Savitribai Phule Pune University,

Pune BOS Mechanical Engineering and Automobile Engineering Question

Bank (End Sem 2022-23)
Subject: Artificial Intelligence & Machine Learning (2019) (302049)
(Mechanical) Class: TE

UNIT 3- CLASSIFICATION AND REGRESSION ( 17 Marks)

Q1 Define following terms of Decision tree with neat sketch. 1. Root Node 2.
Leaf Node 3. Branch/Sub Tree 4. Pruning
Ans:-
Decision Tree Terminology with Sketch
A decision tree is a structure used in machine learning for classification and
regression. It is made up of nodes and branches that split data based on
features.

1. Root Node
• Definition:
o The topmost node of the tree.
o Represents the entire dataset and is split into subsets based on a
feature.
o It is the starting point for the decision-making process.
2. Leaf Node
• Definition:
o The final nodes of the tree that do not split further.
o They represent the output of the decision process, such as a class
label or value.
o Each leaf corresponds to a specific decision or outcome.
3. Branch/Sub-Tree
• Definition:
o The connecting lines from one node to another.
o A branch represents a decision path based on a feature value.
o A sub-tree is any smaller part of the tree, starting from a node and
including all its descendants.
4. Pruning
• Definition:
o A process of reducing the size of a decision tree by removing
unnecessary branches or nodes.
o Helps to prevent overfitting by simplifying the model while
maintaining accuracy.
o Types of pruning:
1. Pre-pruning: Stops tree growth early by setting constraints
(e.g., max depth).
2. Post-pruning: Removes branches after the tree is built,
based on validation results.
Sketch
Below is an example of a simplified decision tree diagram:
Root Node
┃
┌───┴───┐
Branch Branch
/ \
Sub-Tree Leaf Node
Detailed Diagram:
• The Root Node is at the top.
• Branches connect nodes.
 • Smaller trees branching out are Sub-Trees.
• The endpoints are Leaf Nodes.
• Pruned branches are not shown, representing simplification.

Q2 data and improves its generalization ability to make accurate predictions

on unseen data. The Give names of five different decision-tree algorithms?
Explain steps in the ID3 algorithm for learning decision trees. Five different
decision tree algorithms are:
1. ID3 (Iterative Dichotomiser 3) 2. C4.5 (Successor to ID3)
3. CART (Classification and Regression Trees) 4. CHAID (Chi-squared
Automatic Interaction Detection) 5. Random Forest
Ans:- Here are the simplified steps of the ID3 algorithm for building a decision
tree:
1. Select the root attribute:
• Calculate the information gain for each attribute in the dataset.
Information gain measures how well an attribute separates the data into
different classes.
• Choose the attribute with the highest information gain as the root node
of the tree.
2. Partition the dataset:
• Split the dataset into subsets based on the values of the selected
attribute.
• Create a child node for each subset.
3. Repeat for each child node:
• For each child node, use the remaining attributes (excluding the one
used in the parent node).
• Calculate the information gain for each remaining attribute and choose
the one with the highest gain as the next node.
• Split the data again based on this attribute, creating more child nodes.
4. Stopping conditions:
Repeat steps 1-3 until one of the following happens:
• All the data in a subset belongs to the same class (pure subset).
• There are no remaining attributes to split on.
• The tree reaches a maximum depth (you stop growing the tree).
5. Assign class labels to leaf nodes:
• If all instances in a leaf node belong to the same class, assign that class
label.
• If there are instances from different classes, assign the majority class
label in that leaf node.
Final Tree:
The tree formed can be used to classify new data based on their attribute values
by following the paths of the tree.
This is the basic process of building a decision tree using the ID3 algorithm.

Q.4 How does the random forest tree work for classification?
ANS:- Here's a simplified explanation of how Random Forest works:
1. Data Sampling (Bagging):
o Random Forest uses a method called bagging (Bootstrap
Aggregating).
o It randomly selects subsets of the original data by picking data
points with replacement. This means some data points can appear
multiple times in one subset, while others may not appear at all.
2. Building Decision Trees:
o For each subset of data, a decision tree is built. A decision tree is a
model that splits the data based on certain rules.
3. Feature Randomness:
o When building each decision tree, Random Forest introduces
another element of randomness.
 o Instead of looking at all the features (attributes), it randomly
selects a small number of features to consider at each split. This
helps make the trees different from each other.
4. Decision Tree Construction:
o For each subset of data, the decision tree is built by selecting the
best feature to split the data, using a rule like information gain or
the Gini index (which helps decide the best split).
5. Voting for Prediction:
o After all the decision trees are built, when a new data point needs
to be predicted, it goes through all the trees.
o Each tree gives a prediction, and the majority vote (the class that
most trees agree on) becomes the final prediction of the Random
Forest model.
In short, Random Forest combines multiple decision trees, each built with
random data subsets and features, and uses voting to make the final prediction.
This helps make the model more accurate and less prone to overfitting.

Q5 Write Differences between Bagging and Boosting in training of Random

Forest tree ?
ANS:- Here's a simplified explanation of Bagging and Boosting:
Bagging (Bootstrap Aggregating):
1. Sampling:
o Bagging creates multiple random subsets of the original data by
sampling with replacement. This means some data points may
appear more than once in a subset, while others may not appear
at all.
2. Parallel Training:
o Each subset of data is used to train a decision tree. These trees are
trained independently and at the same time, which is called
parallel training.
3. Equal Weighting:
 o All the decision trees in the Random Forest have equal importance
when making predictions. The final prediction is made by
averaging or taking the majority vote from all the trees.
4. Reducing Variance:
o Bagging helps reduce variance (errors caused by random
fluctuations in the data). Even if a decision tree overfits (fits too
closely to the data), combining the predictions of many trees
reduces this overfitting and leads to better accuracy.
5. Parallel Execution:
o Since each decision tree is trained independently, bagging can take
advantage of parallel computing to speed up the process.
Boosting:
1. Sequential Training:
o Boosting builds decision trees one after the other. Each new tree
tries to correct the mistakes made by the previous trees.
2. Weighted Sampling:
o In Boosting, each data point gets a weight. The more mistakes a
point makes, the higher its weight becomes, meaning future trees
will focus more on getting these points right.
3. Weighted Voting:
o When making predictions, Boosting gives more importance to the
trees that perform well. Trees that make fewer mistakes have
more influence on the final decision.
4. Reducing Bias:
o Boosting focuses on reducing bias (errors from not learning the
underlying patterns) by giving more attention to harder-to-classify
data points and learning from past mistakes.
5. Sequential Execution:
o The trees are built in a sequence. Each new tree depends on how
well the previous trees performed, and the process continues to
improve as more trees are added.
In summary:
• Bagging trains trees independently and reduces variance by combining
the predictions.
• Boosting trains trees in sequence, focusing on fixing the mistakes of the
previous trees, and reduces bias.

Q6 Is Naïve Bayes supervised or unsupervised algorithm? Explain Naïve Bayes

algorithm for suitable multiple feature classification example?
Ans:- Naive Bayes is a simple and efficient algorithm used for classification
tasks, where the goal is to predict the class (label) of new data based on
labeled examples. Here's how it works in simple terms:
Steps in Naive Bayes Algorithm:
1. Prepare the Data:
o You start with a labeled training dataset where each data point
(example) has features and belongs to a specific class. For
example, in an email classification task, the email might be labeled
as "spam" or "not spam." Each email also has features like certain
words, length of the email, etc.
2. Calculate Class Probabilities:
o First, calculate the probability of each class. For example, you find
out how many emails in the dataset are "spam" and how many are
"not spam." This gives you the prior probability of each class.
3. Calculate Feature Probabilities:
o Next, for each feature (like certain words in the email), calculate
how often that feature appears in emails of each class. For
example, how often the word "discount" appears in spam emails
vs. non-spam emails.
4. Apply Bayes' Theorem:
o Now, when you get a new email to classify, you calculate the
probability of the email belonging to each class (spam or not
 spam) using Bayes' Theorem. It looks at how likely the email's
features (like words or length) are in each class.
5. Predict the Class:
o Finally, the Naive Bayes algorithm picks the class that has the
higher probability based on the features of the new email. It
assumes that each feature (like the presence of the word
"discount") is independent of the others, which is the "naive"
assumption.
Example:
• Suppose you have a new email with the word "discount," a length of 300
words, and the phrase "buy now." The Naive Bayes algorithm will
calculate the likelihood of the email being "spam" or "not spam" based
on how common these features are in each class. The class with the
higher probability will be the prediction.
Why It's Useful:
• Naive Bayes is fast, simple, and works well even with many features (like
lots of words in a text). However, it assumes that features are
independent of each other, which might not always be true, but it still
often works well in practice.

