Learn machine learning to train a real model in nine months
Nine months of serious work — about an hour a day on one specialization, one practical course, and one personal project — gets a coder who knows Python from no ML to training and evaluating real models on real data. Roughly 270 hours total. You will not be a research scientist. You will be a competent applied ML practitioner.
9 months · ~270 hours · train and deploy a deep learning model on your own dataset
1.Andrew Ng's Machine Learning Specialization
The 2022 rewrite on Coursera is the consensus first ML course on the planet, and it earns the title. Three courses cover supervised learning, advanced learning algorithms, and unsupervised + recommender systems — all in Python with NumPy and scikit-learn. Andrew teaches gradient descent the way nobody else does. Audit free, or pay the subscription if you want graded labs and a certificate.
Free to audit; $49/month for graded labs
Machine Learning Specialization →2.Fast.ai — Practical Deep Learning for Coders
Jeremy Howard's course is the polar opposite of Andrew's: top-down, code-first, train-a-state-of-the-art-model-in-lesson-one. Nine lessons at ninety minutes each, plus the free book that goes deeper. You will train image classifiers, language models, and tabular models on Kaggle datasets using PyTorch and the fastai library. Do not skip the homework. The companion fastbook explains the underpinnings the videos gloss over.
Free; ~$30 in Colab Pro or Paperspace credits
Practical Deep Learning →3.Train a model on your own data
Pick a problem nobody has solved for you — your hometown's traffic accident data, your photos sorted by who is in them, your inbox classified by importance — and build the full pipeline yourself. Collect, clean, train, evaluate, deploy via Hugging Face Spaces or a Streamlit app. This is the project that decides whether ML stays a hobby or becomes a tool. Document your failures in a README. Most learning happens here.
Free–$20/month in compute
Hugging Face Spaces →If this doesn't fit you
If you have a strong math background and want a deeper theoretical foundation rather than the applied path, replace fast.ai with Stanford's CS229 lectures (free on YouTube) plus Murphy's Probabilistic Machine Learning textbook. It is harder, slower, and only worth it if you intend to read research papers or pursue a graduate degree.
Why this path
The two most-recommended ML courses on Earth are also the two best, and they complement each other almost perfectly: Andrew gives you bottom-up intuition, Jeremy gives you top-down practice. Most beginners pick one and stop, which is why most self-taught ML people can talk about models but cannot ship one. Skipping the personal project is the most common mistake. The Coursera certificate impresses nobody; a deployed model with your own data does.