{"pinia":{"ui":{"prefs":{"theme":"system","textScale":"md","fontFamily":"sans","reducedMotion":false,"highContrast":false}},"academy":{"universities":[],"universitySummaries":[{"slug":"rtu","name":"RTU","fullName":"Rajasthan Technical University","branches":[{"slug":"first-year-common","name":"First Year (Common)"},{"slug":"computer-science","name":"Computer Science"},{"slug":"mechanical","name":"Mechanical"},{"slug":"civil","name":"Civil"},{"slug":"ai-data-science","name":"AI & Data Science"},{"slug":"ee-ec-eee","name":"EE / EC / EEE"}]},{"slug":"aktu","name":"AKTU","fullName":"Dr. A. P. J. Abdul Kalam Technical University","branches":[{"slug":"first-year-common","name":"First Year (Common)"},{"slug":"computer-science","name":"Computer Science"}]}],"activePapers":[{"id":"machine-learning-2023","title":"Machine Learning (2023)","subjectSlug":"machine-learning","subjectName":"Machine Learning","universitySlug":"rtu","branchSlug":"computer-science","year":3,"semester":6,"examYear":2023,"paperCode":"6E7102","questions":[{"id":"v2-machine-learning-2023-pA-q1","number":1,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is the difference between supervised and unsupervised learning?"}],"answer":{"summary":"Supervised learning uses labeled data, while unsupervised learning works with unlabeled data.","explanation":[{"type":"paragraph","content":"Supervised learning involves training a model on a labeled dataset where the target variable is known. In contrast, unsupervised learning aims to find hidden patterns or structures in data without predefined labels."}]}},{"id":"v2-machine-learning-2023-pA-q2","number":2,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Define overfitting in machine learning."}],"answer":{"summary":"Overfitting occurs when a model learns the training data too well, capturing noise and failing to generalize.","explanation":[{"type":"paragraph","content":"Overfitting is a modeling error that occurs when a function is too closely fit to a limited set of data points. It results in high accuracy on the training set but poor performance on unseen test data."}]}},{"id":"v2-machine-learning-2023-pA-q3","number":3,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is a confusion matrix?"}],"answer":{"summary":"A confusion matrix is a table used to evaluate the performance of a classification model.","explanation":[{"type":"paragraph","content":"It compares actual target values with those predicted by the machine learning model. The matrix displays true positives, true negatives, false positives, and false negatives."}]}},{"id":"v2-machine-learning-2023-pA-q4","number":4,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Explain the bias-variance trade-off."}],"answer":{"summary":"It is the balance between a model's simplicity (bias) and its flexibility (variance) to minimize total error.","explanation":[{"type":"paragraph","content":"High bias can cause an algorithm to miss the relevant relations between features and target outputs (underfitting). High variance can cause an algorithm to model the random noise in the training data (overfitting)."}]}},{"id":"v2-machine-learning-2023-pA-q5","number":5,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is the purpose of a cost function?"}],"answer":{"summary":"A cost function measures the error between predicted values and actual target values.","explanation":[{"type":"paragraph","content":"The goal of training a machine learning model is to minimize this cost function. Common examples include Mean Squared Error (MSE) for regression and Cross-Entropy for classification."}]}},{"id":"v2-machine-learning-2023-pA-q6","number":6,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Define the term 'epoch' in neural networks."}],"answer":{"summary":"An epoch represents one complete pass of the entire training dataset through the neural network.","explanation":[{"type":"paragraph","content":"During an epoch, the weights of the neural network are updated based on the chosen optimization algorithm. Multiple epochs are usually required for the network to converge to a good solution."}]}},{"id":"v2-machine-learning-2023-pA-q7","number":7,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is the kernel trick in SVM?"}],"answer":{"summary":"The kernel trick transforms data into a higher-dimensional space to make it linearly separable.","explanation":[{"type":"paragraph","content":"It allows SVMs to solve non-linear classification problems efficiently without explicitly computing the coordinates of the data in that higher-dimensional space. Common kernels include Polynomial and RBF."}]}},{"id":"v2-machine-learning-2023-pA-q8","number":8,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Mention two techniques to handle missing data."}],"answer":{"summary":"Missing data can be handled using imputation or deletion.","explanation":[{"type":"paragraph","content":"1. Imputation: Filling missing values with the mean, median, or mode of the column. 2. Deletion: Removing rows or columns that contain significant amounts of missing data."}]}},{"id":"v2-machine-learning-2023-pA-q9","number":9,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is the difference between classification and regression?"}],"answer":{"summary":"Classification predicts categorical labels, while regression predicts continuous numerical values.","explanation":[{"type":"paragraph","content":"Classification algorithms, like Logistic Regression or SVM, categorize data into classes (e.g., spam or not spam). Regression algorithms, like Linear Regression, forecast quantities (e.g., house prices)."}]}},{"id":"v2-machine-learning-2023-pA-q10","number":10,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Name any two evaluation metrics for classification models."}],"answer":{"summary":"Common classification metrics include Accuracy, Precision, Recall, and F1-Score.","explanation":[{"type":"paragraph","content":"Accuracy measures the overall correctness, while Precision measures the ratio of correctly predicted positive observations to total predicted positives. F1-Score is the harmonic mean of Precision and Recall."}]}},{"id":"v2-machine-learning-2023-pB-q11","number":11,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Explain Logistic Regression. How does it differ from Linear Regression?"}],"answer":{"summary":"Logistic Regression is used for binary classification, whereas Linear Regression is used for continuous prediction.","explanation":[{"type":"paragraph","content":"Logistic Regression applies the sigmoid function to the linear combination of inputs to output a probability between 0 and 1. Linear regression, on the other hand, fits a straight line to the data to predict a continuous numerical value. Logistic regression uses cross-entropy loss, while linear regression uses mean squared error."}]}},{"id":"v2-machine-learning-2023-pB-q12","number":12,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Describe the steps involved in the K-Means clustering algorithm."}],"answer":{"summary":"K-Means divides data into K clusters by minimizing the variance within each cluster iteratively.","explanation":[{"type":"paragraph","content":"The steps are: 1. Randomly initialize K cluster centroids. 2. Assign each data point to the closest centroid. 3. Recalculate the centroids as the mean of all points assigned to that cluster. 4. Repeat steps 2 and 3 until the centroids no longer change significantly (convergence)."}]}},{"id":"v2-machine-learning-2023-pB-q13","number":13,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"What is cross-validation? Explain k-fold cross-validation."}],"answer":{"summary":"Cross-validation evaluates a model's performance on unseen data. K-fold splits data into K equal subsets.","explanation":[{"type":"paragraph","content":"In K-fold cross-validation, the dataset is divided into K partitions. The model is trained on K-1 partitions and tested on the remaining 1 partition. This process is repeated K times, with each partition acting as the test set exactly once. The final performance metric is the average of the K results."