The learning path
What to learn, roughly in order. Prefer one guided course? Any of the anchors below covers the fundamentals end to end. Otherwise, follow the path topic by topic.
If you'd rather follow one guided course than assemble your own path, any of these is a solid spine. Pick one and stick with it.
- Andrew Ng — Machine Learning Specialization — the classic first course; gentle and thorough. Coursera →
- fast.ai — Practical Deep Learning for Coders — top-down and code-first; you train real models in lesson one. course.fast.ai →
- Google — Machine Learning Crash Course — short, free, and hands-on. Crash Course →
- Andrej Karpathy — Neural Networks: Zero to Hero — build neural nets (and a small GPT) from scratch. Zero to Hero →
- Kaggle Learn — bite-sized, practical micro-courses (Python, pandas, intro ML). kaggle.com/learn →
A mix of intuition-builders, standard references (several free online), and one on organizing what you learn. You don't need all of them — pick what fits where you are.
- Why Machines Learn — Anil Ananthaswamy. The elegant math behind modern AI, explained for humans; a great conceptual bridge before the heavier texts. Book →
- Mathematics for Machine Learning — Deisenroth, Faisal & Ong. The linear algebra, calculus, and probability you need, in one place. Free online. Read free →
- Deep Learning — Goodfellow, Bengio & Courville. The foundational deep-learning reference. Free online. Read free →
- Reinforcement Learning: An Introduction — Sutton & Barto. The standard RL text. Free PDF. Read free →
- Foundations of Computer Vision — Torralba, Isola & Freeman (MIT Press, 2024). A modern, thorough vision text — and its closing chapters on how to do research, write papers, and give talks are worth reading whatever your field. Free online. Read free →
- Vision Language Models — Noyan, Marafioti, Farré & Zohar (O'Reilly, 2026). A current, hands-on take on multimodal models — the newest book on the list. O'Reilly →
- AI Assurance — Batarseh & Freeman (Academic Press, 2023). Methods for testing and assuring trustworthy, safe, and explainable AI — directly relevant if you head toward AI assurance. Book →
- Build a Second Brain — Tiago Forte. Not about AI — a system for capturing and organizing what you read and learn, which pays off across a whole research career. Book →
Learn to program in Python
Python is the language of AI — start here before any model. Get comfortable with variables, lists and dictionaries, loops, functions, and classes, and with reading error messages. You don't need to be an expert, just fluent enough to express ideas in code.
The math you'll actually use
You don't need a math degree, but an intuition for three areas pays off everywhere: linear algebra (vectors and matrices — how data and model weights are represented), calculus (derivatives and gradients — how models learn), and probability & statistics (distributions, expectation, and how to evaluate results). Aim for intuition first; the notation follows.
Work with data — NumPy, pandas, notebooks
Real ML is mostly data work. Learn to load, clean, and explore data before you model it: NumPy for numerical arrays, pandas for tables, Matplotlib for plots, and Jupyter notebooks for fast, interactive experiments.
Classical machine learning
Start modeling with scikit-learn — a few lines each for linear/logistic regression, SVMs, and random forests. These teach the vocabulary everything else builds on: features and labels, the train/validation/test split, overfitting vs. underfitting, ensembles, and evaluation metrics (especially AUC, and why accuracy alone can mislead on imbalanced data). Master these ideas here and deep learning is far less mysterious later.
Deep learning & neural networks
Write one raw PyTorch training loop by hand so the mechanics aren't a mystery, then move to PyTorch Lightning, which removes the boilerplate so you write the science. Learn what a convolution does and why CNNs suit images, then work up to attention and transformers — the architecture behind modern language and vision models. 3Blue1Brown's Neural Networks series is the best visual companion.
Generative models
How modern AI creates images, audio, and video. Three families worth knowing: VAEs (encoder → compact latent space → decoder), GANs (a generator and a discriminator in a cat-and-mouse game), and diffusion (start from noise and denoise toward an image) — the approach behind today's best image and video generators.
LLMs & agents
Large language models and how to build on them. Learn to prompt well first — clear, specific instructions, and always verify what the model produces. Then understand the ladder: a task is one model call, a workflow is several calls on rails you define, and an agent directs its own steps and tools toward a goal. The building block is the augmented LLM — a model plus tools, retrieval, and memory.
A shared vocabulary makes everything above click faster. You'll meet these again and again — get a rough feel for each.
- Supervised vs. unsupervised vs. reinforcement learning — learning from labeled examples, finding structure in unlabeled data, or learning from reward through trial and error.
- Features & labels — the inputs a model sees and the answer it's trying to predict.
- Train / validation / test split — data you learn from, data you tune on, and data you touch only once to measure honest performance.
- Overfitting vs. underfitting — memorizing the training data vs. being too simple to capture the pattern; the balance is the bias–variance trade-off.
- Loss & gradient descent — a number that measures how wrong the model is, and the algorithm that nudges its parameters to make that number smaller.
- Hyperparameters — the knobs you set before training (learning rate, model size, regularization), as opposed to the weights the model learns.
- Generalization — the whole point: performing well on data you've never seen, not just the training set.
- Evaluation metrics — accuracy, precision/recall, F1, and AUC; which one matters depends on the problem and how balanced the data is.