Q8 Explain Support Vector Machine terminology

Ans:- Here's a simplified explanation of the key concepts in Support Vector
Machines (SVM):
1. Support Vectors: These are the data points that are closest to the
decision boundary (or hyperplane). They help to define where the
boundary should be. Only these points matter in creating the decision
boundary, while other points don't affect it.
2. Hyperplane: A hyperplane is like a dividing line (in 2D), plane (in 3D), or
higher-dimensional boundary that separates data into different
categories. In SVM, the goal is to find the best possible hyperplane that
divides the classes in the best way.
 3. Margins: The margin is the space between the hyperplane and the
closest data points from each class. SVM tries to make this margin as
wide as possible to ensure better accuracy and fewer errors. The points
closest to the hyperplane are called margin points. By maximizing the
margin, SVM helps the model generalize better and be more robust to
noise or outliers.
SVM is also capable of handling complex data (where classes are not linearly
separable) by transforming the data into a higher dimension using techniques
like kernel functions.
In short, SVM works by finding the best hyperplane that separates data and
makes sure this separation is as wide as possible to reduce errors. Only a few
key points (support vectors) affect the decision boundary, making the algorithm
efficient.

Q9 Explain Working of Support Vector Machine? What are hyper parameters

in SVM
Ans:- Here's a simplified explanation of how Support Vector Machines (SVM)
work:
How SVM Works:
1. Data Preparation: SVM starts with a dataset where each example has
features and belongs to one of two classes (for classification). These
examples are used to train the SVM.
2. Feature Transformation: If the data can't be separated with a straight line
(in 2D) or a flat plane (in 3D), SVM can use kernel functions to move the
data into a higher dimension where it can be separated. Common kernel
types include:
o Linear
o Polynomial
o Radial Basis Function (RBF)
o Sigmoid
3. Optimal Hyperplane: SVM tries to find the best "hyperplane" (the
decision boundary) that separates the two classes. The goal is to create a
 boundary that minimizes errors and maximizes the gap (called margin)
between the classes.
4. Margin Calculation: SVM calculates the margin by looking at the support
vectors, which are the closest data points to the decision boundary.
These support vectors determine where the boundary will be placed.
5. Margin Optimization: SVM then optimizes this margin. It solves a math
problem to find the best hyperplane that gives the largest margin and
smallest classification errors.
6. Classification: After training, the SVM uses the hyperplane to classify
new data. It checks which side of the boundary the new data point falls
on and assigns it the corresponding class.
Hyperparameters in SVM:
These are settings you need to define before training an SVM model. They
control how the SVM learns from the data and can impact its performance:
1. Kernel Type: The function that helps SVM transform the data. Different
kernels help separate the data in different ways.
o For example, linear kernels work well when the data is already
linearly separable, while RBF kernels are useful for more complex
data.
2. Regularization Parameter (C): This controls the trade-off between a large
margin and fewer misclassifications.
o A small C means SVM focuses more on getting a large margin,
even if some points are misclassified.
o A large C means SVM tries hard to correctly classify every point,
even if it results in a smaller margin.
3. Kernel Parameter: For certain kernels, you can set parameters like the
degree for polynomial kernels or bandwidth for RBF kernels. These
influence the flexibility of the decision boundary.
4. Class Weights: If some classes are more common than others
(imbalanced data), you can give more importance to the
underrepresented class to improve performance
 5. Tolerance: This controls how precise the optimization needs to be and
how long the algorithm will run before stopping. Lower tolerance gives
higher accuracy but takes longer to compute.
In summary, SVM is a powerful algorithm that separates classes using a
hyperplane and tries to maximize the margin between classes. It uses kernels
for complex data and has several settings (hyperparameters) that help tune its
performance.

Q10 Differentiate between logistic regression and linear regression?

Ans:- Here’s a simple explanation of the differences between logistic
regression and linear regression:
1. Type of Output (Dependent Variable):
o Logistic Regression: Used when the outcome is a category (like
"yes" or "no," or "true" or "false").
o Linear Regression: Used when the outcome is a number (like
predicting a price or temperature).
2. What They Predict:
o Logistic Regression: Gives a probability of something happening. It
predicts whether something falls into one of two categories (like
spam or not spam).
o Linear Regression: Predicts a numeric value, such as predicting
someone's salary based on their years of experience.
3. Mathematical Approach:
o Logistic Regression: Uses a special function (called the sigmoid
function) to turn the output into a probability between 0 and 1.
o Linear Regression: Uses a straight-line equation to make
predictions.
4. How the Coefficients Are Interpreted:
o Logistic Regression: The coefficients tell you how much each factor
(like age or income) affects the likelihood of the outcome
happening. This is often expressed as odds ratios.
 o Linear Regression: The coefficients tell you how much the
dependent variable will change for a one-unit change in the
independent variable (like how much salary increases with years
of experience).
5. Error Function:
o Logistic Regression: Uses a method called maximum likelihood
estimation to find the best-fitting model.
o Linear Regression: Uses mean squared error (MSE) to find the best
line that fits the data.
6. How to Measure Performance:
o Logistic Regression: You use accuracy, precision, recall, and F1-
score to see how well the model is doing.
o Linear Regression: You use mean squared error, mean absolute
error, and R-squared to check how close the predictions are to the
actual values.
In Short:
• Logistic Regression is for classifying things into categories (like yes/no or
true/false).
• Linear Regression is for predicting continuous values (like predicting a
price or a temperature).

Q11 How does K Nearest Neighbour (KNN) algorithm works?

Ans:- Here's a simplified explanation of how the K-Nearest Neighbors (KNN)
algorithm works:
1. Data Preparation:
o KNN uses a labeled training dataset, where each example has
features (characteristics) and is classified into a class (for
classification) or has a numerical value (for regression).
2. Feature Space:
 o Imagine a space where each feature is a dimension. Each instance
(or data point) is like a point in this space. For example, if you have
two features, it’s like a 2D graph where each point represents an
instance.
3. Distance Calculation:
o KNN looks at how far the new data point is from all the other
points in the training dataset. It uses distance measures like
Euclidean distance (straight-line distance) or others, to figure out
how similar the data points are.
4. K Nearest Neighbors:
o The algorithm then finds the k nearest points (neighbors) to the
new data point. You decide k, which is how many neighbors you
want to consider.
5. Majority Voting (Classification) / Averaging (Regression):
o For Classification: KNN looks at the classes (labels) of the k nearest
neighbors and picks the most common class as the prediction.
o For Regression: KNN looks at the values of the k nearest neighbors
and takes the average to predict the numerical value.
6. Decision Rule:
o For Classification: The class with the most neighbors wins.
o For Regression: The average value of the nearest neighbors is the
prediction.
Key Points:
• Non-Parametric: KNN doesn’t make assumptions about the data. It just
looks at the data and makes decisions based on it.
• Choosing k: The number of neighbors (k) is important. Too small a k
might lead to overfitting (too sensitive to noise), while too large a k
might lead to underfitting (not sensitive enough to the data).
• Feature Scaling: Since KNN calculates distances, it’s important to scale
the features so that no single feature dominates the distance calculation.
For example, if one feature is in a range of 0-1 and another is in the
 range of 1000-10000, the larger feature might overly influence the
result.

Q12 Differentiate between K Means and KNN.

