Documenting the Stanford CS229 ML Course and gathering insights

• Although I have an informal understanding of ML and deep learning, I decided to begin this in-depth course on a friend's recommendation and am documenting my progress as part of the "Build in Public" movement.

Lecture 1 - Introduction and Linear Algebra

Course lecture : Lecture - 1

• Slides : Slides
• Linear Algebra Reading : pdf

---

Background & Prerequisites — What You Need to Know Before Completing This Blog

CS229 is an intermediate/advanced course. Before blogging through each lecture, you need a strong foundation in the following areas. Each section explains the topic, why it matters for CS229, and the depth required.

---

1. Linear Algebra (Critical)

Why: ML algorithms are fundamentally linear algebra operations. Gradient descent, PCA, SVD, neural networks — all are matrix/vector operations.

• Vectors — Column vectors, row vectors, norms (L1, L2, L∞). Geometric interpretation: direction + magnitude.
• Matrices — Addition, multiplication (row-by-column), transpose, trace. Understand matrix multiplication as a composition of linear transformations.
• Systems of linear equations — $Ax = b$. Solving via Gaussian elimination, LU decomposition. When solutions exist (rank conditions).
• Eigenvalues & Eigenvectors — $Av = \lambda v$. Eigendecomposition, spectral theorem for symmetric matrices. Used in PCA, covariance analysis, stability analysis.
• Singular Value Decomposition (SVD) — $A = U\Sigma V^T$. Fundamental for dimensionality reduction, matrix approximation, and understanding matrix rank.
• Positive definite matrices — $x^T A x > 0$ for all nonzero $x$. Guarantees convexity of quadratic forms. Hessian positive definite → local minimum.
• Matrix calculus — Gradient of a scalar w.r.t. a vector ($\nabla_x f$), Jacobian (vector-to-vector), Hessian (second derivatives). Chain rule for matrix operations.

2. Probability & Statistics (Critical)

Why: ML is applied statistics. Every model makes probabilistic assumptions about data.

• Probability axioms — Sample space, events, probability measure. Conditional probability, Bayes' theorem: $P(A|B) = \frac\{P(B|A)P(A)\}\{P(B)\}$.
• Random variables — Discrete (PMF) and continuous (PDF). CDF. Expected value $E[X]$, variance $\text\{Var\}(X) = E[X^2] - (E[X])^2$.
• Common distributions — Bernoulli, Binomial, Poisson (discrete). Gaussian/Normal, Exponential, Uniform (continuous). Multivariate Gaussian $\mathcal\{N\}(\mu, \Sigma)$.
• Maximum Likelihood Estimation (MLE) — Given data, find parameters $\theta$ that maximize $P(\text\{data\}|\theta)$. Log-likelihood trick. Applies to linear regression, logistic regression, Gaussian models.
• Maximum A Posteriori (MAP) — MLE with a prior: maximize $P(\theta|\text\{data\}) \propto P(\text\{data\}|\theta)P(\theta)$. Equivalent to regularized MLE.
• Information theory basics — Entropy $H(X) = -\sum p(x)\log p(x)$, KL divergence, cross-entropy (used as loss functions).

3. Multivariable Calculus (Critical)

Why: Optimization (gradient descent) requires computing and understanding gradients.

• Partial derivatives — Rate of change of a function w.r.t. one variable, holding others constant.
• Gradient — Vector of all partial derivatives: $\nabla f = [\frac\{\partial f\}\{\partial x_1\}, \ldots, \frac\{\partial f\}\{\partial x_n\}]^T$. Points in the direction of steepest ascent.
• Chain rule — For composite functions: $\frac\{\partial\}\{\partial x\} f(g(x)) = f'(g(x)) \cdot g'(x)$. Extends to multivariate (Jacobian chain rule). Foundation of backpropagation.
• Hessian matrix — Matrix of second partial derivatives. Determines curvature. Positive definite Hessian → convex function → gradient descent converges.
• Taylor expansion — First-order: linear approximation (used in gradient descent). Second-order: quadratic approximation (used in Newton's method).

4. Optimization Theory (Important)

Why: Training any ML model is an optimization problem — minimizing a loss function.

• Convex functions — A function $f$ is convex if $f(\alpha x + (1-\alpha)y) \leq \alpha f(x) + (1-\alpha)f(y)$. Convex problems have a single global minimum. Linear regression loss is convex; neural network loss is not.
• Gradient descent — Update rule: $\theta := \theta - \alpha \nabla_\theta J(\theta)$. Learning rate $\alpha$ controls step size. Too large → diverge, too small → slow.
• Stochastic Gradient Descent (SGD) — Use a random subset (mini-batch) of data for each gradient update. Faster per iteration, noisier but often converges to good solutions.
• Newton's method — Second-order: $\theta := \theta - H^\{-1\}\nabla J$. Converges faster (quadratically) but expensive to compute/invert the Hessian.
• Lagrange multipliers — For constrained optimization. Used in deriving SVMs.