}]}},{"id":"v2-machine-learning-2023-pB-q14","number":14,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"How does a Support Vector Machine (SVM) handle non-linearly separable data?"}],"answer":{"summary":"SVM uses the kernel trick to map data to a higher-dimensional space where a linear hyperplane can separate the classes.","explanation":[{"type":"paragraph","content":"Standard linear SVMs fail when data points overlap non-linearly. By applying kernel functions (like Radial Basis Function (RBF), Polynomial, or Sigmoid), SVM computes the dot product of data points in a transformed high-dimensional space. This allows it to construct a non-linear decision boundary in the original feature space without explicitly calculating the transformation."}]}},{"id":"v2-machine-learning-2023-pB-q15","number":15,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Explain the Random Forest algorithm. What are its advantages?"}],"answer":{"summary":"Random Forest is an ensemble method that builds multiple decision trees and merges them for a more accurate and stable prediction.","explanation":[{"type":"paragraph","content":"It introduces randomness by training each tree on a random subset of the data (bootstrapping) and considering a random subset of features at each split. Advantages include high accuracy, robustness to overfitting (compared to individual decision trees), and the ability to estimate feature importance automatically."}]}},{"id":"v2-machine-learning-2023-pB-q16","number":16,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Discuss the concept of Gradient Descent with a suitable example."}],"answer":{"summary":"Gradient Descent is an optimization algorithm used to minimize the cost function by iteratively updating model parameters.","explanation":[{"type":"paragraph","content":"It calculates the gradient of the cost function with respect to the parameters. The parameters are then adjusted in the opposite direction of the gradient by a factor of the learning rate. For example, in Linear Regression, it iteratively tweaks the slope and intercept to find the line of best fit that minimizes Mean Squared Error."}]}},{"id":"v2-machine-learning-2023-pB-q17","number":17,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"What is PCA? Explain its significance in dimensionality reduction."}],"answer":{"summary":"Principal Component Analysis (PCA) transforms correlated variables into a smaller number of uncorrelated principal components.","explanation":[{"type":"paragraph","content":"PCA finds the directions (principal components) of maximum variance in high-dimensional data and projects the data onto a lower-dimensional subspace. This significantly reduces computational cost, mitigates the curse of dimensionality, and aids in data visualization while retaining the most critical information."}]}},{"id":"v2-machine-learning-2023-pC-q18","number":18,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Explain the Backpropagation algorithm in Artificial Neural Networks with a detailed step-by-step mathematical derivation."}],"answer":{"summary":"Backpropagation systematically updates neural network weights by calculating the gradient of the loss function using the chain rule.","explanation":[{"type":"paragraph","content":"### Introduction to Backpropagation\nBackpropagation is the backbone of neural network training. It propagates the error backwards from the output layer to the input layer to update weights using Gradient Descent."},{"type":"paragraph","content":"### Step-by-Step Derivation\n1. **Forward Pass:** The input `X` is fed through the network. Pre-activations `z` and activations `a` are computed for each layer.\n2. **Loss Calculation:** The total error is calculated using a loss function like Mean Squared Error or Cross-Entropy.\n3. **Backward Pass:** We compute the gradient of the error with respect to each weight using the chain rule.\n4. **Weight Update:** Weights are updated proportional to the negative gradient, controlled by the learning rate."},{"type":"paragraph","content":"By repeatedly applying these steps over multiple epochs, the network converges to a state with minimal loss."},{"type":"svg","src":"/assets/diagrams/rtu/cs/3/machine-learning/diagram-18.svg","alt":"Diagram","caption":"Diagram for Question 18"}]}},{"id":"v2-machine-learning-2023-pC-q19","number":19,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Discuss Decision Trees in detail. Explain the concepts of Entropy, Information Gain, and the ID3 algorithm."}],"answer":{"summary":"Decision Trees split data based on feature values to maximize information gain and minimize entropy.","explanation":[{"type":"paragraph","content":"### Decision Tree Overview\nA decision tree is a flowchart-like structure where each internal node represents a feature, each branch represents a decision rule, and each leaf node represents an outcome."},{"type":"paragraph","content":"### Entropy and Information Gain\n- **Entropy:** A measure of impurity or randomness in the data.\n- **Information Gain:** The reduction in entropy after splitting the dataset on an attribute. The attribute with the highest gain is chosen as the splitting node."},{"type":"paragraph","content":"### ID3 Algorithm\nThe ID3 (Iterative Dichotomiser 3) algorithm builds trees top-down:\n1. Calculate the entropy of the current dataset.\n2. For each attribute, calculate the Information Gain.\n3. Select the attribute with the highest Information Gain as the decision node.\n4. Split the dataset and recursively apply the process to the child nodes until all instances belong to the same class."},{"type":"svg","src":"/assets/diagrams/rtu/cs/3/machine-learning/diagram-19.svg","alt":"Diagram","caption":"Diagram for Question 19"}]}},{"id":"v2-machine-learning-2023-pC-q20","number":20,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Provide a comprehensive explanation of Ensemble Learning techniques. Compare Bagging and Boosting in detail."}],"answer":{"summary":"Ensemble learning combines multiple weak learners to form a strong learner. Bagging reduces variance, while Boosting reduces bias.","explanation":[{"type":"paragraph","content":"### Ensemble Learning Concept\nInstead of relying on a single model, ensemble methods use multiple models to obtain better predictive performance. The consensus of multiple models often yields a more robust prediction."},{"type":"paragraph","content":"### Bagging (Bootstrap Aggregating)\nBagging builds multiple independent models in parallel using random subsets of the training data (with replacement). The final prediction is an average (regression) or majority vote (classification). Random Forest is a prime example. It primarily reduces variance and prevents overfitting."},{"type":"paragraph","content":"### Boosting\nBoosting builds sequential models where each new model attempts to correct the errors made by the previous ones. Data points that were misclassified get higher weights. Examples include AdaBoost and Gradient Boosting. It primarily reduces bias and improves model accuracy."},{"type":"svg","src":"/assets/diagrams/rtu/cs/3/machine-learning/diagram-20.svg","alt":"Diagram","caption":"Diagram for Question 20"}]}},{"id":"v2-machine-learning-2023-pC-q21","number":21,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Elaborate on Support Vector Machines (SVM). Explain margins, hyperplanes, and the optimization problem involved."}],"answer":{"summary":"SVM finds the optimal hyperplane that maximizes the margin between classes using support vectors.","explanation":[{"type":"paragraph","content":"### Hyperplanes and Margins\nIn an N-dimensional space, a hyperplane is a flat affine subspace of dimension N-1. SVM seeks the hyperplane that separates the classes with the maximum margin, which is the distance between the hyperplane and the nearest data points from either class. These nearest points are called **Support Vectors**."