Ans:- K-Means and K-Nearest Neighbors (KNN) are both machine learning
algorithms, but they serve different purposes:
K-Means (Clustering):
1. Type of Algorithm: It’s an unsupervised learning algorithm, meaning it
doesn't require labeled data.
2. Goal: The aim is to group data points into K clusters. The algorithm tries
to make similar data points belong to the same cluster by finding centers
(centroids) of the clusters and adjusting them until the groups are as
close as possible.
3. How It Works:
o It assigns each data point to a cluster based on the distance from
the cluster's centroid.
o It repeats this process, adjusting centroids until the clusters
stabilize.
4. Number of Clusters (K):
o You need to specify how many clusters (K) you want before
starting.
o Choosing the right K can be tricky, and tools like the elbow method
are used to help decide.
K-Nearest Neighbors (KNN) (Classification/Regression):
1. Type of Algorithm: It’s a supervised learning algorithm, which means it
requires labeled data to train.
2. Goal: The aim is to predict the class (for classification) or value (for
regression) of a new data point by looking at the K nearest neighbors
(similar data points).
3. How It Works:
 o It looks at the K nearest data points and either takes the most
common class (for classification) or the average value (for
regression) of these neighbors to make a prediction.
4. Number of Neighbors (K):
o You define how many neighbors (K) to consider when making
predictions.
o Choosing the right K is important for getting good results.
Summary:
• K-Means is used to group similar data (clustering), while KNN is used to
predict the class or value of a new instance based on its neighbors
(classification or regression).
 UNIT 4 - DEVELOPMENT OF ML MODEL (18 Marks)

Q1 Why data pre-processing is required? Explain techniques of preprocessing

data.
Ans;- Simplified Explanation of Data Preprocessing
Data preprocessing is a crucial step before analyzing data or using it for
machine learning. It involves cleaning and preparing raw data to make it ready
for analysis or modeling. Here's why it's important and what it involves:
Why is Data Preprocessing Important?
1. Improving Data Quality:
Raw data often has missing values, errors, or noise. Preprocessing fixes
these issues, ensuring accurate and reliable results.
2. Creating Better Features:
It helps in modifying or creating new data features (like scaling or
normalizing values) that make models work better.
3. Reducing Data Size:
Sometimes, there’s too much data to handle. Preprocessing can remove
less useful parts while keeping the important ones, improving efficiency
and preventing problems like overfitting.
4. Standardizing Data:
Data may come in different formats or units (e.g., dollars vs.
percentages). Preprocessing ensures everything is on the same scale for
fair comparison.
5. Converting Text to Numbers:
Machine learning models mostly work with numbers. Preprocessing
converts text-based data into numeric formats (e.g., using one-hot
encoding
Common Preprocessing Steps
1. Data Cleaning:
o Fill missing values (e.g., using the average).
o Remove or adjust unusual data points (outliers).
2. Data Transformation:
 o Scale data to ensure all features are on a similar scale (e.g.,
normalization).
o Apply transformations (like log or square root) to make data
distributions smoother.
3. Feature Selection:
o Pick the most important parts of the data for analysis, reducing
unnecessary information.
4. Dimensionality Reduction:
o Use techniques like PCA (Principal Component Analysis) to reduce
the size of the data while keeping the key information.
5. Handling Categorical Data:
o Convert categories into numbers (e.g., “red,” “blue,” “green” into
1, 2, 3).
o Use methods like one-hot encoding to create binary columns for
each category.
6. Splitting Data:
o Divide data into training, validation, and test sets to build, tune,
and test the model properly.
Preprocessing ensures that the data is clean, organized, and ready for accurate
analysis or modeling. The techniques you use depend on the type of data and
the goals of your project.

Q2 Compare Training data vs. Validation data vs. Test data. Training data,
validation data, and test data are all subsets of the overall dataset that serve
different purposes in the machine learning workflow. Here's a comparison of
these three types of data:
Ans:- 1. Training Data
• Purpose: This is the main data used to teach the machine learning model
how to work. It helps the model learn patterns and relationships
between inputs and outputs.
• Size: Usually makes up 60-80% of the total dataset.
• Labels: Includes both input features and their correct outputs (answers).
 • Use: The model learns from this data and adjusts itself to improve
accuracy.
2. Validation Data
• Purpose: This data is used to fine-tune the model's settings
(hyperparameters) and check its performance during training. It helps
prevent overfitting (when the model works well on training data but
poorly on new data).
• Size: Typically 10-20% of the total dataset.
• Labels: Includes both input features and correct outputs.
• Use: Helps decide the best version of the model by checking how well it
performs on this data during training.
3. Test Data
• Purpose: This data is used to evaluate how well the final model works on
new, unseen data.
• Size: Usually 10-20% of the total dataset.
• Labels: May or may not include correct outputs. It's often used only to
check performance.
• Use: The model makes predictions on this data, and its accuracy is
measured to estimate real-world performance.
Key Point:
• Why Split Data? Splitting the data into these three groups ensures that
the model is tested properly and can work well on new data, not just the
data it was trained on.

Q3 Enlist and explain steps involved in development of classification model.

Ans:- Steps to Build a Classification Model
1. Data Collection:
o Gather the data needed to train and test your model.
o Data can come from databases, APIs, or external sources.
2. Data Preprocessing:
o Clean and prepare the data for analysis.
 o Handle missing values, fix outliers, normalize or scale the data,
and encode non-numerical values.
o Balance the dataset if some classes have much more data than
others.
3. Feature Selection/Engineering:
o Identify the most important data features for your task.
o Use methods like correlation analysis or create new useful
features to improve the model’s performance.
4. Train-Test Split:
o Divide the data into two parts: one for training the model and one
for testing it.
o Common splits are 70-30 or 80-20. Alternatively, use cross-
validation to test the model multiple times.
5. Model Selection:
o Choose a classification algorithm based on your data and needs.
o Examples include logistic regression, decision trees, random
forests, SVM, or neural networks.
6. Model Training:
o Train the chosen algorithm using the training data.
o The model learns patterns by adjusting its internal settings
(weights) with techniques like gradient descent.
7. Model Evaluation:
o Test the trained model on the testing data.
o Use metrics like accuracy, precision, recall, F1 score, or AUC-ROC
to measure performance.
8. Model Tuning:
o If performance isn’t good enough, adjust model settings
(hyperparameters).
o Use techniques like grid search or random search to find the best
settings.
9. Model Deployment:
 o Put the model into use.
o This could mean integrating it into a system, creating an API, or
ensuring it works smoothly in real-world settings.
10.Model Maintenance:
• Regularly monitor the model to ensure it stays accurate.
• Update or retrain the model with new data to keep it effective.
This process ensures your classification model is well-built, accurate, and ready
for real-world use.

Q4 Explain 1. Overfitting 2. Under fitting 3. Exact Fit Model with suitable

sketch
Ans:- a. Overfitting
• What is it?
Overfitting happens when a machine learning model learns the training
data too well, including random noise or unnecessary details. This makes
it perform well on training data but poorly on new data.
• Signs of Overfitting:
1. Great results on training data but bad results on validation or
testing data.
2. The model is overly complicated and sensitive to small changes in
the data.
3. It captures irrelevant patterns (like noise or outliers) instead of
useful ones.
4. It struggles to make accurate predictions on new data.
b. Underfitting
• What is it?
Underfitting occurs when the model is too simple and fails to learn the
important patterns in the data. This leads to poor performance on both
training and testing data.
• Signs of Underfitting:
1. The model is too basic to capture the data's complexity.
 2. Poor accuracy on training and testing data.
3. The model misses important relationships in the data.
4. It makes overly generalized predictions with high error rates.
c. Exact Fit Model
• What is it?
An exact fit model is one that perfectly matches the training data with no
errors. While it might seem ideal, it’s usually a problem because it likely
overfits and won't work well on new data.
• Signs of an Exact Fit Model:
1. Predicts everything in the training data perfectly.
2. Has zero errors on training data.
3. Often struggles with new data, as it likely overfits by learning
irrelevant details.
Key Takeaway:
• Overfitting: Model is too complex, fits training data too closely, and fails
on new data.
• Underfitting: Model is too simple, misses important patterns, and
performs poorly overall.
• Exact Fit: Perfect on training data but often overfits, leading to poor
generalization.

