02 — Courses Canon¶
Courses are not the curriculum — they’re scaffolding. Use them to build intuition, then tear them down and work with primary sources (papers, implementations) once the concept is internalized. The failure mode is treating course completion as knowledge acquisition. It isn’t. Finishing a course means you’ve seen the material once. That’s step one of six.
What most people get wrong: They stack courses. Three courses on the same topic is not three times the understanding — it’s one understanding with two distractions. Pick one course per topic, go deep, and build something with it before moving on.
Course Evaluation Criteria¶
Each course below is rated on:
Technical depth (1-5): Does it explain why, or just how?
Recency (as of 2026): Is the content current, or are you learning pre-transformer deep learning?
Community signal (1-5): What are practitioners actually saying vs. what marketing says?
Cost: Free, audit-free, or paid?
Time commitment: Realistic estimate, not the course page estimate
The Canon¶
1. Neural Networks: Zero to Hero¶
Instructor: Andrej Karpathy | Platform: YouTube | Year: 2022-2023 (still the community standard)
What it covers: Builds neural networks from absolute scratch in Python/NumPy/PyTorch. Micrograd → makemore (character-level LM) → GPT-2 from scratch → tokenization. 8 videos, ~15 hours total.
Technical depth: 5/5 — Karpathy codes live, makes mistakes, explains everything. This is what “understanding” looks like.
Community signal: 5/5 — Reddit ML, Discord, Twitter: consistently the top-cited course for anyone who wants to actually understand what’s happening inside a neural network. “After this course, backprop stopped being magic” is a near-universal comment thread reaction.
Recency: 5/5 — Covers GPT, attention, BPE tokenization. Directly relevant to 2024-2026 practice.
Cost: Free (YouTube playlist)
Time commitment: 15-20 hours of watching + 30-40 hours of implementing alongside = 50-60 hours total to do it properly.
⚠️ Note: This is not a beginner course. You need basic Python, basic calculus, and a rough understanding of what a neural network is. If you’re starting from zero, do one of the entry-level courses below first, then return.
Roadmap phase: Months 2-4 (core deep learning phase — this is the spine)
2. fast.ai Practical Deep Learning for Coders¶
Instructor: Jeremy Howard | Platform: fast.ai | Year: 2022 version (Part 1), with ongoing updates
What it covers: Computer vision, NLP, tabular data, collaborative filtering, training tricks, data augmentation, transfer learning — all using PyTorch and the fast.ai library. Top-down philosophy: get something working first, understand math second.
Technical depth: 3/5 — Deliberately trades depth for speed of implementation. The philosophy is correct for getting unstuck, but pair with Karpathy for depth.
Community signal: 5/5 — Fast.ai has an unusually devoted community. The forums are active. Reddit: “Fast.ai is what gets you doing things, not just watching lectures.”
Recency: 4/5 — 2022 version is still the primary version. Jeremy Howard gives occasional talks. fast.ai library is active, but some API calls in the 2022 notebooks are deprecated. Check the forums for updated notebooks.
Cost: Free (fast.ai website)
Time commitment: 7-8 weeks at ~3-4 hours/week = ~25 hours
Roadmap phase: Months 1-3 (entry-level practical course — ideal starting point for a 1-year ML engineer)
3. Machine Learning Specialization (Coursera)¶
Instructor: Andrew Ng + DeepLearning.AI | Platform: Coursera | Year: 2022 revamp
What it covers: Supervised learning (linear/logistic regression, SVMs, decision trees, ensembles), neural networks, unsupervised learning, recommender systems, RL. Updated 2022 Python/NumPy version.
Technical depth: 2/5 — Deliberately simplified. Ng’s strength is clarity, not depth. “Hand-holding for math” is the frequent criticism.
Community signal: 3/5 — Most-mentioned course on Reddit, but with significant caveats. Beginners love it. Experienced engineers often call it “fine but insufficient.” Community verdict: “Good for your first exposure, but don’t stop here.”
Recency: 4/5 — 2022 version replaced the old Octave-based version. ⚠️ Pre-2022 versions are obsolete — if you see recommendations for the original ML course using Octave/MATLAB, that content is from 2011. Avoid.
Cost: ~$49/month (Coursera) or free audit. No legitimate free PDF. Financial aid available for genuine hardship cases via Coursera’s application.
Time commitment: 3 months at 5-7 hours/week
⚠️ India note: Coursera’s Financial Aid option is available; the application takes ~2 weeks to process.
