Artificial Intelligence Module

Foundations of Machine Learning & AI

Artificial Intelligence encompasses computational systems capable of performing tasks typically requiring human intelligence, formalized at the 1956 Dartmouth Summer Research Project organized by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. Machine learning (ML) is broadly categorized into supervised learning (training on labeled datasets for classification and regression), unsupervised learning (identifying latent patterns, clustering, and dimensionality reduction like PCA), and reinforcement learning (optimizing agent actions via reward signals in Markov Decision Processes). Fundamental algorithms include linear regression, logistic regression, support vector machines (SVM), decision trees, and random forests, evaluated using statistical metrics such as accuracy, precision, recall, and F1-score.

Practice Topic MCQs All Category Topics

Key Concepts & Examination Highlights

The term 'Artificial Intelligence' was officially coined at the 1956 Dartmouth Summer Research Project organized by John McCarthy.
Supervised learning algorithms train mathematical models on labeled input-output datasets to predict continuous or categorical outcomes.
Reinforcement learning optimizes agent decision-making policies through trial-and-error interactions governed by numerical reward and penalty functions.
Arthur Samuel coined the term Machine Learning in 1959, defining it as the subfield of computer science that gives computers the ability to learn without being explicitly programmed.
Unsupervised learning techniques include clustering algorithms (such as K-Means and DBSCAN) and dimensionality reduction techniques (such as Principal Component Analysis, PCA).
Reinforcement learning relies on an agent interacting with an environment through Markov Decision Processes (MDPs), optimizing a cumulative reward function using policies and value functions.
Decision trees use criteria such as Information Gain (based on Shannon Entropy) or Gini Impurity to determine optimal feature splits at each decision node.
Random Forest is an ensemble learning method that builds multiple decision trees using bootstrap aggregating (bagging) and feature randomization to reduce model variance.
Support Vector Machines (SVM) find the optimal separating hyperplane that maximizes the geometric margin between different classes in a high-dimensional feature space.
Gradient descent is a first-order optimization algorithm that iteratively updates model weights in the direction opposite to the gradient of the loss function to minimize error.
Evaluation metrics in classification include Accuracy, Precision, Recall (Sensitivity), and the F1-Score, which represents the harmonic mean of Precision and Recall.
Supervised learning algorithms are categorized into regression (predicting continuous numerical values) and classification (predicting discrete class labels).
Linear regression models the relationship between dependent and independent variables by fitting a linear equation of the form $y = wx + b$ , minimizing the Mean Squared Error (MSE) loss function.
Logistic regression uses the sigmoid activation function ( $\sigma(z) = \frac{1}{1 + e^{-z}}$ ) to map linear inputs into probabilities between 0 and 1 for binary classification.
Overfitting occurs when a machine learning model learns the training data and noise too closely, failing to generalize to unseen test data, characterized by high variance and low bias.
Underfitting occurs when a model is overly simplistic and fails to capture underlying data patterns, characterized by high bias and low variance.
Regularization techniques like L1 (Lasso) introduce absolute penalty terms ( $\lambda \sum |w|$ ) leading to sparse feature selection, while L2 (Ridge) adds squared penalty terms ( $\lambda \sum w^2$ ) to shrink weight magnitudes.
K-Nearest Neighbors (KNN) is an instance-based, non-parametric, 'lazy' learning algorithm that classifies new data points based on the majority class of their $k$ nearest geometric neighbors.
Naive Bayes classifiers apply Bayes' theorem assuming conditional feature independence, calculating posterior probability $P(A|B) = \frac{P(B|A)P(A)}{P(B)}$ for efficient text classification.
Principal Component Analysis (PCA) performs orthogonal linear transformations to project high-dimensional data onto orthogonal principal components with maximum variance.
Cross-validation techniques, such as $k$ -fold cross-validation, partition training datasets into $k$ subsets to validate model performance and assess generalization ability reliably.
Gradient Boosting algorithms (such as XGBoost, LightGBM, and CatBoost) build predictive models sequentially by training each new weak learner to predict the residual errors of preceding trees.
Hierarchical clustering groups data into a nested tree of clusters called a dendrogram, executed through either agglomerative (bottom-up) or divisive (top-down) approaches.
Receiver Operating Characteristic (ROC) curves plot the True Positive Rate against the False Positive Rate across varying classification thresholds, evaluated by the Area Under the Curve (AUC-ROC).
Hyperparameter tuning utilizes techniques such as Grid Search, Random Search, and Bayesian Optimization to find optimal model configurations that minimize validation loss.