Q5 What is hyper parameter tuning? Enlist different hyper parameter tuning

algorithms. Explain any two hyper parameters in Random Forest algorithm.
Ans:- Hyperparameter Tuning
• What is it?
It's the process of finding the best settings (hyperparameters) for a
machine learning model to improve its performance. Unlike other
parameters, hyperparameters are set by the user before training and are
not learned from the data.
• Why is it important?
Good hyperparameters make the model more accurate and better at
handling new data.
Techniques for Hyperparameter Tuning
1. Grid Search:
o Tests every possible combination of predefined hyperparameter
values.
o Example: Trying all combinations of learning rates and tree depths
for a model.
o Pro: Finds the best combination.
o Con: Time-consuming for many hyperparameters.
2. Random Search:
o Randomly picks hyperparameter combinations from a defined
range.
o Example: Testing 20 random values for tree depth or learning rate.
o Pro: Faster than grid search.
o Con: May miss the best combination.
3. Bayesian Optimization:
o Uses a smart strategy to guess which hyperparameters are likely to
perform best.
o Balances between testing new areas and improving known good
ones.
o Pro: Efficient for large hyperparameter spaces.
4. Genetic Algorithms:
o Inspired by evolution: creates a group of hyperparameter sets,
evaluates them, and evolves better ones over generations.
o Uses methods like mutation (small changes) and crossover (mixing
good sets).
o Pro: Good for complex models with many hyperparameters.
5. Gradient-based Optimization:
o Adjusts hyperparameters step by step based on how much they
improve the model.
o Pro: Works well for specific, continuous hyperparameters.
Example: Tuning Random Forest Hyperparameters
1. n_estimators (Number of Trees):
o Determines how many decision trees to include in the forest.
o More trees can improve accuracy but may also increase
computation time.
2. max_depth (Tree Depth):
o Controls how deep each tree can grow.
o Deeper trees capture complex patterns but risk overfitting.
o Balance depth to avoid overfitting or underfitting.
Key Takeaway:
Hyperparameter tuning ensures the model performs its best by adjusting
settings like the number of trees or learning rate. Choosing the right technique
depends on the complexity of the model and the resources available.

Q6 Explain with neat sketch K-fold cross-validation mode.

Ans:-
K-Fold ross-Validation
K-Fold Cross-Validation is a method to check how well a machine learning
model works on new data. It splits the dataset into parts (called folds) and
trains and tests the model multiple times to get a reliable performance
measure.

Steps in K-Fold Cross-Validation

1. Split the Dataset:
o Divide the dataset into K equal parts (e.g., 5 folds if K = 5).
2. Train and Validate:
o For each fold:
▪ Use one fold as the validation set (to test the model).
▪ Use the other K-1 folds as the training set (to train the
model).
 o Repeat this process K times, so each fold gets a chance to be the
validation set.
3. Measure Performance:
o After testing on the validation set in each fold, calculate
performance metrics like accuracy, precision, recall, or mean
squared error.
4. Average the Results:
o Take the average of the metrics from all K folds to get an overall
score for the model.
5. Hyperparameter Tuning:
o Test different hyperparameter settings (like learning rate or tree
depth) using K-Fold Cross-Validation to find the best combination.
Advantages of K-Fold Cross-Validation
1. Reliable Performance Estimate:
o Since the model is tested on different parts of the data, the results
are more trustworthy than a single train-test split.
2. Efficient Use of Data:
o Every data point is used for both training and testing, which makes
the most of the available data.
3. Detect Overfitting or Underfitting:
o If the model does well on training data but poorly on validation
data, it is overfitting.
o If the model performs poorly on both, it is underfitting.
4. Hyperparameter Optimization:
o Helps find the best settings for the model by testing multiple
configurations across different folds.
Important Note
• K-Fold Cross-Validation doesn’t replace a separate test set.
After using K-Fold for tuning and evaluation, always test the model on a
completely new, unseen test set to check its performance in real-world
scenarios.
Key Takeaway:
K-Fold Cross-Validation splits data into parts, trains and tests the model
multiple times, and averages the results for reliable evaluation. It's a great way
to improve model accuracy, tune hyperparameters, and prevent overfitting or
underfitting.

Q7 What is confusion matrix? Write any 2 x 2 confusion matrix showing, True

Positive, True Negative, False Positive, False negative? Define 1. Accuracy 1.
Precision 2. Recall
Ans:- Confusion Matrix
A confusion matrix is a table that helps evaluate how well a classification model
works by comparing its predictions to the actual results. It's especially useful in
binary classification (two classes: positive and negative).

Components of a 2x2 Confusion Matrix

Predicted Positive Predicted Negative
Actual True Positive (TP): Correctly False Negative (FN): Predicted as
Positive predicted as positive. negative but is actually positive.
Actual False Positive (FP): Predicted as True Negative (TN): Correctly
Negative positive but is actually negative. predicted as negative.

Key Metrics
1. Accuracy
o What it is: The percentage of correct predictions out of all
predictions.
o Formula: Accuracy=TP+TNTP+TN+FP+FN\text{Accuracy} = \frac{TP
+ TN}{TP + TN + FP + FN}
o Use: A simple measure of overall performance, useful for balanced
datasets.
2. Precision
o What it is: Of all the predictions labeled as positive, how many are
actually positive?
 o Formula: Precision= TPTP+FP\text{Precision} = \frac{TP}{TP + FP}
o Use: Important when false positives (FP) are costly, like in spam
detection.
3. Recall (Sensitivity/True Positive Rate)
o What it is: Of all the actual positive cases, how many were
correctly predicted?
o Formula: Recall=TPTP+FN\text{Recall} = \frac{TP}{TP + FN}
o Use: Critical when false negatives (FN) are costly, like in disease
detection.
Key Takeaway:
• A confusion matrix provides a detailed breakdown of a model's
performance.
• Metrics like accuracy, precision, and recall help evaluate the model's
strengths and weaknesses.
• Use the right metric based on the problem's needs (e.g., precision for
spam emails, recall for medical tests).

Q8 Explain following performance evaluators used for

interpretation/assessment of classification model 1. Cohen's Kappa
Coefficient 2. F Score 3. ROC Curve.
Ans:-
Key Performance Metrics for Classification Models
1. Cohen's Kappa Coefficient
o What it is: A measure of agreement between a model’s predictions
and actual labels, considering the possibility of random
agreement.
o Range:
▪ 1: Perfect agreement.
▪ 0: Agreement by chance.
▪ -1: Complete disagreement.
o When to use:
 ▪ Useful for imbalanced datasets or when class distribution is
uneven.
▪ More reliable than accuracy in such cases.
2. F1 Score
o What it is: A metric that combines precision and recall into one
number.
o Formula: F1 Score=2×Precision×RecallPrecision+Recall\text{F1
Score} = 2 \times \frac{\text{Precision} \times
\text{Recall}}{\text{Precision} + \text{Recall}}
o Range:
▪ 1: Perfect balance between precision and recall.
▪ 0: Poor performance.
o When to use:
▪ Ideal when both precision (avoiding false positives) and
recall (avoiding false negatives) are important, such as in
imbalanced datasets.
3. Receiver Operating Characteristic (ROC) Curve
o What it is: A graph showing the trade-off between the True
Positive Rate (TPR) and the False Positive Rate (FPR) at different
thresholds.
o Key Metric:
▪ AUC (Area Under the Curve): A single number summarizing
the model's ability to distinguish between positive and
negative cases.
▪ AUC close to 1: Excellent performance.
▪ AUC ~ 0.5: Poor discrimination (like random
guessing).
o When to use:
▪ Helps in evaluating the model’s performance across various
thresholds.
▪ Common in medical testing, fraud detection, and risk
assessment.
Why These Metrics Matter
• Cohen’s Kappa: Useful for datasets with uneven class distribution.
• F1 Score: Balances precision and recall, especially for imbalanced
problems.
• ROC and AUC: Shows how well the model distinguishes between classes
and its robustness to threshold changes.
These metrics provide a detailed understanding of a model's strengths and
weaknesses, guiding you to choose the best model for your task.
 UNIT 5 - REINFORCED AND DEEP LEARNING (18 Marks)
Q1 Explain how reinforcement learning is different from other machine learning paradigms?
Explain Reinforcement learning with a neat sketch.

Ans:- Reinforcement learning (RL) is a type of machine learning where an agent learns by interacting
with its environment. Here’s a simple breakdown of how it works and how it differs from other types
of machine learning:

1. Feedback Signal:

• The agent learns through feedback, like rewards or punishments.

• Unlike supervised learning (where answers are given), the agent figures things out by
interacting with the environment.

2. Sequential Decision-Making:

• The agent makes decisions step by step, based on its current state.

• It learns which actions lead to better outcomes over time.