},{"type":"paragraph","content":"### The Optimization Problem\nFor linearly separable data, the goal is to maximize the margin, which mathematically converts to a constrained convex optimization problem. We use Lagrange multipliers to solve it. The dual formulation allows the use of kernels for non-linear decision boundaries."},{"type":"paragraph","content":"### Soft Margin\nWhen data is not perfectly separable, SVM introduces slack variables to allow for some misclassifications, balancing the trade-off between maximizing the margin and minimizing the classification error using a regularization parameter C."},{"type":"svg","src":"/assets/diagrams/rtu/cs/3/machine-learning/diagram-21.svg","alt":"Diagram","caption":"Diagram for Question 21"}]}},{"id":"v2-machine-learning-2023-pC-q22","number":22,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Explain Principal Component Analysis (PCA) algorithm step-by-step. How does it compute the principal components?"}],"answer":{"summary":"PCA performs eigenvalue decomposition on the covariance matrix to find the directions of maximum variance.","explanation":[{"type":"paragraph","content":"### Introduction to PCA\nPCA is an unsupervised linear dimensionality reduction technique. It transforms the original variables into a new set of orthogonal variables called Principal Components, ordered by the amount of variance they explain."},{"type":"paragraph","content":"### Algorithm Steps\n1. **Standardization:** Scale the data to have a mean of 0 and variance of 1.\n2. **Covariance Matrix:** Compute the covariance matrix of the standardized data to understand feature relationships.\n3. **Eigen Decomposition:** Calculate the eigenvalues and eigenvectors of the covariance matrix. The eigenvectors represent the directions (principal components), and eigenvalues represent their magnitude (variance explained).\n4. **Sort and Select:** Sort the eigenvectors by decreasing eigenvalues. Choose the top k eigenvectors to form a projection matrix W.\n5. **Transform Data:** Multiply the original standardized dataset by W to obtain the new k-dimensional feature subspace."},{"type":"svg","src":"/assets/diagrams/rtu/cs/3/machine-learning/diagram-22.svg","alt":"Diagram","caption":"Diagram for Question 22"}]}}]},{"id":"machine-learning-2024","title":"Machine Learning (2024)","subjectSlug":"machine-learning","subjectName":"Machine Learning","universitySlug":"rtu","branchSlug":"computer-science","year":3,"semester":6,"examYear":2024,"paperCode":"6E7102","questions":[{"id":"v2-machine-learning-2024-pA-q1","number":1,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is Reinforcement Learning?"}],"answer":{"summary":"Reinforcement Learning is a behavioral mathematical framework where an Agent learns to make optimal sequential decisions in an Environment purely by maximizing a numerical Reward signal.","explanation":[{"type":"paragraph","content":"Reinforcement Learning (RL) is structurally distinct from supervised learning. It mathematically models an Agent interacting with a stateful Environment using Markov Decision Processes (MDPs). The agent executes an Action, observes the new State, and receives a positive or negative numerical Reward, learning a Policy (via algorithms like Q-Learning) to maximize total cumulative reward."}]}},{"id":"v2-machine-learning-2024-pA-q2","number":2,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Define precision and recall."}],"answer":{"summary":"Precision measures the mathematical accuracy of positive predictions (True Positives / Predicted Positives), while Recall measures the ability to detect all actual positive instances (True Positives / Actual Positives).","explanation":[{"type":"paragraph","content":"In binary classification, Precision ($TP / (TP + FP)$) violently penalizes False Positives (e.g., falsely diagnosing a healthy patient with cancer). Recall ($TP / (TP + FN)$) aggressively penalizes False Negatives (e.g., missing an actual cancer diagnosis). Engineers use the F1-Score to mathematically balance these conflicting metrics."}]}},{"id":"v2-machine-learning-2024-pA-q3","number":3,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is the role of a learning rate in gradient descent?"}],"answer":{"summary":"The learning rate is a critical mathematical hyperparameter in Gradient Descent that strictly dictates the absolute magnitude of the step taken toward the global minimum of the loss function.","explanation":[{"type":"paragraph","content":"The learning rate ($\\alpha$) multiplies the gradient vector: $\\theta_{new} = \\theta_{old} - \\alpha \\nabla J(\\theta)$. If $\\alpha$ is too massive, the algorithm violently overshoots the minimum and mathematically diverges to infinity. If $\\alpha$ is too microscopic, the algorithm converges so catastrophically slowly that it stalls in local minima."}]}},{"id":"v2-machine-learning-2024-pA-q4","number":4,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Differentiate between L1 and L2 regularization."}],"answer":{"summary":"L1 regularization mathematically forces absolute feature selection by driving weights exactly to zero, while L2 regularization smoothly penalizes massive weights without completely eradicating them.","explanation":[{"type":"paragraph","content":"L1 (Lasso) regularization adds the absolute sum of weights ($\\sum |w|$) to the cost function, creating a mathematically sharp diamond constraint that violently forces non-critical feature weights to exactly $0.0$. L2 (Ridge) adds the squared sum of weights ($\\sum w^2$), creating a smooth circular constraint that shrinks weights toward zero but rarely exactly to zero."}]}},{"id":"v2-machine-learning-2024-pA-q5","number":5,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is one-hot encoding?"}],"answer":{"summary":"One-hot encoding is an absolute architectural requirement for mathematically translating categorical string variables into orthogonal binary vectors that a machine learning algorithm can process.","explanation":[{"type":"paragraph","content":"Algorithms cannot multiply the word \"Red\". One-Hot Encoding violently expands a categorical column with $N$ unique labels into $N$ separate binary columns. \"Red\" becomes `[1, 0, 0]`, \"Green\" becomes `[0, 1, 0]`. This prevents the algorithm from mathematically assuming false ordinal relationships (e.g., assuming Green > Red)."}]}},{"id":"v2-machine-learning-2024-pA-q6","number":6,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Define a perceptron."}],"answer":{"summary":"A perceptron is the absolute fundamental mathematical unit of a neural network; it calculates the weighted sum of inputs, adds a bias, and applies a step activation function.","explanation":[{"type":"paragraph","content":"Invented by Frank Rosenblatt, the Perceptron is a primitive binary linear classifier. It mathematically calculates the dot product of the input vector $x$ and weight vector $w$, adds a scalar bias $b$, and feeds the result into a Heaviside Step function. It can mathematically solve linearly separable problems like AND/OR, but catastrophically fails at XOR."}]}},{"id":"v2-machine-learning-2024-pA-q7","number":7,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is the Silhouette score used for?"}],"answer":{"summary":"The Silhouette Score is a strict mathematical metric used exclusively in unsupervised clustering to evaluate how dense and well-separated the generated clusters are.","explanation":[{"type":"paragraph","content":"The Silhouette Score ranges from $-1$ to $+1$. It mathematically compares the mean intra-cluster distance (how tightly packed the points in Cluster A are) against the mean nearest-cluster distance (how far Cluster A is from Cluster B). A score near $+1$ mathematically proves absolute optimal clustering."}]}},{"id":"v2-machine-learning-2024-pA-q8","number":8,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"Name two common activation functions."