Back to the fundamentals: I'm taking Andrew Ng's Machine Learning Specialization

I spend my days building products with LLMs and agents. And yet I decided to go back to the very beginning: linear regression, gradient descent, bias and variance.

Andrew Ng's Machine Learning Specialization (Stanford + DeepLearning.AI, on Coursera) is the course I'm taking now. This post is less a review and more a statement of intent — plus a map of what lives inside each course, for anyone thinking about taking it too.

Why a "classic" course in the age of agents

It's tempting to assume that, with ready-made models behind an API, understanding what happens underneath has become optional. I disagree.

Working well with AI today requires intuition about how models learn: why a model memorizes instead of generalizing, what a high training error versus a high validation error means, when more data helps and when it doesn't. That's the basic vocabulary — and it didn't come from LLMs, it came from classic ML.

I want to stop treating these ideas as a black box. Gradient descent, regularization, the difference between bias and variance: they all show up, wearing different clothes, in everything I already do with agents.

What the specialization is

It's three courses, designed for people who can code but don't necessarily have an ML background. The focus is intuition first, the right amount of math, and hands-on practice in Python with NumPy, scikit-learn and TensorFlow.

Andrew Ng's real edge is pedagogical: he builds intuition before the formula. You understand why before you see the how.

Course 1 — Supervised Machine Learning: Regression and Classification

The foundation. This is where the "hello world" of ML lives:

• Linear regression — predicting a continuous number, the cost function, and gradient descent as the engine that tunes the model.
• Multiple regression — many features, feature scaling, feature engineering and polynomial regression.
• Classification with logistic regression — decision boundary, and your first serious encounter with overfitting and regularization.

If you only take this one, you'll already understand the full cycle of training a supervised model.

Course 2 — Advanced Learning Algorithms

This is where it gets interesting:

• Neural networks — inference and training, activation functions, and why they unlocked so much.
• Multiclass classification — softmax and the practical details.
• ML best practices — for me the most valuable part: bias/variance in practice, train/validation/test splits, error analysis, and how to iterate on an ML project without spinning your wheels.
• Decision trees — trees, random forests and XGBoost, which still beat neural nets on a lot of tabular data.

The "how to debug a model that isn't good enough" section alone justifies the course.

Course 3 — Unsupervised Learning, Recommenders, Reinforcement Learning

The closer, covering the rest of the map:

• Unsupervised learning — clustering with K-means and anomaly detection.
• Recommender systems — collaborative filtering and a content-based approach with deep learning.
• Reinforcement learning — a practical introduction, with state-action value and a Deep Q-Network.

It's the course that talks most directly to product: recommendation and anomaly detection show up constantly in the real world.

How I'm going to study

My plan is simple:

1. Follow along with the YouTube playlist of the lectures.
2. Do the labs and exercises by hand before looking at any solution.
3. Use this repository of solutions only as a check, not a shortcut. The difference between learning and copying lives exactly there.

The rule I'll follow: if I can't explain a concept out loud, I haven't really understood it.

What I expect to take from this

I'm not doing this to switch fields. I'm doing it to become a better AI engineer — the kind who understands what's happening beneath the abstraction.

When I tune a RAG system, evaluate an agent, or stare at an eval metric, I want the foundation that turns intuition into decisions. Going back to the fundamentals isn't a step backward. It's building the groundwork that was missing.

As I progress, I'll write about whatever surprises me along the way. If you're thinking about taking it too — let's do it together.