3. Delayed Rewards:
• Unlike in supervised learning (where feedback is immediate), in RL, rewards may
come later.
• The agent needs to consider long-term effects of its actions.
4. Exploration vs. Exploitation:
• The agent has to decide whether to explore new actions (to discover better
strategies) or exploit what it already knows (to get rewards quickly).
• Balancing exploration and exploitation is crucial for learning.
5. Modeling the Environment:
• The agent builds a mental model of how its actions affect the environment.
• It learns to predict the outcome of actions and the rewards they will bring.
6. Sequential Training:
• The agent improves gradually by interacting with the environment over multiple
episodes.
• It learns from trial and error, adjusting its actions to perform better over time.
These aspects make RL different from other machine learning approaches and allow it to be applied
to dynamic and uncertain situations like robotics, games, or autonomous driving.

Q2 Define following terms for Reinforcement learning

Ans :-

1. Agent: The agent is like a decision-maker or learner. It interacts with the world (the
environment), makes choices (takes actions), and tries to do things that give it rewards. The
 goal of the agent is to figure out how to make the best decisions over time to get the most
rewards.

2. Environment: This is everything outside the agent that it interacts with. It could be a real-
world environment (like a robot moving around) or a computer simulation. The environment
shows the agent what’s going on (providing observations), accepts its actions, and gives
rewards based on what the agent does.
3. Reward: A reward is feedback that tells the agent if it did something good or bad. Positive
rewards encourage the agent to keep doing the same thing, negative rewards tell the agent
to avoid that action, and zero rewards give no feedback. The agent tries to get as many
rewards as possible.
4. State: A state is like a snapshot of the environment at a given time. It’s all the information
the agent needs to make a decision. In a good setup, the current state is enough for the
agent to figure out what to do next, without needing past information.
5. Policy: A policy is the set of rules or strategy the agent follows to decide what action to take
in a given state. It could be simple, where one state always leads to the same action, or
random, where the agent picks different actions with some probability.
6. Value: The value is like a prediction of how good a state or action is in the long run. It tells
the agent how much reward it can expect if it starts from a certain state or takes a certain
action. The agent uses this to decide the best choices to make over time.

Q3 Define Markov property. Explain why Markov property is applicable in solving

Reinforcement learning
Ans :- The Markov property is a concept in math and computer science, especially in reinforcement
learning and stochastic processes, that says:
Key Idea
At any given time, if you know where you are right now (the current state), you don't need to
remember how you got there (the past). The future only depends on the present, not the past.
For example:
If you're walking in a park and you need to decide your next step, all you need to know is where you
currently are. You don't need to think about how you entered the park or which paths you already
walked on.
Why is it useful in reinforcement learning?
1. Simplifies the problem
• Instead of keeping track of everything that happened before, the system only needs
to focus on the current situation.
• This makes the problem easier to work with and faster to solve.
2. No need for memory
• Decisions are made based only on the "now," not the "past." This makes calculations
simpler and avoids the hassle of storing and analyzing a lot of data.
3. Compact representation
 • By assuming the Markov property, all the important details about the system are
already captured in the current state.
• This means we can represent the system using simple tables or formulas instead of
needing to describe the entire history.
4. Breaking the problem into smaller parts
• The Markov property lets us think about each situation (or state) separately.
• This makes it easier to design algorithms that figure out the best actions and learn
from experience.
The Markov property helps by focusing only on the "present moment," making learning, planning,
and decision-making faster and easier. This is why it's so important in reinforcement learning, where
algorithms need to make decisions efficiently in complex environments.

Q4 Explain Bellman Equation in Reinforcement Learning. How Bellman equation significant in

maximization of rewards in Reinforcement learning.
Ans:- The Bellman equation is a key concept in reinforcement learning. It helps an agent figure out
how good a state or action is by connecting the immediate rewards and the future rewards it can
expect.
There are two main versions of the Bellman equation:
1. Bellman Expectation Equation
This version tells us how good a state is, assuming the agent follows a specific behavior or
policy ππ (a set of rules for choosing actions).
Simple formula:
V(s)=(Immediate Reward)+(Future Rewards with discount).V(s)=(Immediate Reward)+(Future Reward
s with discount).
In words:
• V(s) is the value of the current state ss.
• R(t+1)R(t+1) is the reward for taking an action.
• γγ (gamma) is the discount factor—it balances how much we care about future rewards
compared to immediate rewards.
• V(S(t+1))V(S(t+1)) is the value of the next state.
The Bellman Expectation Equation says:
"The value of a state ss is the current reward plus the discounted value of the next state, assuming
the agent keeps following the same policy."
2. Bellman Optimality Equation
This version helps the agent find the best possible strategy (optimal policy) by always choosing the
action that leads to the highest future reward.
Simple formula:
Q∗(s,a)=(Immediate Reward)+(Best Future Rewards).Q∗(s,a)=(Immediate Reward)+(Best Future Rewa
rds).
In words:
• Q(s, a)* is the best value for taking action aa in state ss.
• The equation looks at all possible future actions and picks the one with the highest value
(that’s the "max" part).
The Bellman Optimality Equation says:
"To find the best action in a state, combine the immediate reward and the maximum value of future
rewards, assuming the agent always makes the best choice."
Why is the Bellman equation important?
• Learning: It helps the agent learn how to behave by estimating the value of states and
actions.
• Decision-making: It allows the agent to figure out the best choices to maximize its total
reward.
• Iterative updates: Algorithms like value iteration and policy iteration solve the Bellman
equation repeatedly to improve the agent's understanding of the environment.
In simple terms, the Bellman equation is like a guidebook that teaches an agent how to balance
immediate rewards with long-term benefits, helping it make smarter decisions over time.

Q5 Explain Epsilon greedy policy in selecting dilemma of exploration and exploitation

trade off.
Ans :- The exploration-exploitation trade-off is fundamental in reinforcement learning (RL), as it
determines how an agent balances gaining new knowledge (exploration) versus utilizing what it
already knows to maximize rewards (exploitation). The epsilon-greedy policy is a widely used
method to manage this balance effectively.
Key Points of the Epsilon-Greedy Policy:
1. Exploration Rate (εε):
• A small positive constant (εε) is set initially (e.g., 0.1 or 0.2).
• It represents the probability of exploring new actions.
2. Exploration:
• With probability εε, the agent selects a random action from all available actions.
• This ensures the agent doesn't get stuck in suboptimal actions by discovering
potentially better alternatives.
3. Exploitation:
• With probability 1−ε1−ε, the agent chooses the action with the highest estimated
value based on its current knowledge.
• Exploitation maximizes the expected immediate reward by leveraging the agent's
experience.
How it Works:
• Initialization: Start with a relatively high εε to encourage exploration in the early stages.
• Learning: As the agent gathers experience and improves its knowledge, εε is often reduced (a
process called decay) to prioritize exploitation over time.
• Dynamic Balance: The agent alternates between trying new strategies and sticking to the
best-known actions, optimizing its learning and decision-making process.
Importance of the Exploration Rate (εε):
• Higher εε:
• Promotes exploration, allowing the agent to better understand the environment.
• Risk: The agent might spend too much time on suboptimal actions.
• Lower εε:
• Focuses on exploitation, leveraging the agent's current knowledge to maximize
rewards.
• Risk: The agent may miss discovering better actions due to limited exploration.
Practical Considerations:
• The value of εε depends on the problem's nature:
• Static Environments: A fixed or decaying εε is often suitable.
• Dynamic Environments: Adaptive exploration strategies may be required to handle
changing reward patterns.
The epsilon-greedy policy's simplicity and effectiveness make it a cornerstone in RL, ensuring a
systematic approach to achieving optimal decisions by balancing the exploration-exploitation
dilemma.

Q6 What do you understand from On policy and Off policy algorithm in reinforcement learning?
Explain Q- learning algorithm with flow diagram.
Ans:- On-Policy vs. Off-Policy
1. On-Policy Algorithms:
• These algorithms learn while following their current policy.
• The agent explores the environment and updates its decisions (policy) based on
what it does itself.
• Example: It’s like someone improving their basketball skills only by playing games
themselves.
2. Off-Policy Algorithms:
• These algorithms learn from other policies (could be from past actions or even
another agent's actions).
• The agent can explore in one way (behavior policy) but use the knowledge to
improve a different, better policy (target policy).
 • Example: Watching a coach or another player and learning better moves while still
practicing your own way.

What is Q-Learning?
Q-learning is a specific off-policy algorithm in reinforcement learning. It's like teaching an agent to
make the best decisions over time by learning from trial and error.