}],"answer":{"summary":"Two absolute dominant activation functions are the Sigmoid function (used for binary probability) and the Rectified Linear Unit (ReLU) function (used aggressively in deep hidden layers).","explanation":[{"type":"paragraph","content":"The Sigmoid function, $f(x) = 1 / (1 + e^{-x})$, violently squashes any real number strictly into the range $(0, 1)$, perfect for outputting probabilities. ReLU, $f(x) = \\max(0, x)$, mathematically solves the catastrophic vanishing gradient problem by passing positive values unchanged and flatlining negative values."}]}},{"id":"v2-machine-learning-2024-pA-q9","number":9,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is the naive assumption in Naive Bayes?"}],"answer":{"summary":"The naive assumption is the strict, mathematically flawed presumption that all input features are absolutely conditionally independent of one another given the target class.","explanation":[{"type":"paragraph","content":"In Naive Bayes, calculating the exact joint probability $P(x_1, x_2, ..., x_n | y)$ is mathematically impossible for large datasets. The \"naive\" assumption violently simplifies this by assuming independence, calculating it as $P(x_1|y) \\times P(x_2|y) \\times ... P(x_n|y)$. Despite this horrific logical flaw, the algorithm performs exceptionally well in text classification."}]}},{"id":"v2-machine-learning-2024-pA-q10","number":10,"marks":2,"tags":["Part A"],"text":[{"type":"paragraph","content":"What is the curse of dimensionality?"}],"answer":{"summary":"The curse of dimensionality is a catastrophic mathematical phenomenon where adding more features exponentially increases the volume of the data space, rendering distance metrics completely meaningless.","explanation":[{"type":"paragraph","content":"As the number of dimensions (features) increases past 50, the mathematical distance between ANY two random points becomes nearly identical. Algorithms relying on Euclidean distance (like K-Nearest Neighbors or K-Means) completely break down because the concept of \"closeness\" mathematically ceases to exist in hyper-dimensional space."}]}},{"id":"v2-machine-learning-2024-pB-q11","number":11,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Explain Ridge and Lasso regularization techniques."}],"answer":{"summary":"A definitive mathematical comparison of Ridge (L2) and Lasso (L1) regularization techniques. Details how adding penalty terms to the loss function aggressively prevents catastrophic model overfitting and manipulates feature weights.","explanation":[{"type":"heading","content":"Ridge (L2) and Lasso (L1) Regularization Architectures"},{"type":"paragraph","content":"In Machine Learning, if a Linear Regression model is fed a dataset with 500 features and only 100 rows, it will mathematically memorize the noise (Overfitting). The model's weight vector $\\theta$ will explode to massive values (e.g., $w_1 = 45000$, $w_2 = -44900$) just to perfectly fit the training points. Regularization is the absolute architectural solution to this catastrophe; it mathematically penalizes massive weights by modifying the Cost Function $J(\\theta)$."},{"type":"subheading","content":"1. Ridge Regularization (L2 Penalty)"},{"type":"paragraph","content":"Ridge aggressively adds the squared magnitude of the weights to the loss function."},{"type":"list","items":["**Mathematical Equation:** $J(\\theta) = \\text{MSE} + \\lambda \\sum_{i=1}^{n} \\theta_i^2$","**Mechanism:** The hyperparameter $\\lambda$ dictates the violence of the penalty. Because the penalty is squared, a weight of $10$ incurs a massive penalty of $100$. The gradient descent algorithm is mathematically forced to shrink all weights smoothly toward zero to minimize this penalty.","**Result:** All features remain in the model, but their impact is suppressed. It is exceptionally good at handling highly correlated features (multicollinearity)."]},{"type":"subheading","content":"2. Lasso Regularization (L1 Penalty)"},{"type":"paragraph","content":"Least Absolute Shrinkage and Selection Operator (Lasso) adds the absolute magnitude of the weights."},{"type":"list","items":["**Mathematical Equation:** $J(\\theta) = \\text{MSE} + \\lambda \\sum_{i=1}^{n} |\\theta_i|$","**Mechanism:** The geometry of the L1 absolute value penalty creates a sharp mathematical \"diamond\" constraint space. When the loss function contour intersects this diamond, it almost always hits exactly on a corner.","**Result:** Absolute Feature Selection. Lasso violently forces the weights of irrelevant features to be EXACTLY $0.0$. If you input 500 features, Lasso might output a model using only 12 features, creating a highly sparse, interpretable mathematical architecture."]}]}},{"id":"v2-machine-learning-2024-pB-q12","number":12,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Describe the Naive Bayes classifier theorem and its applications."}],"answer":{"summary":"An advanced theoretical breakdown of the Naive Bayes Classifier. Executes the mathematics of Bayes' Theorem, details the horrific yet highly effective conditional independence assumption, and highlights its absolute dominance in NLP applications.","explanation":[{"type":"heading","content":"The Architecture of the Naive Bayes Classifier"},{"type":"paragraph","content":"Naive Bayes is a highly aggressive, probabilistic classification architecture strictly founded upon Bayes' Theorem. Unlike Logistic Regression, which attempts to find a mathematical boundary between classes, Naive Bayes is a Generative Model. It mathematically calculates the absolute probability of a data point belonging to a specific class, and violently assigns the point to the class with the highest probability (Maximum A Posteriori - MAP)."},{"type":"subheading","content":"The Mathematical Theorem"},{"type":"paragraph","content":"Bayes' Theorem is mathematically defined as: \n$$ P(Y|X) = \\frac{P(X|Y) \\cdot P(Y)}{P(X)} $$"},{"type":"list","items":["$P(Y|X)$: The Posterior Probability (The probability that an email is Spam $Y$, given that it contains the word \"Viagra\" $X$).","$P(X|Y)$: The Likelihood (The probability of seeing the word \"Viagra\" given that the email is already known to be Spam).","$P(Y)$: The Prior Probability (The absolute overall percentage of emails that are Spam in the dataset)."]},{"type":"subheading","content":"The \"Naive\" Assumption"},{"type":"paragraph","content":"If the input $X$ contains 100 features (words), calculating the exact likelihood $P(x_1, x_2 ... x_{100} | Y)$ is mathematically impossible due to exponential combinations. The algorithm makes a \"Naive\" (and physically incorrect) assumption: it violently assumes every single feature is completely mathematically independent of every other feature."},{"type":"paragraph","content":"$$ P(X|Y) \\approx P(x_1|Y) \\times P(x_2|Y) \\times ... \\times P(x_n|Y) $$"},{"type":"subheading","content":"Applications"},{"type":"paragraph","content":"Despite this horrific logical flaw (words in a sentence are obviously highly dependent), the mathematical multiplication works phenomenally well for classification. It is the absolute industry standard for Natural Language Processing (NLP), specifically Spam Filtering, Sentiment Analysis, and Document Categorization, because it executes blisteringly fast on massive datasets and requires very little training data."}]}},{"id":"v2-machine-learning-2024-pB-q13","number":13,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Explain Agglomerative Hierarchical Clustering with an example."}],"answer":{"summary":"A detailed algorithmic trace of Agglomerative Hierarchical Clustering. Explains the \"bottom-up\" mathematical fusion of data points using distance linkages to generate a massive tree-like Dendrogram architecture.","explanation":[{"type":"heading","content":"Agglomerative Hierarchical Clustering Architecture"},{"type":"paragraph","content":"Unlike K-Means, which forces the engineer to blindly guess the number of clusters ($K$) upfront, Hierarchical Clustering mathematically builds a massive, multi-level hierarchy of clusters. \"Agglomerative\" is the strictly bottom-up approach to building this architecture."