Machine Learning : Basic to Core (Hindi & English)

This lesson explains one of the most important concepts in machine learning: why we divide data into training and testing sets before building a model. Students learn how splitting the dataset prevents overfitting, gives a realistic measure of model performance, and ensures the model can generalize to new, unseen examples.

We walk through the purpose of the training set, the role of the test set, common split ratios, and what can go wrong if the data is mixed incorrectly. The session also covers data leakage, overfitting symptoms, and the correct workflow to evaluate an ML model properly. By the end of this lecture, learners will understand how to apply a train–test split, why it matters, and how it protects a model from producing misleading results.

This lesson introduces learners to Linear Regression, one of the most fundamental and widely used algorithms in machine learning. The session breaks down how a regression model finds relationships between inputs and numeric outputs, how the “best-fit line” is calculated, and why linear regression remains a strong baseline model for many real-world prediction tasks.

Students explore practical examples, understand concepts like slope, intercept, error, and the Mean Squared Error (MSE) loss function, and learn how gradient descent helps the model improve over time. The lecture also highlights when linear regression works well, when it fails, and what its convex loss surface means for optimization. By the end, learners will have a solid intuition for how linear regression functions and where to apply it in real projects.

This session introduces Gradient Descent, the core optimization algorithm that powers almost every modern machine learning model—from linear regression to deep neural networks. Students learn how models improve by iteratively reducing error, how gradients indicate the direction of steepest descent, and why the learning rate controls the size of each update step.

The lesson covers the intuition behind the loss landscape, the difference between batch, stochastic, and mini-batch gradient descent, and common challenges such as overshooting, slow convergence, and getting stuck in non-convex regions. By the end of this lecture, learners will understand how parameters are optimized and why Gradient Descent is fundamental to model training.

This lesson explains Logistic Regression, one of the simplest and most effective algorithms for binary classification. Students explore how logistic regression transforms linear combinations of features into probabilities using the sigmoid function, and how a decision boundary is created to separate classes.

The lecture covers why linear regression fails for classification, how Binary Cross-Entropy loss guides the learning process, and how Gradient Descent optimizes weights for better predictions. Real-world examples—such as spam detection and medical diagnosis—illustrate how probability outputs can be interpreted and adjusted using thresholds. By the end, learners will understand when logistic regression works well, when it struggles, and how to use it confidently in classification tasks.

This lesson breaks down the most important metrics used to evaluate classification models and explains why accuracy alone is not enough to judge real performance. Students learn how a confusion matrix captures the true behavior of a model through true positives, false positives, true negatives, and false negatives. From there, the lecture explores accuracy, precision, recall, F1-score, and how each metric reveals different strengths and weaknesses of a classifier.

Practical examples—such as spam detection, fraud detection, and medical screening—help learners understand why certain metrics are preferred in specific real-world scenarios. The session also introduces ROC curves and AUC to show how ranking-based evaluation works. By the end of this lecture, learners will know how to select the right metric for a given problem and how to interpret model performance beyond a single number.

This lesson introduces learners to Decision Trees, one of the most intuitive and widely used models in classical machine learning. The session explains how trees split data into smaller, more meaningful groups by asking feature-based questions, how impurity measures like Gini and Entropy guide each split, and why the structure of a tree mirrors real decision-making processes.

Through clear visual intuition and practical examples, students learn how root nodes, branches, and leaves translate into classification or regression outputs. The lecture also explores common pitfalls—such as overfitting and instability—and demonstrates how techniques like pruning help improve generalization. By the end of this session, students will understand when decision trees work well, when they struggle, and why they serve as the foundation for powerful ensemble methods like Random Forests and Gradient Boosted Trees.

This lecture introduces Random Forests and explains why combining many simple decision trees produces a model that is far more accurate and stable than any single tree. Learners will understand how randomness in data sampling and feature selection helps reduce overfitting, improve generalization, and make tree-based models more reliable in real-world scenarios.

The session breaks down the intuition behind bagging, how individual trees vote to make final predictions, and why randomness is a strength rather than a weakness. Practical examples show how Random Forests handle noisy data, nonlinear relationships, and mixed feature types better than standalone decision trees. By the end of this lecture, students will have a clear mental model of how Random Forests work, when to use them, and why they are one of the most widely used algorithms in applied machine learning.

This lecture introduces the k-Nearest Neighbors algorithm and explains how predictions can be made without training a traditional model. Learners explore the core intuition behind kNN—making decisions based on similarity—and understand how distance, feature scaling, and the choice of k directly affect model behavior. The session highlights why kNN is simple in theory yet surprisingly powerful in practice.

Through practical examples, the lecture shows how kNN handles classification and regression tasks, why it is sensitive to noise and high-dimensional data, and how different distance metrics influence results. Students also learn when kNN is a good choice, when it becomes inefficient, and how it compares to other algorithms in terms of interpretability and performance. By the end of this session, learners will have a strong conceptual foundation to use kNN effectively and avoid common pitfalls.

This lecture builds an intuitive understanding of Support Vector Machines by focusing on the geometric principles that drive the algorithm. Learners explore how SVM finds the most confident decision boundary by maximizing the margin between classes and why only a small subset of data points—called support vectors—determine the final model. Rather than emphasizing formulas, the session explains how distance, margins, and boundary placement influence generalization.

The lecture also covers how SVM handles overlapping data through soft margins, the role of the penalty parameter, and the intuition behind kernel methods for separating non-linear data. Practical examples help learners understand when SVM performs well, when it struggles, and how it differs from probabilistic classifiers. By the end of this session, students will have a strong conceptual framework for using SVM effectively and understanding its strengths and limitations in real-world applications.

Today’s focus is on understanding why and when dimensionality reduction matters, and how PCA helps simplify complex data without throwing away what’s important.

We’ll start by looking at the problem PCA is designed to solve—high-dimensional data that is hard to visualize, interpret, or model effectively. From there, you’ll build intuition around variance, directions of maximum information, and how PCA transforms data into a lower-dimensional space.

The emphasis is on conceptual clarity. You’ll learn what principal components represent, how they are chosen, and what you gain and lose when applying PCA. We’ll also discuss common misunderstandings, such as when PCA helps model performance and when it doesn’t.

By the end of this session, you should be able to explain PCA in plain terms, recognize appropriate use cases, and confidently include it as part of a Machine Learning workflow—especially during exploratory analysis and feature preprocessing.

This lecture explains why machine learning models often perform extremely well on training data yet fail when exposed to new, unseen data. Learners explore the difference between genuine learning and memorization, and how overfitting emerges from excessive model complexity, limited data, and noise. The session introduces the bias–variance tradeoff and shows how training and validation performance reveal whether a model is underfitting or overfitting.

The lecture then focuses on regularization as a practical way to control model complexity. Students learn how techniques such as L1 and L2 regularization, parameter constraints, early stopping, and model-specific controls help improve generalization and reduce performance gaps between training and validation data. By the end of this session, learners will understand how to diagnose overfitting and apply regularization strategies to build more reliable, production-ready models.