Steps of Q-Learning:
1. Initialize:
• Start by assigning random guesses (Q-values) for all possible situations (states) and
actions.
2. Choose an Action:
• The agent picks an action based on a rule, like trying something new sometimes
(exploration) or using the best-known choice (exploitation).
3. Take the Action:
• The agent performs the chosen action in the environment and sees:
• What happens next (new state).
• How much reward it gets immediately.
4. Update Knowledge:
• The agent adjusts its Q-value (knowledge) for the action it just tried, using this
formula:Q(s,a)←Q(s,a)+α×[r+γ×max⁡Q(s′,a′)−Q(s,a)]Q(s,a)←Q(s,a)+α×[r+γ×maxQ(s′,
a′)−Q(s,a)]
• Q(s,a)Q(s,a): Current value for action aa in state ss.
• αα: Learning rate (how fast we update our knowledge).
• rr: Immediate reward.
• γγ: Discount factor (how much we value future rewards).
• max⁡Q(s′,a′)maxQ(s′,a′): The best value we could get from the next state.
5. Repeat:
• Keep doing this until the agent becomes really good at choosing actions (Q-values
stabilize).

Why is Q-Learning Off-Policy?

It’s off-policy because:
• The agent learns the best way to act (target policy) by using the maximum reward estimates,
even if it’s exploring in a different way (behavior policy).
Why is Q-Learning Useful?
• It’s a flexible and powerful way to teach agents to solve problems.
• Applications include playing games, controlling robots, or optimizing systems like traffic
lights.

Q7 Explain SARSA algorithm.

Ans:- SARSA, which stands for State-Action-Reward-State-Action, is a method used in artificial
intelligence to help an agent (like a robot or a game character) learn how to make the best decisions
in various situations. Here’s a simplified breakdown of how SARSA works:
Step-by-Step Process of SARSA
1. Start with Random Guesses
• The agent begins by making random guesses about how good each possible action is
in different situations. These guesses are stored in a table known as the Q-table.
2. Choose an Action
• Based on its current guesses, the agent selects an action. It can either:
• Pick the action it believes is the best (using its Q-table).
• Try something random to explore new possibilities.
• This selection often uses a strategy called epsilon-greedy, which balances between
exploring new actions and exploiting known good actions.
3. Take the Action
• The agent performs the chosen action in its environment. For example, if it's
navigating a maze, it might move left, right, up, or down.
4. Observe Results
• After taking the action, the agent observes:
• The reward it received (for instance, +10 points for moving closer to the
goal).
• The new situation it finds itself in (referred to as the "next state").
5. Choose the Next Action
• In this new situation, the agent picks its next action using the same method as
before (e.g., epsilon-greedy).
6. Repeat
• The agent continues to repeat steps 3–6 until it completes its task, such as finding an
exit in a maze or reaching a goal.
Why is SARSA Unique? ;- SARSA is special because it learns based on the actions that the agent
actually takes rather than just potential actions. This means it adapts its learning to fit the strategy
(or policy) it's currently following, making it particularly useful in scenarios where sticking to a
specific approach is important.
Example: A Robot in a Maze
To illustrate SARSA's functionality:
• Step 1: The robot starts with random ideas about which direction might be good.
• Step 2: It tries moving up and realizes that this direction brings it closer to the exit; it updates
its knowledge accordingly.
• Step 3: Next, it moves left and learns from that result too.
• Step 4: Over time, through many trials and adjustments, it figures out the best path to
escape the maze efficiently.
SARSA is beneficial for various real-world applications like robot navigation, game playing, and any
task where learning through trial and error helps find optimal solutions.

Q8 What are characteristics of deep learning? What is the difference between deep learning and
machine learning?
Ans :- Deep learning is a specialized area within machine learning that utilizes deep neural networks
to process data. Here, we break down key concepts of deep learning and contrast them with
traditional machine learning approaches.
Key Concepts of Deep Learning
1. Deep Neural Networks (DNNs):
• Deep learning relies on deep neural networks, which consist of multiple layers of
nodes (neurons). Each layer transforms the input data into more abstract
representations, enabling the model to learn complex patterns and relationships in
the data
2. Representation Learning:
• Unlike traditional methods that require manual feature extraction, deep learning
algorithms automatically learn hierarchical representations from raw input data. This
means they can identify important features without human intervention
3. End-to-End Learning:

• Deep learning models can perform end-to-end learning, processing input data
directly to produce output without needing explicit feature engineering. This
streamlined approach simplifies the modeling process
4. Large-Scale Data Handling:
• These algorithms excel with large datasets, leveraging vast amounts of data to
enhance their performance and ability to generalize across new examples
5. Feature Extraction:
• DNNs automatically extract relevant features from raw data, allowing them to
uncover intricate patterns that might be missed by simpler models
 6. Non-Linear Transformations:
• Deep learning models can capture non-linear relationships through activation
functions and the layered structure of the network, enabling them to model complex
phenomena effectively
Differences Between Deep Learning and Traditional Machine Learning
1. Representation Learning:
• Deep learning automatically learns data representations, while traditional machine
learning often depends on manually crafted features
2. Feature Engineering:
• In deep learning, feature engineering occurs within the model itself, whereas
traditional methods typically involve manual input from domain experts
3. Hierarchy of Features:
• Deep learning can learn multiple levels of abstraction, allowing it to build a hierarchy
of features. In contrast, traditional methods usually work with a single level of
features
4. Data Requirements:
• Deep learning generally requires large amounts of labeled training data for effective
performance, while traditional machine learning can yield good results with smaller
datasets
5. Computational Resources:
• Training deep learning models is computationally intensive and often necessitates
powerful hardware like GPUs, whereas traditional algorithms can run on less
powerful machines
6. Problem Domains:
• Deep learning has achieved significant success in areas such as image and speech
recognition, natural language processing, and computer vision. Traditional machine
learning remains widely used for tasks like regression, classification, and clustering
across various domains

Q9 Explain working of biological neuron? Explain with neat diagram equivalence of biological
neuron and artificial neuron.
Ans :- Working of a Biological Neuron:
A biological neuron works by receiving, processing, and transmitting information through electrical
and chemical signals. Here's how it works step by step:
1. Dendrites: These are branched structures that receive signals from other neurons. These
signals are usually in the form of electrical impulses.
2. Cell Body (Soma): The cell body processes all the signals it receives. If the signals are strong
enough (i.e., they exceed a certain threshold), it generates an electrical impulse called
an action potential.
 3. Axon: The action potential travels along the axon, a long, slender fiber. The axon is coated
with myelin, a fatty substance that helps speed up the transmission of the electrical signal.
4. Synapse: At the end of the axon, there are tiny gaps called synapses. When the electrical
signal reaches the synapse, it triggers the release of chemical messengers called
neurotransmitters.
5. Neurotransmitters: These chemicals travel across the synapse and bind to receptors on the
dendrites of the next neuron. This transfers the signal to the next neuron, continuing the
communication.

Equivalence of Biological Neuron and Artificial Neuron:

While biological neurons and artificial neurons are quite different in their structure and complexity,
they share similarities in how they process and transmit information. Here's a comparison:

Component Biological Neuron Artificial Neuron

Dendrites receive electrical signals from other Inputs are numerical values (data or
Input neurons. features).

Not explicitly defined, but the strength of the Weights are numerical values that
Weights signal can vary. determine the importance of inputs.

Summing The artificial neuron sums up the

Signals The cell body (soma) sums up the inputs. weighted inputs.

The activation function (e.g., step

Activation Action potential is generated when the sum of function, sigmoid) decides whether the
Function inputs reaches a threshold. neuron fires.

Electrical signal (action potential) is sent down Output is calculated from the activation
Output the axon. function and sent to other neurons.

Neurons adapt through synaptic plasticity Neurons adjust weights through learning
Learning (strengthening/weaken of synapses). algorithms (e.g., gradient descent).

