patterns from data.
- Reinforcement learning models are used when we want to implement machine learning through a reward mechanism.
Everyday example: A person learning by reward and correction changes their next action based on what worked and what did not.
Subcategories of Supervised Learning: Classification and Regression
There are two types of Supervised Learning models:
- Classification model
- Regression model
Classification Model:
- In classification, data is classified according to labels.
- In a grading system, students are classified using the grades they get according to their marks.
- This model works on a discrete dataset.
- Discrete data need not be continuous.
A classification model can be trained on historical weather data labeled as hot or cold. It learns patterns that separate hot and cold weather using factors such as location, season, high and low temperatures, and humidity. When weather data for tomorrow is given, it predicts the category as hot or cold.
Another classification example is classifying emails as spam or not spam. The model is shown many emails, including real emails and spam emails. It learns what makes an email look like spam. After training, it reads a new email and assigns it to the category spam or not spam. Email classifiers may also identify categories such as social, advertisement and notifications.
Regression Model:
- Regression models work on continuous data.
- A regression algorithm predicts a continuous value using input variables.
- Continuous values include Temperature, Price, Income and Age.
A regression model can predict temperature because temperature is a continuous variable. It can take any value within a range.
Another example is house price prediction. The model predicts the price of a house using features or independent variables such as:
- No of bedrooms
- Carpet Size
- Garage Area
The label or dependent variable is:
- Price
A used car price prediction model is also regression. It predicts an approximate selling price using parameters such as fuel type, years of service, number of previous owners, kilometers driven and transmission type, manual or automatic.
Everyday example: Sorting a day as hot or cold is classification, while estimating the exact temperature is regression.
Subcategories of Unsupervised Learning: Clustering and Association
Unsupervised learning models are further divided into two categories:
- Clustering model
- Association model
Clustering:
- Clustering works with input data that has no class labels.
- The model divides data into clusters using similarity of characteristics.
- In an example with birds and animals, the model can create one cluster for animals and another cluster for birds.
- The model forms groups even though class labels were not given.
Difference between Clustering and Classification:
- Classification uses predefined classes in which objects are assigned.
- Clustering finds similarities between objects and places them in the same cluster, while separating them from objects in other clusters.
In the music example, Jim likes music with slow tempo and soft intensity. He dislikes music with fast tempo and high intensity. Songs with slow tempo and soft intensity can be grouped into one cluster that he likes. Songs with fast tempo and high intensity can be grouped into another cluster. When a new song has slow tempo and soft intensity, the clustering model can use the pattern to predict that he may like it. Here, tempo and intensity are known features, but the clusters based on likes and dislikes are found from patterns.
Association:
- Association Rule is an unsupervised learning method.
- It is used to find interesting relationships between variables from a database.
- It can be useful for recommending items to customers.
In the purchase pattern example, if customers who buy bread are also likely to buy butter, then the model can find this meaningful association. Based on other customers’ purchase patterns, it can predict that a customer who buys bread will most probably buy butter.
Everyday example: Grouping similar songs together is clustering, while noticing that two items are often bought together is association.
Subcategories of Deep Learning: ANN and CNN
Deep Learning helps software train itself to do tasks using huge amounts of data. In deep learning, the machine is trained with large data, and this helps it train itself around the data. Such machines are intelligent enough to develop algorithms for themselves.
There are two types of Deep Learning models:
- Artificial Neural Networks (ANN)
- Convolutional Neural Network (CNN)
Artificial Neural Networks (ANN):
- Artificial Neural Networks are modelled on the human brain and nervous system.
- They can automatically extract features without input from the programmer.
- Every neural network node is basically a machine learning algorithm.
- ANN is useful for solving problems where the dataset is very large.
Convolutional Neural Network (CNN):
- Convolutional Neural Network is a Deep Learning algorithm.
- It can take an input image.
- It assigns importance, through learnable weights and biases, to different aspects or objects in the image.
- It is able to differentiate one object from another.
Everyday example: When a learner looks at many pictures and slowly learns what parts of a picture matter most, it is similar to the idea of extracting important features.
Artificial Neural Networks: Meaning and Layers
Neural networks are loosely based on the way neurons in the human brain behave. Their main advantage is that they can extract data features automatically without needing the programmer to provide those features.
A neural network is a way of organizing machine learning algorithms so that they can perform tasks. It is a fast and efficient way to solve problems where the dataset is very large, such as images.
A Neural Network has multiple layers. Each layer has several blocks called nodes. Each node has its own task. After completing its task, it passes information to the next layer.
The main layers are:
- Input layer
- Hidden layer
- Output layer
Input layer:
- The input layer is the first layer of a Neural Network.
- Its job is to receive data and feed it to the Neural Network.
- No processing happens in the input layer.
Hidden layer:
- Hidden layers come after the input layer.
- They are the layers where the main processing happens.
- They are called hidden because they are not visible to the user.
- Each node in a hidden layer has its own machine learning algorithm.
- It executes that algorithm on the data received from the input layer.
- Hidden layers perform computation using weights and biases.
- Information passes from one layer to another after the value from this calculation goes through a selected activation function.
- There can be many hidden layers.
- The number of hidden layers depends on the complexity of the function for which the network is configured.
- The number of nodes in each layer can also vary.
Output layer:
- The last hidden layer sends the final processed data to the output layer.
- The output layer gives the final output to the user.
- Like the input layer, the output layer does not process the data it receives.
- It is meant for user-interface.
The process of finding the correct output begins with trial and error. With each try, the weights are adjusted according to the error between the desired output and the network output. The network keeps learning through this process.
Real-world applications of neural networks include facial recognition, customer support chatbot and vegetable price prediction.
Everyday example: A group of students passing short notes from one row to another to identify an image is similar to information moving through layers.
How AI Makes a Decision
AI can make a decision using a structure like a perceptron. A perceptron uses inputs, weights, bias, a threshold and an output.
In the park example, the decision is whether to go out to the park. The factors that may influence the decision are:
- Do I have a jacket?
- Do I have an umbrella?
- Is it sunny now?
- What is the weather forecast for later?
These factors become inputs. Not all factors are equally important. Some are more important and some are less important. Importance is represented using weights. For example, being sunny now may be considered more important than the weather forecast later. Having a jacket may be considered more important than having an umbrella.
In the perceptron:
- The four factors become inputs.
- The inputs are represented as X values.
- The weights are represented as W values.
- Bias is also included.
- Bias has its own weight.
- The weighted values are added.
- The result is compared with a threshold.
- The output decides the action.
The actual values of weights may come from experience. A person may feel that if it is sunny now, the day is likely to remain bright, so that input may get higher importance. The values may also come from personal preference. One person may give more importance to the weather forecast than another person.
The value of the bias weight can also depend on personal preference. A cautious person may keep the bias weight higher, so the decision may move toward staying indoors unless the person is sure it will not rain. A daring person may keep it lower and may go out more easily.
There is no single right or wrong set of weights. The values affect the final result. This is why different people can make different decisions even when the situation is the same.
For calculation, yes and no can be converted into numbers 1 and 0. Bias is always taken as 1. After adding the weighted inputs and bias, the value is compared with the threshold. If the result is higher than the threshold, the decision can be to go out. If the result is lower than the threshold, the decision can be to not go out.
Everyday example: Two people may look at the same sky but make different choices because they give different importance to the same factors.
About these notes
These notes and questions were written by Syllabax from the official CBSE Class X Artificial Intelligence (417) Facilitator Handbook and the 2026-27 curriculum. This is not an official CBSE publication.