},{"type":"subheading","content":"The Mathematical Algorithm"},{"type":"list","items":["**Step 1 (Initialization):** Treat every single data point in the dataset as its own absolute independent cluster. If there are $N=500$ points, there are strictly 500 clusters.","**Step 2 (Proximity Calculation):** Mathematically calculate the Distance Matrix between all existing clusters. This is typically done using Euclidean distance.","**Step 3 (The Violent Fusion):** Identify the two clusters that are mathematically the absolute closest to each other. Aggressively fuse them together into a single new cluster. The total number of clusters is now $N-1$.","**Step 4 (Linkage Update):** Mathematically recalculate the distances between the new fused cluster and all remaining clusters using a Linkage Criteria (e.g., Single Linkage = distance between the closest points; Complete Linkage = distance between the furthest points).","**Step 5:** Repeat Steps 3 and 4 endlessly until every single data point is fused into ONE massive, singular root cluster."]},{"type":"subheading","content":"The Dendrogram"},{"type":"paragraph","content":"The algorithm does not output a simple label. It outputs a Dendrogram—a massive mathematical tree graph. The vertical y-axis represents the absolute mathematical distance at which two clusters were fused. The engineer visually inspects the Dendrogram and draws a horizontal line across the longest vertical branches to violently \"slice\" the tree into the absolute mathematically optimal number of clusters."}]}},{"id":"v2-machine-learning-2024-pB-q14","number":14,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Discuss various methods to deal with highly imbalanced datasets."}],"answer":{"summary":"A critical engineering review of handling Imbalanced Datasets. Violently analyzes the failure of accuracy metrics, and details the mathematical execution of SMOTE (Oversampling) and strict Algorithmic Penalization (Class Weights).","explanation":[{"type":"heading","content":"Architectural Strategies for Imbalanced Datasets"},{"type":"paragraph","content":"An Imbalanced Dataset is a catastrophic machine learning crisis. If a dataset contains 99,000 legitimate credit card transactions and only 1,000 fraudulent ones, a \"dumb\" model that violently predicts `Legitimate` 100% of the time will achieve 99% Mathematical Accuracy, but will fail spectacularly in the real world. Standard algorithms are mathematically biased toward the majority class."},{"type":"subheading","content":"1. Metric Alteration (Abandoning Accuracy)"},{"type":"paragraph","content":"The engineer must instantly abandon \"Accuracy.\" The architecture must be evaluated using the **F1-Score**, **Precision-Recall AUC**, or the **Confusion Matrix**. These metrics violently expose the model's catastrophic failure to detect the minority class (Fraud)."},{"type":"subheading","content":"2. Data-Level Resampling Architectures"},{"type":"list","items":["**Random Undersampling:** Violently delete 98,000 legitimate transactions at random until both classes have exactly 1,000 samples. *Demerit:* Massive loss of critical mathematical data.","**Random Oversampling:** Blindly duplicate the 1,000 fraud samples 99 times. *Demerit:* Causes horrific model Overfitting.","**SMOTE (Synthetic Minority Over-sampling Technique):** The absolute optimal mathematical solution. SMOTE does not duplicate data; it uses K-Nearest Neighbors to analyze the fraud points and mathematically generates brand new, synthetic, highly realistic fraud data points to balance the dataset."]},{"type":"subheading","content":"3. Algorithm-Level Modifications"},{"type":"list","items":["**Class Weights (Cost-Sensitive Learning):** The algorithm's Cost Function is mathematically violently altered. A false prediction on a legitimate transaction incurs a penalty of `1`. However, a false prediction on a Fraud transaction incurs a massive penalty of `100`. The Gradient Descent algorithm is mathematically terrified of missing a fraud case, forcing it to focus on the minority class."]}]}},{"id":"v2-machine-learning-2024-pB-q15","number":15,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"What is a ROC curve and how is the AUC calculated?"}],"answer":{"summary":"A rigorous exposition on the ROC Curve and AUC metric. Details how graphing True Positive Rate against False Positive Rate mathematically evaluates binary classifiers independently of threshold selection.","explanation":[{"type":"heading","content":"The ROC Curve and AUC Architecture"},{"type":"paragraph","content":"When a Logistic Regression model predicts cancer, it does not output a binary `Yes/No`. It outputs a continuous mathematical probability from `0.0` to `1.0` (e.g., `0.85`). The engineer must artificially select a Threshold (e.g., `0.5`). If Probability > 0.5, predict `Yes`. Changing this threshold violently alters the Precision and Recall. The **Receiver Operating Characteristic (ROC) Curve** is a graphical architecture designed to evaluate the absolute quality of the model across EVERY SINGLE mathematically possible threshold simultaneously."},{"type":"subheading","content":"Constructing the ROC Curve"},{"type":"list","items":["**Y-Axis:** True Positive Rate (Sensitivity / Recall) $\\Rightarrow \\frac{TP}{TP + FN}$.","**X-Axis:** False Positive Rate (1 - Specificity) $\\Rightarrow \\frac{FP}{FP + TN}$.","**Execution:** The algorithm mathematically iterates the threshold from 0.0 to 1.0. At each threshold, it calculates the TPR and FPR and plots a single point on the 2D graph. A perfect model shoots violently straight up the Y-axis to `(0, 1)`. A useless model (random guessing) plots a straight diagonal 45-degree line from `(0, 0)` to `(1, 1)`."]},{"type":"subheading","content":"Area Under the Curve (AUC)"},{"type":"paragraph","content":"Looking at a graph is not mathematically precise. The **AUC (Area Under the Curve)** is the exact mathematical integral (the physical 2D area) underneath the ROC curve."},{"type":"list","items":["An AUC of exactly `1.0` indicates absolute mathematical perfection (100% TPR with 0% FPR).","An AUC of `0.5` mathematically proves the model is garbage (no better than flipping a coin).","The AUC represents the strict probability that the model will score a randomly chosen positive instance mathematically higher than a randomly chosen negative instance."]}]}},{"id":"v2-machine-learning-2024-pB-q16","number":16,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Explain the architecture of a Multilayer Perceptron (MLP)."}],"answer":{"summary":"An architectural breakdown of the Multilayer Perceptron (MLP). Violently details the strict mathematical feedforward propagation of inputs through dense hidden layers utilizing non-linear activation functions to solve complex non-linear problems.","explanation":[{"type":"heading","content":"The Architecture of a Multilayer Perceptron (MLP)"},{"type":"paragraph","content":"A single Perceptron is a strictly linear classifier; it mathematically fails instantly when presented with a non-linear problem like the XOR logic gate. The Multilayer Perceptron (MLP) is the absolute foundational architecture of Deep Learning (a Feedforward Artificial Neural Network). It aggressively stacks thousands of perceptrons into highly structured mathematical layers."},{"type":"subheading","content":"The Architectural Layers"},{"type":"list","items":["**1. The Input Layer:** A passive layer. It contains exactly one node for every single feature in the dataset (e.g., 784 nodes for a $28 \\times 28$ pixel image). It performs absolutely zero mathematical computation.","**2. The Hidden Layers:** The absolute computational core of the network. An MLP must have at least one hidden layer to be considered \"Deep.\" It is a \"Dense\" (Fully Connected) architecture, meaning every single node in Layer 1 is mathematically wired via a unique Weight ($W$) to every single node in Layer 2.","**3. The Output Layer:** Represents the final prediction. For binary classification, it is 1 node utilizing a Sigmoid activation. For 10-class classification, it is 10 nodes utilizing a Softmax activation."]