Andrew Ng's courses (CS229, Coursera ML Specialization) vs fast.ai's Practical Deep Learning: approach, math, coding and who should start where.
Free for up to 3 courses. Pro adds AI help for $9/month.
Andrew Ng and Jeremy Howard (fast.ai) are probably the two most recommended machine learning teachers on the internet, and they teach in opposite directions. Ng starts with the math and builds up to applications. Howard starts with a working model and digs down into how it works. Both approaches produce people who understand ML. The question is which one keeps you going long enough to get there.
"Andrew Ng's course" can mean several things: - **Stanford CS229 (lectures on YouTube):** A graduate-level course with full derivations. Math-heavy. - **Machine Learning Specialization (Coursera, with DeepLearning.AI and Stanford):** The beginner-friendly successor to his original Coursera course, using Python. Can be audited; certificates are paid. - **Deep Learning Specialization (Coursera):** Neural networks, CNNs, sequence models. Also auditable. This comparison focuses on the Machine Learning Specialization and CS229 versus fast.ai's Practical Deep Learning for Coders, which is free on YouTube and course.fast.ai.
| Andrew Ng (ML Specialization / CS229) | fast.ai Practical Deep Learning | |
|---|---|---|
| Direction | Bottom-up: theory first | Top-down: working models first |
| Math | Moderate (Specialization) to heavy (CS229) | Introduced only when needed |
| Coding | Guided notebooks; NumPy and scikit-learn style | Real models in PyTorch and the fastai library from lesson one |
| Topics | Regression, classification, trees, clustering, recommender systems, RL basics | Vision, NLP, tabular data, collaborative filtering, then the internals |
| Pace | Steady, very clear | Fast; lessons are long and dense |
| Best for | People who want to understand before they build | Coders who want results and learn by tinkering |
**Clarity.** Ng is famous for patient, clear explanations. Concepts like gradient descent, regularization, and bias-variance tradeoffs are explained so well that many people never need another source. **Foundation for the classics.** Much real-world ML is still tabular data with linear models, trees and gradient boosting. Ng covers that ground thoroughly. **Confidence with math.** If you plan to read papers, pursue research, or take a master's degree, the habit of understanding derivations pays off. **The downside:** you may spend weeks before building anything that feels impressive, and the theory-first approach loses some people before the fun part.
**Early wins.** You train a real image classifier in the first lesson. Seeing it work is motivating in a way equations are not. **Practical judgement.** Jeremy Howard spends a lot of time on what practitioners actually do: picking learning rates, handling data, avoiding overfitting, deploying models. These are the things that separate working ML from homework ML. **Real tools.** You use PyTorch, which is the dominant framework in research and widely used in industry. **The downside:** the fastai library abstracts a lot, and if you never go deeper you may not understand what is happening underneath. Lessons are long and assume you are comfortable with Python.
- **New to programming:** neither yet. Learn Python first. - **Comfortable with Python, weak in math:** fast.ai, then backfill math with 3Blue1Brown and StatQuest as questions come up. - **Strong in math, want rigor:** CS229, or the ML Specialization if you want a gentler start. - **Want a job in data science on tabular data:** Ng's ML Specialization, plus scikit-learn projects. - **Want to work with deep learning or LLMs:** fast.ai, then Karpathy's Neural Networks: Zero to Hero. A sequence that works for many people: fast.ai part 1 for motivation and practical skill, then Ng's material to firm up the foundations.
Both courses ask for real effort: fast.ai lessons can take a full day each if you do the notebooks properly, and CS229 problem sets are hard. Keep notes with timestamps for the moments that confused you and test yourself before moving on. If you paste the YouTube lectures into Filow, you get progress tracking, timestamped notes and AI quizzes per lecture.
Neither is better in general. fast.ai suits coders who learn by doing; Andrew Ng suits people who want to understand theory first. Many learners benefit from both.
Yes. Practical Deep Learning for Coders is free on course.fast.ai and YouTube, along with its notebooks and book.
CS229 lectures are free on YouTube. The Coursera Machine Learning Specialization can generally be audited for free, while the certificate requires payment.
Not at the start. The course introduces the math as needed. For deeper understanding later, linear algebra and calculus help.
or $79 a year
Get ProTurn the CS229 or fast.ai lectures into a tracked course with notes and quizzes on Filow.
Or browse community courses