Machine Learning and Computer Vision 1) Machine Learning (ML: Field of Al where computers learn patterns from data and improve performance without explicit programming. Key Elements: Data - Input examples used for training. Model - Mathematical representation of relationships. Features - Input variables describing data. _Algorithm - Procedure used to learn patterns. Training - Process of learning model ___parameters. Prediction - Model output on new data.Evaluation - Measuring model accuracy, Basic ML Flow Diagram Data — Preprocessing — Model Training = Learned Model — Prediction 0) Types of Machine Learning 1. Supervised Learning Model trained with labeled data Ciaput + output. Tasks: Classification, Regression wi Examples: EXOT DC ____ Email spam detection ____House price prediction 2. Unsupervised Learning _____Uses unlabeled data, Finds hidden patterns. Tasks:Clustering Association Example: Customer segmentation 3. Semi-Supervised Learning Combination of small labeled + large unlabeled data. Used when labeling data is expensive. Example: Image classification with few labeled images. 4, Reinforcement Learning Agent learns by interaction with environment using rewards. Goal: maximize cumulative reward. Example: Game playing (AlphaGo)Robotics navigation © Statistical Description of Data Mean Average value of dataset Variance Measures spread of data from mean Standard Deviatio Square root of variance Shows dispersion of dataset \mportance of Data Visualization Helps understand patterns and trends Detects outliers and anomalies Improves feature selection Co on tools: Histogram Scatter plot Box plot Heatmap d) Mathematical Foundations for MLVectors Ordered list of numbers Represent data samples. Matrices Rectangular array of numbers. Used for storing datasets. Linear Algebra Concepts Matrix multiplication Dot product Eigenvalues Eigenvectors Used in: Neural networks PCA Linear regression €) Bayesian Learning and Naive Bayes Bayesian Learning Based on Bayes Theorem Updates probability when new evidence appears.Naive Bayes Classifier ___ Probabilistic classifier based on Bayes theorem. Assumes feature independence. Example: Spam Detection Features: Words like free, offer, win Model calculates probability: PSparm | Words) Email classified as Spam or Not Spam based on highest probability. 2) Parametric Learning a) Hypothesis Function for Linea Regression Used for predicting continuous values. E : | Exo pe Predicting house price from area $e i — CoOptimization algorithm to minimize cost function. Steps: Initialize parameters Calculate gradient Update parameters © Single vs Multiple Linear Regression Sinale Variable One input feature. Example: House price using \ze. bedroo locatioMethods: Min-Max Scaling Standardization Benefits: Faster convergence Prevents dominance of large-value features \mproves gradient descent performance Example: Height (cm) vs Salary @) e) Logistic Regression and Regularization Logistic Regression Used for classification problems. Output between 0 and 1. Example: Disease prediction (Yes/No Reqularization Prevents overfitting. Types:Ui (Lasso) — feature selection L2 (Ridge) > weight shrinkage Utility: \mproves model generalization Reduces complexit S) Non-Linear Learning a) Perceptron Model Basic binary linear classifier. Output: =signw-x+b) b) Stochastic Gradient Descent (SGD) Updates weights using one sample ata time. Advantages: Faster for large dataset: Requires le emory Helps escape local minima M ayer Perceptron (MLP) ArchitectureIaput Layer Hidden Layers, Output Layer Backpropagation Algorithin to train neural networks. Steps: Forward pass Calculate error Propagate error backward Update weights Used in: \mage recognition Speech processing d) Bias-Variance Dilemma Bias Error due to oversimplified model. Leads to_underfitting. Variance Error due to complex model sensitive To noise.e€) Support Vector Machines (SVM) and Nearest Neighbor Support Vector Machine Finds optimal hyperplane separating classes. $e Maximizes margin. Example: Face recognition. Nearest Neighbor (KANN) Classifies data based on closest neighbors. oteps: Choose K value Calculate distance Assign majority class Example: Recommender systems.4) Unsupervised Learning Learning from unlabeled data, 0) K-Means Clustering Steps: Choose number of clusters Initialize centroids Assign points to nearest centroid Update centroid Repeat until convergence Applications: Customer segmentation \mage compression © Hierarchical Clustering Builds cluster hierarchy. Types: Agglomerative (cottom-up) Divisive Gop-down) _Dendrogram Tree diagram showing cluster relationships. d) DBSCANi” Density-Based Spatial Clustering Key Concepts: Core points ____ Border points ___Noise points a Advantages: __| Advantages: Detects arbitrary shape clusters Handles noise/outliers ___No need to specify number of clusters iW e) Model-Based & Fuzzy Clustering Model-Based Clustering __Assumes data generated by probability __model. "Example: ee Gaussian Mixture Model. Fuzzy Clusterin. UZZY CiUsTering Data points belong to multiple clusters with probabilities. Example:Fuzzy C-Means algorithm. 5) Model Evaluation and Ensemble Learning a) Errors Training Error Error on training dataset. Test Error Error on unseen data. Generalization Error Model's ability to perform on new data. b) Cross Validation Technique for reliable model evaluation. K-Fold Cross Validatio Split data into K parts. Train on K-1 parts. Jest on remaining part. Repeat K times. Benefits: Reduces overfitting Better performance estimation©) Decision Trees and Random Forest Decision Tree Tree structure with decision nodes and leaves. Advantages: Easy to interpret Handles nonlinear data Random Forest Ensemble of many decision trees, Uses bagging + random feature selection. Benefits: Higher accurac Reduced overfitting d) Bagging vs Boosting __ Bagging Models trained independently Reduces variance Example: Randorm ForestBoostin Models trained sequentially Reduces bias Example: AdaBoost, Gradient Boost e) Dataset Challenges imbalanced Dataset _One class dominates. Solutions: SMOTE Oversampling Undersampling Outlier Detection Identifying abnormal data points. Methods: Score Isolation Forest Large Scale Data Handling D outed computing Mint-batch training Oud comp ing Spa K, hadoop