Roadmap phase: Month 1 (classical ML foundations — skip if you already have solid fundamentals)
4. Deep Learning Specialization (Coursera)¶
Instructor: Andrew Ng | Platform: Coursera / DeepLearning.AI | Year: 2017, minor updates
What it covers: Neural network basics, optimization, hyperparameter tuning, batch norm, CNNs, RNNs, sequence models, attention. 5 courses.
Technical depth: 2.5/5 — Better than the ML specialization but still skims over implementation details. Ng is excellent at conceptual clarity and weak at production-grade code.
Community signal: 4/5 — 120K+ reviews, 4.9/5 on Coursera. But this rating partly reflects accessibility, not depth. Practitioners note: “Good survey but go deeper after this.”
Recency: 2.5/5 — ⚠️ The content is from 2017-2018. Transformers are covered only superficially in Course 5. The RNN section teaches architectures that are now rarely used in practice. Treat as historical foundation, not current practice.
Cost: ~$49/month Coursera subscription
Time commitment: ~4 months at 5 hours/week
Roadmap phase: Months 2-3 (neural network foundations — after this, go to Karpathy for depth)
5. Stanford CS231n: Convolutional Neural Networks for Visual Recognition¶
Platform: Stanford (free) | Year: 2024 latest version, 2017 Karpathy version widely cited
What it covers: Image classification, CNNs (convolution, pooling, architectures), training tricks, RNNs, detection, segmentation, generative models. Graduate-level computer vision.
Technical depth: 5/5 — Lecture notes are among the best-written technical documents in ML education. Assignments require implementing backprop through convolutional layers from scratch.
Community signal: 5/5 — Reddit and Discord consistently cite CS231n as the gold standard for computer vision fundamentals. Karpathy’s 2017 version is frequently recommended alongside the 2024 version.
Recency: 4/5 — 2024 version available. Covers modern architectures. Note: Vision Transformers (ViT) are newer than earlier versions but are covered in recent iterations.
Cost: Free — lectures on YouTube, assignments on GitHub, notes at cs231n.github.io
Time commitment: 10-12 weeks at 8-10 hours/week (it’s a full Stanford course)
Roadmap phase: Months 4-6 (computer vision deep dive — the assignments are more valuable than the lectures)
6. Stanford CS224n: Natural Language Processing with Deep Learning¶
Platform: Stanford (free) | Year: 2024 latest version
What it covers: Word vectors, neural classifiers, RNNs, LSTMs, transformers (deep coverage), BERT, GPT, pretraining, fine-tuning, QA systems, summarization, coreference.
Technical depth: 5/5 — Christopher Manning’s course. Assignment 5 requires implementing a transformer from scratch.
Community signal: 5/5 — The NLP equivalent of CS231n. Reddit: “If you want to understand transformers mechanistically, CS224n and Karpathy’s video are the combination.”
Recency: 5/5 — Updated annually. 2024 version covers modern LLMs, instruction tuning, RLHF.
Cost: Free — web.stanford.edu/class/cs224n/
Time commitment: 10-12 weeks at 8-10 hours/week
Roadmap phase: Months 6-8 (NLP/transformer deep dive)
7. Hugging Face NLP Course¶
Platform: HuggingFace (free) | Year: 2021, continuously updated
What it covers: HuggingFace Transformers library end-to-end. Tokenizers, datasets library, model fine-tuning, sharing models, building pipelines. Practical.
Technical depth: 3/5 — Library-focused, not theory-focused. Assumes you understand transformers; teaches you how to use the HF ecosystem.
Community signal: 5/5 — This is the standard onboarding for the HF ecosystem. No alternative for learning HF-specific APIs.
Recency: 5/5 — Actively maintained, includes PEFT, fine-tuning, alignment chapters.
Cost: Free at huggingface.co/course
Time commitment: 20-30 hours
Roadmap phase: Months 5-7 (once you understand transformers conceptually, learn the HF tooling)
8. MIT 6.S191: Introduction to Deep Learning¶
Platform: MIT OpenCourseWare / YouTube | Year: Updated annually (2025 version covers diffusion models, LLMs, RL)
What it covers: 1-week bootcamp format. Deep learning fundamentals, CNNs, RNNs, transformers, generative models (VAEs, GANs, diffusion), RL, fairness. Updated to include the latest models each year.
Technical depth: 3/5 — Breadth over depth. Excellent survey; weak on implementation.
Community signal: 4/5 — 60+ Reddit mentions in 2024-2025. Strong for its annual updates and clean lecture style. Reddit: “Best for getting current — they cover what’s happening this year.”
Recency: 5/5 — Literally updated in 2025. Includes content that books haven’t caught up to yet.