Diagram of Biological Neuron:

scss
Copy code
[Dendrites] [Cell Body (Soma)] [Axon] [Synapse] → [Next Neuron] | | | | (Receives (Processes (Sends
signal (Releases neurotransmitters) signals) signals) through myelin)
Diagram of Artificial Neuron:
css
Copy code
[Input 1] → [Weight 1] ↑ [Input 2] → [Weight 2] → [Summation] → [Activation Function] → [Output]
↑ [Input 3] → [Weight 3]
 • Inputs: Data values received by the artificial neuron.
• Weights: Values that adjust the importance of each input.
• Summation: The total of weighted inputs.
• Activation Function: Decides whether the neuron should activate or not (e.g., sigmoid,
ReLU).
• Output: The result that gets sent to other neurons or used for decision-making.
Summary:
• Biological neurons work using electrical and chemical signals to transmit information across
the nervous system.
• Artificial neurons are mathematical models that simulate this process, receiving input,
applying weights, summing them up, and using an activation function to decide the output.

Q12 Explain working of Convolutional Neural Network (CNN) with flow diagram. Define 1. Padding
2. Striding in CNN
Ans:- Simple Explanation of Convolutional Neural Networks (CNNs)
Convolutional Neural Networks (CNNs) are a type of deep learning model that is especially good at
analyzing visual data, such as images or videos. They are designed to automatically learn patterns
and features from input images and use these patterns to make predictions, like recognizing objects,
faces, or text. Here's a breakdown of how CNNs work:
1. Convolutional Layer:
This is the first step in processing an image.
• Input: The CNN starts with an image (or a feature map from a previous layer).
• Filters (Kernels): The CNN applies small filters (also called kernels) to the image. These filters
slide across the image, one piece at a time, looking for patterns like edges, corners, or
textures.
• Result: After applying these filters, the network creates feature maps. These feature maps
contain information about where specific patterns (e.g., edges) are located in the image.
Flow:
Input Image → Convolution Layer → Feature Maps
2. Activation Function (ReLU):
• After the convolution, the feature maps are passed through an activation function (usually
ReLU).
• ReLU simply changes negative values to zero and leaves positive values as they are. This step
helps the network learn complex patterns by adding non-linearity.
Flow:
Feature Maps → Activation Function (ReLU) → Transformed Feature Maps
3. Pooling Layer:
 • The Pooling Layer reduces the size of the feature maps, which helps the network run faster
and reduces the amount of computation.
• Pooling also makes the network more robust by making it less sensitive to small changes (like
slight shifts in the position of objects in an image).
• Common methods are Max Pooling (which takes the highest value in each region) and
Average Pooling (which takes the average).
Flow:
Feature Maps → Pooling Layer → Pooled Feature Maps
4. Fully Connected Layer:
• After several convolution and pooling steps, the network has learned many useful features
from the image. These features are passed to the Fully Connected Layer.
• In this layer, every neuron is connected to every neuron in the previous layer, and the
network uses these connections to make final predictions (like classifying the image as a cat
or dog).
• The output layer typically uses softmax for classification tasks (e.g., "cat", "dog").
Flow:
Pooled Feature Maps → Fully Connected Layer → Output (e.g., Classification Resul
5. Padding:
• Padding is the process of adding extra pixels (usually zeros) around the edges of the image.
• This helps preserve the spatial size of the image and ensures that the convolution process
can capture features at the borders of the image without losing information.
Purpose: To keep the image's dimensions and avoid shrinking after each convolution.
6. Striding:
• Striding refers to how much the filter moves when scanning the image.
• A stride of 1 means the filter moves one pixel at a time, and a stride of 2 means it moves two
pixels at a time.
• A larger stride reduces the size of the output feature map but speeds up computation, while
a smaller stride keeps more detail but requires more processing power.
Purpose: To control how much the feature map shrinks after convolution and pooling.
Summary:
A CNN works by passing an image through several layers that perform operations like convolution
(looking for patterns), activation (adding complexity), pooling (reducing size), and finally, fully
connected layers that make the prediction. Padding and striding help control how much the image
shrinks and ensure the network captures important details.
 UNIT 6:- APLLICATIONS
Q1 Explain human and machine interaction? Explain with any example.

Ans :- Human and machine interaction means how humans and machines communicate and work
together. Humans give instructions or commands, and machines process them to provide results or
perform tasks. This interaction happens in many ways, like typing, speaking, or using gestures.

Example: Using a Voice Assistant

Imagine you have a voice assistant like Siri, Google Assistant, or Alexa. Here's how it works:

1. You give a command: You wake up the voice assistant by saying "Hey Siri" or pressing a
button, then ask something like, "What's the weather today?"

2. Voice to text: The assistant listens to your question and converts your voice into text using a
program.

3. Understanding your request: The text is analyzed to figure out what you want. For example,
the assistant understands you're asking about the weather.

4. Processing the task: The assistant finds the weather information from the internet.

5. Responding to you: It turns the answer into speech and says something like, "Today's
weather is sunny, with a high of 25°C."

6. You can ask more: You hear the answer and can ask more questions or give another
command.

This example shows how humans and machines work together. The human asks a question, the
machine uses technology to understand and respond, and the process repeats until the task is done.

Human and machine interaction makes life easier by helping us quickly get information, perform
tasks, or control devices.

Q2 What is predictive maintenance? Explain different steps in predictive maintenance.

Ans :- Predictive maintenance means using data and smart algorithms (machine learning) to predict
when a machine or equipment might break down or need servicing. By doing this, we can fix
problems before they happen, saving time, money, and effort.

Steps of Predictive Maintenance

1. Data Collection
• Gather information from the machine.
• Use sensors to track things like temperature, vibrations, or pressure.
• Look at old maintenance records or logs.
2. Data Preprocessing
• Clean up the data so it’s usable.
• Fix errors, remove outliers (strange values), and fill in any missing pieces.
 • Make sure everything is consistent and easy to analyze.
3. Feature Engineering
• Pull out the important details from the data.
• For example, instead of just raw vibration readings, calculate averages or spot
patterns that show wear and tear.
4. Model Development
• Use machine learning techniques to create a model (like a smart program).
• Train this model with past data to understand patterns leading to equipment failure.
• Use methods like decision trees, regression, or even advanced tools like neural
networks.
5. Model Evaluation
• Test the model to check how well it predicts failures.
• Use performance metrics to ensure it’s accurate and reliable before using it live.
6. Deployment
• Put the model to work!
• Connect it to live machine data so it can monitor everything in real time.
7. Alert Generation
• If the model sees something unusual or predicts a problem, it sends an alert.
• This warning tells the maintenance team that something might go wrong soon.
8. Maintenance Action
• Maintenance teams use the alerts to take action.
• This could mean fixing a part, replacing it, or running a full check on the machine.
9. Continuous Monitoring and Refinement
• The process keeps going!
• The system gets smarter as it learns from new data and feedback. This makes future
predictions even better.
Why Is This Important?

Instead of waiting for machines to break or spending extra money on regular maintenance,
predictive maintenance helps fix things at just the right time. It makes work smoother, avoids
unexpected downtime, and reduces costs.
Q3 Explain any one mechanical engineering application where image-based classification can be
adopted.

Ans :- Image-based classification in mechanical engineering for quality control can be used to
automatically check if manufactured parts meet the required quality. Here's a simple breakdown of
how it works:

1. Capturing Images
Cameras take pictures of the manufactured parts from different angles to see them clearly.
2. Cleaning Up Images
The images are adjusted to remove blurriness, fix colors, and make the details more visible.
3. Identifying Features
Important details like the shape, texture, or color of the parts are pulled out. For example:
• Is the surface smooth?
• Are the edges sharp?
4. Teaching the System (Training)
The system is taught to recognize:
• Good parts (that meet quality standards).
• Defective parts (with issues).
Using labeled examples of "good" and "bad" parts, the system learns the difference.
5. Building the Model
A computer program (like a neural network) is created to analyze these features. This program learns
to spot defects by studying lots of example parts.
6. Testing the System
The program is tested to see if it can correctly classify new parts as good or defective. If it makes
mistakes, it is improved.
7. Using the System
Once the program is ready, it looks at new images of parts, decides if they are good or defective, and
gives a result.
8. Taking Action
Based on the result:
• Good parts go forward for use.
• Defective parts are sent back for fixing or discarded.

Why is this useful?