},{"type":"subheading","content":"The Mathematical Execution (Forward Pass)"},{"type":"paragraph","content":"For a single node $j$ in a hidden layer, it receives the outputs from all $N$ nodes of the previous layer ($x_i$). It computes a massive linear algebra dot product: \n$$ Z_j = \\sum_{i=1}^{N} (w_{ij} \\cdot x_i) + b_j $$"},{"type":"paragraph","content":"Crucially, $Z_j$ is then violently forced through a strict **Non-Linear Activation Function** (like ReLU or Tanh). If the activation function was removed, the entire massive 100-layer network would mathematically collapse back into a single, useless linear regression equation."}]}},{"id":"v2-machine-learning-2024-pB-q17","number":17,"marks":4,"tags":["Part B"],"text":[{"type":"paragraph","content":"Discuss the difference between Standardization and Normalization for feature scaling."}],"answer":{"summary":"A definitive mathematical comparison of Feature Scaling architectures. Contrasts the strict $0$ to $1$ bounding of Normalization against the Z-score variance alteration of Standardization required by gradient descent algorithms.","explanation":[{"type":"heading","content":"Feature Scaling: Standardization vs. Normalization"},{"type":"paragraph","content":"Machine Learning algorithms (especially Gradient Descent, SVMs, and K-Means) are mathematically blind to real-world units. If Feature A is \"Salary\" (ranging from 50,000 to 200,000) and Feature B is \"Age\" (ranging from 18 to 80), the massive numbers in Feature A will violently dominate the Euclidean distance calculations, completely eradicating the impact of Feature B. Feature Scaling is the mandatory architectural process of mathematically forcing all features into a similar numerical scale."},{"type":"subheading","content":"1. Normalization (Min-Max Scaling)"},{"type":"list","items":["**Mathematical Equation:** $$ X_{norm} = \\frac{X - X_{min}}{X_{max} - X_{min}} $$","**Mechanism:** It mathematically crushes the entire dataset strictly into the exact boundary of `[0.0, 1.0]`.","**Demerit:** It is catastrophically vulnerable to Outliers. If the maximum Salary is usually $100,000$, but one outlier is $10,000,000$, the $X_{max}$ becomes 10M. The entire normal dataset gets violently squashed into a tiny microscopic range between `0.0` and `0.01`."]},{"type":"subheading","content":"2. Standardization (Z-Score Scaling)"},{"type":"list","items":["**Mathematical Equation:** $$ X_{std} = \\frac{X - \\mu}{\\sigma} $$","**Mechanism:** It subtracts the dataset Mean ($\\mu$) and divides by the Standard Deviation ($\\sigma$).","**Result:** The new data is mathematically centered exactly at a Mean of $0.0$, with a Standard Deviation of $1.0$. The values usually range from roughly $-3.0$ to $+3.0$.","**Merit:** It is vastly superior to Normalization because it does NOT bind the data to a strict limit, making it highly robust against massive outliers. It is the absolute mandatory requirement for Deep Learning and Gradient Descent."]}]}},{"id":"v2-machine-learning-2024-pC-q18","number":18,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Elaborate on the architecture and working of Convolutional Neural Networks (CNNs) for image classification."}],"answer":{"summary":"A massive architectural exposition on Convolutional Neural Networks (CNNs). Violently details the mathematical execution of Convolutional layers, feature maps, ReLU activation, Max Pooling downsampling, and the final Dense classification layers.","explanation":[{"type":"heading","content":"Architecture of Convolutional Neural Networks (CNNs)"},{"type":"paragraph","content":"A standard Multilayer Perceptron (MLP) fails catastrophically at Image Classification. If you input a $1000 \\times 1000$ pixel RGB image into an MLP, the first hidden layer requires 3,000,000 weights PER NODE. The network will instantly run out of RAM and violently overfit. Furthermore, MLPs completely destroy spatial structure (flattening the 2D image into a 1D line). **Convolutional Neural Networks (CNNs)** are a highly specialized architecture mathematically engineered to preserve 2D spatial relationships and drastically slash the number of parameters using weight sharing."},{"type":"heading","content":"The Mathematical Pipeline of a CNN"},{"type":"paragraph","content":"A CNN executes a rigid architectural pipeline: Convolution $\\rightarrow$ Activation (ReLU) $\\rightarrow$ Pooling $\\rightarrow$ Fully Connected."},{"type":"subheading","content":"1. The Convolutional Layer (Feature Extraction)"},{"type":"paragraph","content":"This is the absolute core mathematical engine. Instead of dense weights, the CNN utilizes small $3 \\times 3$ or $5 \\times 5$ mathematical matrices called **Filters (Kernels)**."},{"type":"list","items":["**The Sliding Window:** The $3 \\times 3$ Filter physically slides (convolves) across the massive image matrix, moving by a set `Stride` (e.g., 1 pixel at a time).","**The Dot Product:** At every stop, it mathematically calculates the dot product between the 9 filter weights and the 9 underlying image pixels.","**The Feature Map:** The output of this sliding dot product is a brand new 2D matrix called a Feature Map. If the network uses 64 filters, it generates 64 distinct Feature Maps.","**Architectural Genius (Weight Sharing):** A single $3 \\times 3$ filter only has 9 mathematical parameters, yet it processes the entire $1000 \\times 1000$ image. The CNN looks for the exact same feature (e.g., a vertical edge) everywhere in the image using the same 9 weights, providing massive translation invariance."]},{"type":"subheading","content":"2. The Activation Layer (ReLU)"},{"type":"paragraph","content":"Every single number in the resulting Feature Map is violently forced through a Rectified Linear Unit (ReLU) function: $f(x) = \\max(0, x)$. It mathematically annihilates all negative numbers to 0, injecting critical non-linearity into the network without triggering the vanishing gradient catastrophe."},{"type":"subheading","content":"3. The Pooling Layer (Downsampling)"},{"type":"paragraph","content":"After convolution, the 64 Feature Maps contain too much redundant data. The network executes **Max Pooling**."},{"type":"list","items":["A $2 \\times 2$ window with a Stride of 2 slides across the Feature Map.","It looks at 4 pixels and violently extracts ONLY the absolute maximum value, discarding the other 3.","This mathematically shrinks the spatial dimensions of the image by exactly 75%, slashing computational load and aggressively controlling overfitting, while preserving the dominant features."]},{"type":"subheading","content":"4. The Fully Connected (Dense) Layer (Classification)"},{"type":"paragraph","content":"After passing through multiple blocks of (Conv $\\rightarrow$ ReLU $\\rightarrow$ Pool), the image has been mathematically reduced from a $1000 \\times 1000$ pixel array into a deep, tiny $7 \\times 7 \\times 512$ block of highly abstract semantic features (e.g., \"eyes\", \"wheels\")."},{"type":"paragraph","content":"This final 3D block is violently **Flattened** into a single 1D vector and fed into a standard MLP (Dense Layer). The final output layer utilizes a **Softmax** mathematical function to output a strict probability distribution (e.g., 98% Cat, 2% Dog)."},{"type":"svg","content":""}]}},{"id":"v2-machine-learning-2024-pC-q19","number":19,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Explain K-Means vs Hierarchical clustering. Give a detailed mathematical explanation of both approaches."}],"answer":{"summary":"An exhaustive theoretical comparison of K-Means and Hierarchical Clustering. Violently breaks down the Expectation-Maximization mathematics of K-Means centroids against the linkage-based distance matrices of Agglomerative trees, highlighting their absolute computational complexity differences.","explanation":[{"type":"heading","content":"K-Means vs. Hierarchical Clustering Architectures"},{"type":"paragraph","content":"Unsupervised Machine Learning relies on clustering algorithms to mathematically discover hidden structures in unlabeled datasets. The two absolute dominant architectural paradigms are K-Means (a partition-based algorithm) and Hierarchical Clustering (a tree-based algorithm). They execute fundamentally different mathematical mechanics."