Study guides
Each lecture rewritten as a structured guide, with clean diagrams, worked examples and practice questions, cross-checked against the slides and the textbook.
- W02 Evaluating a modelTrain/validation/test, cross-validation, bootstrapping, confusion matrix, precision/recall, F1, PR & ROC curves, overfitting Open →
- W03 Linear modelsLeast-squares linear regression (closed form and gradient descent), feature scaling, linear decision boundaries, the perceptron, linear separability and XOR, the differentiable perceptron vs. logistic regression Open →
- W04 k nearest neighborsInstance-based learning, the k-NN vote, choosing k, distance metrics and feature scaling, k-NN regression, k-d trees and inverted lists, the curse of dimensionality Open →
- W05 Support vector machinesMaximum margin, point-to-hyperplane distance, hard-margin SVM as a quadratic program, support vectors, soft margin and C, hinge loss, the dual problem and the kernel trick Open →
- W06 Decision trees, bagging and random forestsDecision tree anatomy, greedy top-down tree growing (ID3/CART), entropy, Gini and misclassification impurity, information gain, the PlayTennis example, continuous thresholds and axis-aligned partitions, overfitting and pruning, high variance, ensembles and aggregation, bootstrapping and bagging, out-of-bag error, random forests Open →
- W07 Naive Bayes classifierBayes' rule, prior, likelihood and posterior, generative vs discriminative classifiers, MLE vs MAP, the naive Bayes conditional-independence assumption, training by counting, the play-tennis example, log-probabilities, zero counts and Laplace smoothing, Gaussian naive Bayes, spam filtering, missing values Open →
- W08 K-means clusteringUnsupervised learning and clustering, what makes a good clustering, Lloyd's k-means algorithm, the distortion objective and why it converges, initialization and local minima, k-means++, choosing k with the elbow and silhouette, limits of k-means, k-means vs k-NN Open →
- W09 Gaussian mixture models and EMk-means recap and distortion, hard vs soft assignment, mixture models and mixing coefficients, covariance and ellipse shapes, likelihood and negative log-likelihood, responsibilities, the EM algorithm, k-means as a limit of EM, choosing K with BIC and AIC, singularities, anomaly detection Open →
- W10 PCA, kernel PCA and ICAWhy reduce dimensions, PCA as maximum-variance projection, the eigenvector derivation, later components and deflation, maximum variance equals minimum reconstruction error, explained variance and choosing k, eigenfaces, eigendecomposition vs SVD, the power method, kernel PCA, ICA Open →
- W11 RegularizationOverfitting and its fixes, the regularized cost function and lambda, L2 (ridge) and weight decay, the regularized normal equation, L1 (lasso) sparsity and the diamond-vs-disk picture, elastic net, regularized logistic regression, choosing lambda, dropout, data augmentation, early stopping Open →
- W12 Semi-supervised learningLabeled plus unlabeled data, why unlabeled data needs an assumption, transductive SVM and low-density separation, the label-flipping heuristic, co-training with two views, agreement regularization, graph-based methods, min-cut and harmonic label propagation Open →
- W13 The multilayer perceptron and backpropagationThe multilayer perceptron, hidden layers and why they need a nonlinear activation, sigmoid/tanh/ReLU, solving XOR, softmax and cross-entropy for multiclass outputs, the chain rule and computation graphs, backpropagation with a full worked example, non-convex training and local minima Open →
- W14 Convolutional neural networksWhy images need CNNs, sparse connectivity and weight sharing, the convolution (cross-correlation) operation, feature maps and edge filters, padding and stride and the output-size formula, channels and parameter counts, pooling and translation invariance, receptive fields, backpropagation with shared weights, LeNet-5 and AlexNet, what CNN layers learn Open →
- W15 Recurrent neural networks and LSTMSequence data and variable-length inputs, the vanilla RNN cell, unrolling and weight sharing over time, input-output scenarios, backpropagation through time, vanishing and exploding gradients, gradient clipping, deep and bidirectional RNNs, LSTM gates and the cell state, peephole LSTM, GRU Open →
- W16 Reinforcement learningAgent-environment loop, policies and rewards, discounted return, Markov decision processes, Bellman optimality, value iteration, Q-learning, epsilon-greedy exploration, deep Q-networks (experience replay, target network, Double DQN), policy gradients and REINFORCE, AlphaGo Zero Open →