5. CS229 Course Structure — Lecture Roadmap

Why: To plan which lectures to blog and in what order.

• Supervised Learning — Linear regression, logistic regression, GDA (Generative Discriminative Analysis), Naive Bayes, SVMs, neural networks, decision trees, ensemble methods.
• Unsupervised Learning — K-means clustering, EM algorithm, PCA, ICA, factor analysis.
• Learning Theory — Bias/variance tradeoff, VC dimension, PAC learning, regularization, model selection.
• Reinforcement Learning — MDPs, value iteration, policy iteration, Q-learning, policy gradient.
• Practical Advice — Debugging ML algorithms, error analysis, ceiling analysis.

6. Python & NumPy (Practical)

Why: Implementing algorithms from the course requires efficient numerical computing.

• NumPy arrays — Vectorized operations, broadcasting, slicing. Avoid Python loops for numerical work.
• Matrix operations — np.dot(), np.linalg.inv(), np.linalg.eig(), np.linalg.svd().
• Matplotlib — Plotting decision boundaries, loss curves, data distributions.
• Jupyter notebooks — Interactive development for experimentation and documentation.

---

TODO / Remaining Work

• Complete Lecture 1 notes: Linear algebra review with worked examples
• Blog Lecture 2: Linear regression, normal equations, gradient descent derivation
• Blog Lecture 3: Locally weighted regression, logistic regression
• Blog Lecture 4: Newton's method, perceptron, exponential family
• Blog Lecture 5: Generative learning algorithms (GDA, Naive Bayes)
• Blog Lecture 6-7: Support Vector Machines (kernels, SMO algorithm)
• Blog Lecture 8-9: Neural networks and backpropagation
• Blog Lecture 10: Bias/variance, regularization, model selection
• Blog Lecture 11-12: Unsupervised learning (K-means, EM, PCA)
• Blog Lecture 13-14: Reinforcement learning (MDPs, Q-learning)
• Add Python implementations for each lecture's key algorithm

---

Lecture 1 — Summary & Key Takeaways

The Three Types of Machine Learning (Andrew Ng's Framing)

1. Supervised Learning — given labelled pairs $(x^\{(i)\}, y^\{(i)\})$, learn $h: X \to Y$.
   • Regression: $y \in \mathbb\{R\}$ (house prices, watch time).
   • Classification: $y \in \{0, 1, \ldots, K\}$ (spam/not-spam, digit recognition).
2. Unsupervised Learning — only $x^\{(i)\}$, no labels. Find structure: clusters, lower-dim representations, density.
3. Reinforcement Learning — an agent interacts with an environment, gets rewards, learns a policy $\pi(a|s)$ to maximise long-run reward.

Core Notation Used Throughout the Course

Linear Algebra Essentials for the Rest of the Course

Lecture 1's linear-algebra review exists because almost every algorithm in CS229 can be written as matrix/vector operations:

• Vector inner product: $u^T v = \sum_i u_i v_i$ — used for every prediction $h_\theta(x) = \theta^T x$.
• Matrix-vector product: $(Ax)_i = \sum_j A_\{ij\} x_j$ — how linear layers of a neural net work.
• Gradient: $\nabla_\theta J \in \mathbb\{R\}^n$, the vector of partial derivatives. Gradient descent steps in $-\nabla_\theta J$.
• Hessian: $H_\{ij\} = \partial^2 J / \partial \theta_i \partial \theta_j$. Newton's method uses $\theta \leftarrow \theta - H^\{-1\} \nabla J$.
• Positive semi-definiteness: $x^T A x \geq 0$ for all $x$. Guarantees convex $J$ when $J$ is quadratic with Hessian $A$.

My Plan for Blogging the Rest of the Course

I'm following the official lecture order (linked per-lecture above). For each lecture I commit to producing:

1. A one-page summary — key concepts, definitions, and intuitions.
2. The math derivation of the central algorithm (not a copy of the slides; re-derived in my own notation).
3. A Python/NumPy implementation of the algorithm from scratch — no scikit-learn except for sanity-checking.
4. A worked example on a real dataset (UCI, Kaggle, or a small dataset I generate).
5. "Where I got stuck" — the specific concept that required re-watching or extra reading, and how I unstuck myself.

Build-in-public reality: the value is not in re-explaining Andrew Ng — the value is in the friction points I hit and how a practising engineer (not a PhD student) works through them.

When every lecture has a published post with all five artefacts above, flip this post to status: published.

• Create summary comparison table of all algorithms covered