},{"type":"heading","content":"1. K-Means Clustering Architecture"},{"type":"paragraph","content":"K-Means is a highly aggressive, iterative algorithm that utilizes the **Expectation-Maximization (EM)** mathematical framework to violently partition the dataset into exactly $K$ distinct, non-overlapping clusters."},{"type":"subheading","content":"The Mathematical Algorithm"},{"type":"list","items":["**Step 1 (Initialization):** The engineer MUST manually define the integer hyperparameter $K$. The algorithm mathematically drops $K$ random points into the data space to act as initial Centroids (the mathematical center of mass).","**Step 2 (Expectation / Assignment):** The algorithm calculates the absolute Euclidean distance between every single data point $x_i$ and every centroid $\\mu_j$: \n $$ D = || x_i - \\mu_j ||^2 $$ \n Each data point is violently assigned to the cluster of the centroid it is mathematically closest to.","**Step 3 (Maximization / Update):** The algorithm completely recalculates the position of all $K$ centroids. The new centroid $\\mu_j$ is moved to the exact mathematical Mean (average) of all the data points currently assigned to it: \n $$ \\mu_j = \\frac{1}{|S_j|} \\sum_{x \\in S_j} x $$","**Step 4 (Convergence):** Steps 2 and 3 repeat endlessly. The centroids physically move across the space until they mathematically stop shifting (Convergence). The loss function (Within-Cluster Sum of Squares) is minimized."]},{"type":"subheading","content":"Merits and Demerits"},{"type":"paragraph","content":"It executes blisteringly fast with a time complexity of $O(n \\cdot K \\cdot I)$. However, guessing the exact $K$ is often mathematically impossible upfront, and K-Means fails catastrophically if the physical clusters are not perfectly spherical."},{"type":"heading","content":"2. Hierarchical Clustering Architecture"},{"type":"paragraph","content":"Hierarchical Clustering completely abandons the concept of predefined $K$. It builds a massive mathematical taxonomy (a Dendrogram) representing the exact distance relationships of every point."},{"type":"subheading","content":"The Mathematical Algorithm (Agglomerative)"},{"type":"list","items":["**Step 1:** Initialize all $N$ data points as $N$ independent clusters.","**Step 2:** Compute the massive $N \\times N$ Distance Matrix. This defines the exact mathematical distance between every cluster.","**Step 3:** Find the two clusters $A$ and $B$ that have the absolute minimum mathematical distance. Violently fuse them into a single cluster $(A \\cup B)$.","**Step 4 (Linkage Update):** Recalculate the distance from the new fused cluster to all other remaining clusters. The mathematical \"Linkage\" defines how this is calculated:\n - *Single Linkage:* Distance is the absolute minimum distance between any point in $A$ and any point in $B$: $\\min(dist(x, y))$.\n - *Ward's Method:* Fuses clusters that mathematically minimize the total variance within the clusters.","**Step 5:** Repeat until only exactly ONE massive root cluster remains. The engineer then mathematically \"cuts\" the resulting Dendrogram tree at a specific height to obtain the desired number of clusters."]},{"type":"subheading","content":"Merits and Demerits"},{"type":"paragraph","content":"It is highly interpretable via the Dendrogram and requires zero guessing of $K$. However, its mathematical time complexity is a catastrophic $O(N^3)$, making it completely physically impossible to execute on massive datasets containing millions of rows."}]}},{"id":"v2-machine-learning-2024-pC-q20","number":20,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Discuss Logistic Regression cost function derivation and how gradient descent minimizes it."}],"answer":{"summary":"A rigorous mathematical derivation of the Logistic Regression Cost Function. Violently details the failure of Mean Squared Error, the necessity of Maximum Likelihood Estimation to derive Log-Loss, and the exact Gradient Descent calculus required to minimize it.","explanation":[{"type":"heading","content":"Logistic Regression and the Cost Function Derivation"},{"type":"paragraph","content":"Despite its name, Logistic Regression is an absolute foundational Binary Classification algorithm. A standard Linear Regression outputs a line $z = \\theta^T x$. For classification, this output must be violently squashed strictly into a probability range $[0, 1]$. This is achieved by passing $z$ through the Sigmoid (Logistic) activation function:"},{"type":"paragraph","content":"$$ \\hat{y} = h_\\theta(x) = \\frac{1}{1 + e^{-\\theta^T x}} $$"},{"type":"heading","content":"1. The Failure of Mean Squared Error (MSE)"},{"type":"paragraph","content":"If an engineer foolishly attempts to use the standard Linear Regression Cost Function (MSE) for Logistic Regression, the math fails catastrophically. Because the Sigmoid function is highly non-linear, inserting it into the MSE equation creates a horrific, non-convex cost function surface. It will contain thousands of \"Local Minima\". Gradient Descent will instantly get trapped, mathematically failing to ever find the true global minimum."},{"type":"heading","content":"2. Derivation of the Log-Loss (Cross-Entropy) Cost Function"},{"type":"paragraph","content":"To solve this, the architecture utilizes **Maximum Likelihood Estimation (MLE)**."},{"type":"paragraph","content":"The probability of the target $y$ (where $y \\in \\{0, 1\\}$) given the input $x$ can be mathematically written as a single Bernoulli distribution equation:"},{"type":"paragraph","content":"$$ P(y | x; \\theta) = (\\hat{y})^y \\cdot (1 - \\hat{y})^{(1-y)} $$"},{"type":"list","items":["If the true label $y=1$, the right half mathematically vanishes ($something^0 = 1$), leaving only the probability $P = \\hat{y}$.","If the true label $y=0$, the left half vanishes, leaving $P = (1 - \\hat{y})$."]},{"type":"paragraph","content":"To find the optimal weights $\\theta$, we must maximize the total likelihood of the entire dataset. Because multiplying thousands of probabilities causes a catastrophic floating-point underflow in the CPU, we apply a mathematical logarithm to convert the multiplication into addition (Log-Likelihood). We then multiply by $-1$ to convert the maximization problem into a standard Minimization problem (Gradient Descent requires a loss to minimize)."},{"type":"paragraph","content":"This strictly derives the absolute **Binary Cross-Entropy (Log-Loss) Cost Function** for a single data point:"},{"type":"paragraph","content":"$$ J(\\theta) = - [y \\log(\\hat{y}) + (1-y) \\log(1-\\hat{y})] $$"},{"type":"paragraph","content":"This newly derived function is mathematically guaranteed to be perfectly **Convex** (a single smooth bowl shape), completely eradicating the Local Minima crisis."},{"type":"heading","content":"3. Minimization via Gradient Descent"},{"type":"paragraph","content":"To minimize the cost, Gradient Descent mathematically calculates the partial derivative (gradient) of the massive cost function with respect to every single weight $\\theta_j$. Through the brutal application of the Chain Rule of calculus across the Log and Sigmoid functions, the derivative beautifully collapses into an incredibly simple mathematical equation:"},{"type":"paragraph","content":"$$ \\frac{\\partial J(\\theta)}{\\partial \\theta_j} = \\frac{1}{m} \\sum_{i=1}^{m} (\\hat{y}^{(i)} - y^{(i)}) x_j^{(i)} $$"},{"type":"paragraph","content":"The algorithm then violently updates the weights using the learning rate $\\alpha$:"},{"type":"paragraph","content":"$$ \\theta_j := \\theta_j - \\alpha \\frac{\\partial J(\\theta)}{\\partial \\theta_j} $$"},{"type":"paragraph","content":"This loop executes relentlessly until the gradient mathematically reaches zero (Convergence at the absolute global minimum)."