Feature scaling techniques in machine learning include Min-Max Normalization (scaling features to $[0, 1]$ ) and Standardization (Z-score scaling to mean 0 and variance 1).
The Curse of Dimensionality refers to the exponential increase in volume of space and data sparsity that occurs when adding more feature dimensions to machine learning models.
The bias-variance tradeoff describes the fundamental tension between a model's error from erroneous assumptions (bias) and sensitivity to fluctuations in training data (variance).
Stochastic Gradient Descent (SGD) computes loss gradients and updates weights using a single randomly selected training sample per step, increasing optimization speed and noise.
Mini-batch Gradient Descent strikes a balance between SGD and full-batch gradient descent by updating model parameters on small subsets of training data (typically 32 to 512 samples).
AdaGrad (Adaptive Gradient Algorithm) scales learning rates adaptively for each parameter based on historical squared gradients, suitable for sparse data like text embeddings.
RMSprop modifies AdaGrad by using an exponentially decaying average of past squared gradients to prevent aggressive, monotonically decreasing learning rates.
The learning rate hyperparameter ( $\alpha$ or $\eta$ ) controls the step size taken during parameter updates; too high causes divergence, while too low causes slow convergence.
Lasso regression (L1 regularization) drives non-essential feature weights to exactly zero, effectively performing embedded feature selection during model training.
Ridge regression (L2 regularization) shrinks collinear feature weights toward zero but never sets them exactly to zero, preventing matrix singularity in multicollinear data.
Elastic Net regularization combines L1 and L2 penalty terms ( $\lambda_1 \sum |w| + \lambda_2 \sum w^2$ ) to handle groups of correlated features robustly.
The Gini Impurity index ( $I_G = 1 - \sum p_i^2$ ) measures the probability of misclassifying a randomly chosen element if it were labeled according to class distribution in the node.
Information Gain is defined as the reduction in Shannon Entropy ( $H(S) = -\sum p_i \log_2 p_i$ ) achieved by partitioning a dataset on a given attribute.
Pruning in decision trees (pre-pruning and post-pruning / cost-complexity pruning) removes leaf nodes that provide little predictive value to prevent overfitting.
AdaBoost (Adaptive Boosting), formulated by Freund and Schapire in 1995, trains weak learners sequentially by assigning higher weights to instances misclassified by earlier rounds.
The Kernel Trick in Support Vector Machines implicitly maps non-linearly separable data into higher-dimensional feature spaces using functions like the Radial Basis Function (RBF) kernel.
Support vectors are the critical data points residing closest to the separating hyperplane that define the margin boundaries of the classification boundary.
K-Means clustering minimizes the within-cluster sum of squares (WCSS / inertia) by iteratively updating cluster centroids until mathematical convergence.
The Elbow Method and Silhouette Analysis (scoring between -1 and +1) are standard heuristics used to determine the optimal number of clusters ( $k$ ) in unsupervised clustering.
DBSCAN (Density-Based Spatial Clustering of Applications with Noise) groups closely packed points based on density parameters ( $\varepsilon$ radius and MinPts), identifying arbitrary shapes and noise.
Singular Value Decomposition (SVD) decomposes a real matrix $X$ into $U \Sigma V^T$ , forming the mathematical foundation of latent semantic analysis and matrix factorization.
The t-SNE (t-Distributed Stochastic Neighbor Embedding) algorithm is a non-linear dimensionality reduction technique widely used for visualizing high-dimensional datasets in 2D or 3D scatter plots.
UMAP (Uniform Manifold Approximation and Projection) preserves both local and global topological data structure faster than t-SNE in high-dimensional manifold learning.
Confusion matrices tabulate True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN) to calculate precision, recall, and specificity.
Stratified $k$ -fold cross-validation ensures that each partitioned data fold preserves the exact class percentage distribution of the target variable present in the complete dataset.

Curriculum & Reference Sources: Association for the Advancement of Artificial Intelligence (AAAI), Stanford AI Lab, MIT CSAIL

Sample Solved Questions & Concept Explanations

8 Verified Concept Questions

Q1.EASY

What type of Machine Learning algorithm learns from labeled training data containing input features and ground-truth target outputs?

Q2.EASY

What type of Machine Learning algorithm groups and identifies hidden patterns or clusters in unlabeled data without predefined output tags?

Q3.EASY

In machine learning classification evaluation, what metric is defined as the harmonic mean of Precision and Recall?

Q4.EASY

What type of Machine Learning involves an agent learning optimal actions through trial-and-error interaction with an environment to maximize cumulative rewards?

Q5.EASY

What does the machine learning acronym "SVM" stand for in classification algorithms?

Q6.EASY

In statistics and machine learning, what is "Linear Regression" used for?

Q7.EASY

In machine learning model evaluation, what is the "Confusion Matrix" used for in classification tasks?

Q8.EASY

In machine learning, what is a "Hyperparameter" compared to a standard model "Parameter"?