Cost: Free
Time commitment: 1-2 weeks intensive (~20 hours)
Roadmap phase: Month 1 (orientation) and revisit annually to update on latest developments
9. Practical Deep Learning for Coders Part 2 (fast.ai)¶
Instructor: Jeremy Howard | Platform: fast.ai | Year: 2022
What it covers: Goes deeper into the fast.ai library internals, implementing from scratch (training loop, optimizer, callbacks), diffusion models, stable diffusion implementation.
Technical depth: 4/5 — Significantly deeper than Part 1. Howard implements the training loop, custom optimizers, and diffusion models.
Community signal: 4/5 — Smaller audience than Part 1 but very positive. “Part 2 is where you start to really understand fast.ai and deep learning.”
Recency: 4/5 — 2022 version covers diffusion models, which is still highly relevant.
Cost: Free
Time commitment: 30-40 hours
Roadmap phase: Months 5-7 (after completing Part 1 and Karpathy — for those doing the vision track)
10. Full Stack Deep Learning (FSDL)¶
Instructor: Pieter Abbeel et al. (Berkeley) | Platform: FSDL website / YouTube | Year: 2022 cohort (latest full course)
What it covers: The production ML stack — data collection, labeling, experiment tracking, model training, deployment, CI/CD for ML, monitoring in production. Not theory — engineering.
Technical depth: 4/5 — Uniquely focused on the gap between research and deployment. Covers tools (W&B, MLflow, DVC, FastAPI) that no academic course touches.
Community signal: 4/5 — The only production-focused curriculum that Reddit and Discord consistently recommend. “If you want to know how to ship models, FSDL is the only course that teaches this.”
Recency: 3/5 — ⚠️ 2022 cohort is the latest full version. No 2023 or 2024 cohort has been announced as of mid-2026. Some tooling content is showing age (e.g., specific version references). Check fullstackdeeplearning.com for updates.
Cost: Free (YouTube), was paid during live cohort
Time commitment: ~30 hours lectures + projects
Roadmap phase: Months 10-13 (production/MLOps phase)
11. DeepLearning.AI Short Courses¶
Platform: DeepLearning.AI (learn.deeplearning.ai) | Year: Ongoing releases 2023-2026
What it covers: 1-2 hour micro-courses on specific topics: LangChain, RAG, function calling, fine-tuning, AI agents, LLM evaluation, vector databases, multimodal AI.
Technical depth: 2/5 — Very introductory. Good for a quick mental model of a new API/concept.
Community signal: 3/5 — “Good for quickly understanding what something is, not for learning how it actually works.” Treat as a first 30-minute orientation, not a course.
Recency: 5/5 — Released frequently, covering latest tools and frameworks.
Cost: Free
Roadmap phase: Throughout — use as orientation before diving into a new topic
12. Practical Reinforcement Learning (fast.ai / Spinning Up)¶
Platform: OpenAI Spinning Up | Year: 2018 (still canonical for RL foundations)
What it covers: RL fundamentals — Markov decision processes, policy gradients, Q-learning, actor-critic methods. OpenAI’s Spinning Up implementation includes clean PyTorch implementations of PPO, SAC, DDPG, TD3.
Technical depth: 5/5 — The code is production-quality and educational simultaneously.
Community signal: 4/5 — Spinning Up is the go-to RL implementation reference. For learning RL from scratch alongside theory, combine with David Silver’s UCL RL lectures (free on YouTube).
Recency: 3/5 — 2018 content is still foundational for classic RL, but doesn’t cover recent RLHF or model-based RL advances.
Cost: Free
Roadmap phase: Months 10-12 (RL/RLHF phase — optional track)
Course Decision Matrix¶
Goal |
Primary Course |
Supplement With |
|---|---|---|
Start fast, build something |
fast.ai Part 1 |
Karpathy Week 1 |
Understand deeply |
Karpathy Zero to Hero |
CS231n assignments |
Classical ML foundations |
Ng ML Specialization |
ISL book |
Computer vision track |
CS231n |
fast.ai Part 1 |
NLP/LLM track |
CS224n → HF NLP Course |
Karpathy |
Production/MLOps |
FSDL |
Chip Huyen book |
Current landscape |
MIT 6.S191 (annually) |
- |
Outdated Resources to Avoid¶
Resource |
Problem |
|---|---|
Original Andrew Ng ML Course (Octave/MATLAB) |
Pre-2022 version, obsolete language |
Fast.ai 2019 and earlier versions |
Deprecated library API |
cs231n before 2017 |
Architecture references are outdated |
Any TF 1.x course |
TF 2.x is not backward-compatible |
Return to README.md · Previous: 01_books_canon.md · Next: 03_papers_canon.md