}]}},{"id":"v2-machine-learning-2024-pC-q21","number":21,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Provide a detailed explanation of Hidden Markov Models (HMM) and their applications in sequential data."}],"answer":{"summary":"A massive architectural breakdown of Hidden Markov Models (HMM). Explains the strict mathematical separation of hidden state transitions and observable emissions, detailing the execution of the Viterbi Algorithm to deduce the most probable sequence of hidden realities.","explanation":[{"type":"heading","content":"Architecture of Hidden Markov Models (HMM)"},{"type":"paragraph","content":"Standard classification algorithms completely assume that every data point is independent. They fail catastrophically on Sequential or Time-Series data (like DNA sequences, Speech Recognition, or Stock Markets) where Data Point 2 is mathematically highly dependent on Data Point 1. The **Hidden Markov Model (HMM)** is an absolute foundational probabilistic architecture engineered specifically to model sequence dependencies where the true \"State\" of the system is mathematically invisible to the observer."},{"type":"heading","content":"The Dual-Matrix Mathematical Architecture"},{"type":"paragraph","content":"An HMM mathematically operates in two completely separate spaces simultaneously:"},{"type":"subheading","content":"1. The Hidden States (The Markov Process)"},{"type":"list","items":["The system possesses a set of unobservable (Hidden) states (e.g., the actual Weather: `Sunny` or `Rainy`).","It strictly obeys the **Markov Assumption**: The mathematical probability of transitioning to State $t$ depends ABSOLUTELY ONLY on State $t-1$. The entire rest of history is violently mathematically ignored.","This is governed by the **Transition Matrix ($A$)**: A 2D matrix detailing the exact probability of jumping from any state to any other state (e.g., $P(\\text{Sunny tomorrow} | \\text{Rainy today}) = 0.3$)."]},{"type":"subheading","content":"2. The Observable Emissions"},{"type":"list","items":["Because the true state is hidden, the engineer can only see \"Emissions\" or \"Observations\" (e.g., you are locked in a windowless room, you can only observe if your coworker brings an `Umbrella` or `No Umbrella`).","This is governed by the **Emission Matrix ($B$)**: A 2D matrix detailing the exact mathematical probability of generating a specific Observation given a specific Hidden State (e.g., $P(\\text{Umbrella} | \\text{Rainy}) = 0.9$)."]},{"type":"heading","content":"The Three Fundamental Mathematical Problems of HMMs"},{"type":"paragraph","content":"To execute an HMM, scientists formulated three massive dynamic programming algorithms:"},{"type":"list","items":["**1. The Evaluation Problem (The Forward Algorithm):** Given an HMM ($A$ and $B$ matrices) and a sequence of observations (e.g., `Umbrella, Umbrella, No Umbrella`), what is the absolute mathematical probability that this specific HMM generated this sequence?","**2. The Decoding Problem (The Viterbi Algorithm):** The absolute most critical use case. Given a sequence of visible observations (`Umbrella, Umbrella, No Umbrella`), the Viterbi algorithm executes a highly complex dynamic programming trace to violently deduce the *absolute most probable sequence of Hidden States* that occurred (e.g., it deduces the weather was actually `Rainy, Rainy, Sunny`). It mathematically calculates the path of maximum probability through the state space.","**3. The Learning Problem (The Baum-Welch Algorithm):** Given only raw sequences of observations, this algorithm utilizes Expectation-Maximization to mathematically adjust and perfectly optimize the unknown Transition ($A$) and Emission ($B$) matrices."]},{"type":"subheading","content":"Real-World Applications"},{"type":"paragraph","content":"HMMs dominated **Speech Recognition** (Siri) for decades. The audio waveform is the visible Observation. The actual English words spoken by the human are the Hidden States. The Viterbi algorithm violently calculates the most probable sequence of English words that produced those specific audio waveforms."}]}},{"id":"v2-machine-learning-2024-pC-q22","number":22,"marks":10,"tags":["Part C"],"text":[{"type":"paragraph","content":"Explain the concept of Q-Learning in Reinforcement Learning with a detailed algorithmic breakdown."}],"answer":{"summary":"An exhaustive algorithmic exposition on Q-Learning in Reinforcement Learning. Violently details the strict mathematical update of the Q-Table using the Bellman Equation to balance immediate rewards against discounted future expectations.","explanation":[{"type":"heading","content":"Q-Learning Architecture in Reinforcement Learning"},{"type":"paragraph","content":"Q-Learning is the absolute foundational \"Model-Free, Off-Policy\" algorithm in Reinforcement Learning. The Agent is dropped into a completely unknown Environment. It does not possess a map (No Model). Its absolute goal is to mathematically deduce the ultimate \"Policy\" ($\\pi$)—the exact perfect Action to take in every possible State to violently maximize the total cumulative numerical Reward."},{"type":"heading","content":"The Mathematical Architecture: The Q-Table"},{"type":"paragraph","content":"The algorithm initializes a massive 2D matrix called the **Q-Table**. The rows represent every possible State ($S$), and the columns represent every possible Action ($A$). The cells contain the \"Q-Value\" $Q(S, A)$, which represents the absolute mathematical \"Quality\" (expected future reward) of taking Action $A$ while in State $S$. Initially, this entire table is filled with absolute zeros (complete ignorance)."},{"type":"heading","content":"The Algorithmic Execution"},{"type":"paragraph","content":"The agent executes thousands of \"Episodes\" (lifetimes) in the environment, looping through a strict mathematical process:"},{"type":"subheading","content":"1. Action Selection (Exploration vs. Exploitation)"},{"type":"paragraph","content":"The agent uses the **Epsilon-Greedy ($\\epsilon$) Strategy**. It generates a random number. If the number $< \\epsilon$, the agent violently executes a completely random action (Exploration) to discover new areas. If $> \\epsilon$, it looks at its Q-Table and strictly chooses the action with the absolute highest mathematical Q-Value for its current state (Exploitation)."},{"type":"subheading","content":"2. Interaction and Observation"},{"type":"paragraph","content":"The agent executes the chosen Action ($A$). The Environment mathematically responds by pushing the agent into a brand new State ($S'$) and giving it an immediate numerical Reward ($R$)."},{"type":"subheading","content":"3. The Bellman Equation (The Mathematical Update)"},{"type":"paragraph","content":"This is the absolute engine of Q-Learning. The agent must update the Q-Value in its table for the previous $(S, A)$ pair. It uses the Bellman Equation:"},{"type":"paragraph","content":"$$ Q^{new}(S, A) = Q^{old}(S, A) + \\alpha \\left[ R + \\gamma \\max_{a} Q(S', a) - Q^{old}(S, A) \\right] $$"},{"type":"list","items":["$\\alpha$ (Learning Rate): How violently new information overrides old information (0 to 1).","$R$: The immediate, physical reward just received.","$\\gamma$ (Discount Factor): Crucial parameter. It mathematically determines how much the agent cares about *future* rewards versus immediate gratification.","$\\max_{a} Q(S', a)$: The absolute genius of the algorithm. The agent looks at its *new* state $S'$, scans the Q-Table for all possible actions it *could* take next, and violently extracts the maximum possible future Q-Value. It mathematically propagates this future knowledge backward in time to update the previous state."]},{"type":"subheading","content":"4. Convergence"},{"type":"paragraph","content":"After millions of iterations, the Q-Table mathematically converges. The values stop changing. The agent now possesses perfect knowledge. To navigate the world perfectly, it simply drops exploration entirely, looks at the Q-Table, and blindly follows the mathematically highest number in every state."}]}}]}],"loading":false,"error":null}}}

Q15Machine Learning

Question

Answer