• Faster inspections: Machines work quicker than humans.
 • Better accuracy: Machines don’t get tired or make mistakes easily.
• Consistent quality: Ensures every part meets the same standard.
This method helps factories save time, money, and deliver better products.

q.4) Explain the steps involved in material inspection? How machine learning can be implemented
in material inspection.
Ans:- Material inspection involves assessing the quality and properties of materials to ensure they
meet specific standards and requirements. Machine learning can be implemented in material
inspection to automate and enhance the inspection process. Here are the steps involved in material
inspection and how machine learning can be applied:
1.
1. Collecting Data:
Gather information about the material using tools like cameras, sensors, or other
testing methods. This can include pictures, readings from sensors, or any data that
shows the material’s quality.
2. Cleaning the Data:
Before using the data, clean it up! Remove unnecessary bits (like noise or errors), fix
any missing parts, and organize it into a format that’s easy to analyze.
3. Identifying Important Details:
Pick out the most useful information from the data. For example, in images, this
could be color, texture, or shape—things that can tell you if the material is good or
bad.
4. Labeling Data for Training:
Label the data to show what "good" and "bad" materials look like. This helps the ML
model understand what it should learn.
5. Building a Smart Model:
Use ML techniques (like decision trees, neural networks, etc.) to create a system that
can predict material quality. It "learns" the patterns from the labeled data.
6. Teaching and Testing the Model:
Train the model on one part of the labeled data and test it on another part to see
how well it predicts the quality of materials. Evaluate its accuracy to ensure it works
properly.
7. Using the Model:
Once the model is ready, use it to inspect new materials. It analyzes the features and
gives a verdict: does the material meet the quality standards or not?
8. Making Decisions:
Based on the model's results, decide what to do next. If the material doesn’t meet
the required standards, take actions like rejecting it, inspecting it further, or sorting it
out.
Why Use Machine Learning?;- Machine learning automates this whole process, making inspections
faster, more accurate, and less dependent on manual work. It can handle large volumes of data,
adapt to changes, and provide consistent results!
Q5 Explain different applications in health care where AIML can be used.
Ans:- Artificial Intelligence (AI) and Machine Learning (ML) have numerous applications in healthcare
that can CO6 BL2 5 revolutionize the industry. Here are some areas where AI and ML can be used in
healthcare:
1. Medical Image Analysis
• Analyze X-rays, MRIs, and CT scans.
• Detect abnormalities like tumors or fractures.
• Assist radiologists in accurate diagnosis.
2. Disease Diagnosis and Risk Prediction
• Analyze symptoms, medical history, and lab results.
• Predict or diagnose diseases early (e.g., cancer, diabetes, heart disease).
• Help in risk assessment based on patient data patterns.
3. Drug Discovery and Development
• Analyze biological data to find potential drug targets.
• Predict drug effectiveness and side effects.
• Design new molecules for faster drug development.
4. Health Monitoring with Wearables
• Process data from smartwatches and health sensors.
• Track vital signs, activity levels, and sleep patterns.
• Identify anomalies for early health issue detection.
5. Electronic Health Record (EHR) Analysis
• Identify trends in patient data.
• Predict disease progression.
• Optimize treatment plans for personalized care.
6. Virtual Assistants and Chatbots
• Answer medical queries and provide recommendations.
• Assist in scheduling appointments.
• Triage patients and provide initial diagnosis or self-care tips.
7. Precision Medicine
• Analyze genetic and molecular data.
• Develop personalized treatments tailored to individuals.
• Reduce the risk of adverse drug reactions.
8. Clinical Decision Support Systems
 • Offer evidence-based treatment recommendations.
• Alert doctors about potential drug interactions or risks.
• Support healthcare professionals in decision-making.
9. Healthcare Resource Optimization
• Forecast patient demand using predictive models.
• Plan staff schedules and manage hospital beds efficiently.
• Optimize inventory of medicines and supplies.
These sub-points highlight the many ways AI and ML are transforming healthcare, making it more
efficient, accurate, and personalized.

Q6 Write a short note on use of AIML in traffic control

Ans :- AI (Artificial Intelligence) and ML (Machine Learning) can make traffic control much smarter
and more efficient. Here's how:
1. Traffic Flow Prediction: AI and ML can look at past traffic data, weather, and other factors to
predict how traffic will move. This helps control traffic signals and road configurations,
making traffic flow smoother.
2. Smart Traffic Lights: AI can change the timing of traffic lights in real time, based on how
much traffic is on the road. This reduces waiting times and keeps traffic moving.
3. Intelligent Traffic Management: AI systems can automatically detect accidents or road
problems and adjust traffic patterns to avoid delays. It can also alert authorities quickly so
they can fix the issue.
4. Detecting Traffic Incidents: AI can watch live video from cameras to spot accidents or other
problems on the road. This helps authorities respond faster and clear issues quickly.
5. Route Optimization: AI can suggest the best routes for drivers based on real-time traffic and
other conditions. This helps avoid traffic jams and saves time.
6. Smart Parking: AI helps find parking spots by using sensors and cameras. It guides drivers to
available spaces, reducing the time spent looking for parking.
7. Managing Traffic Demand: AI looks at when and where traffic is busiest, then suggests
solutions like encouraging people to use public transport, adjusting toll prices, or spreading
out traffic to avoid congestion.
8. Self-Driving Cars: AI helps self-driving cars communicate with each other and traffic systems
to avoid accidents and move more efficiently, improving traffic flow.
In short, AI and ML can make traffic systems smarter, reduce congestion, improve safety, and help
everything run more smoothly by reacting to traffic changes in real time.

Q7 Explain different steps in Dynamic system reduction.

Ans :- Dynamic system reduction is about making complex systems simpler while still keeping the
most important features and behaviours intact. Here's how it's done in simple terms:
 1. System Analysis:
• Study the structure and components of the system.
• Identify the important variables and relationships.
• Understand the system's inputs, outputs, and dynamics.
2. Variable Selection:
• Choose the most relevant variables that affect system behaviour.
• Exclude variables with minimal impact or those that can be approximated.
3. Order Reduction:
• Simplify the system by reducing the number of equations or state variables.
• Techniques used include modal analysis, balanced truncation, and model order
reduction.
4. Model Approximation:
• Approximate the dynamics of remaining variables.
• Use methods like linearization, curve fitting, interpolation, or regression analysis.
5. Validation and Error Analysis:
• Compare the behaviour of the reduced model with the original system.
• Perform error analysis to check how accurate the reduced model is.
6. Refinement and Iteration:
• Revisit previous steps if the reduced model isn’t accurate enough.
• Modify variable selection, order reduction, or approximation techniques as needed.
7. Model Integration:
• Integrate the reduced model into practical applications.
• Use it for simulation, control design, or optimization where simplicity is needed.
The goal is to balance simplicity with accuracy, so the reduced model captures the essential dynamics
of the original system.

Q8 Explain the use of Machine learning in process Optimization.

Ans :- Machine learning is a powerful tool used to improve processes in industries by making them
more efficient, cost-effective, and productive. Here’s a simple explanation of how it helps in process
optimization:
1. Predictive Analytics:
• Machine learning analyzes historical data.
• It predicts future outcomes like equipment failures or product quality.
• Helps in better scheduling, resource allocation, and decision-making.
 • Reduces downtime and improves efficiency.
2. Anomaly Detection:
• Machine learning detects unusual patterns in data.
• Identifies issues like equipment malfunctions or process deviations.
• Timely intervention to prevent failures.
• Reduces waste and optimizes resource usage.
3. Optimization Algorithms:
• Algorithms (e.g., genetic algorithms, reinforcement learning) find the best process
settings.
• They explore a wide range of possible solutions.
• Helps improve productivity, yield, and efficiency.
4. Quality Control and Defect Detection:
• Analyzes sensor data, images, or measurements.
• Detects defects or variations in product quality.
• Allows real-time adjustments to minimize waste and maintain high quality.
5. Energy Efficiency:
• Analyzes historical energy usage data.
• Identifies opportunities to save energy.
• Optimizes energy use patterns and suggests efficient strategies.
• Reduces costs and environmental impact.
6. Resource Allocation:
• Analyzes data to optimize the use of raw materials, inventory, and labor.
• Dynamically adjusts allocation based on real-time data.
• Minimizes waste and reduces operational costs.
7. Process Control and Automation:
• Integrates machine learning with control systems.
• Adjusts process variables (e.g., temperature, pressure, flow rates) in real time.
• Maintains optimal operating conditions, reduces variations, and improves outcomes.
These points highlight how machine learning can be applied to different areas of process
optimization for better efficiency, cost savings, and improved performance