The first week of January is generous with resolve and vague on specifics. You decide to learn machine learning this year, open six browser tabs, and by the middle of February the tabs are gone and so is the resolve. The problem is rarely motivation. It is the absence of edges: no defined scope, no finish line, no artifact to point at when someone asks what you have been doing with your evenings.
What follows is a twelve-month plan built for people learning on their own, alongside a job, with maybe six to eight hours a week to spend. One theme per month, one thing built by the end of it. Adjust the pace, keep the shape.
Set the Terms Before You Start
Decide the rules now, while you are still enthusiastic, because in week nine you will not feel like negotiating with yourself.
- Six to eight hours a week, split into three or four sessions. Two 90-minute blocks beat seven distracted 20-minute ones.
- Every month ends with an artifact — a notebook, a script, a small app, a written summary. Something a stranger could open.
- Keep a learning log. One paragraph a week describing what confused you and what you did about it. This becomes your best diagnostic tool later.
- Protect one deep session with no phone in the room. Reading documentation while half-watching something is not studying.
Months 1–3: Python, Data, and the Language of Models
Month one is Python and nothing else: variables, loops, functions, dictionaries, files, and enough error reading to stop fearing a traceback. Then NumPy for arrays, pandas for tables, and matplotlib for looking at what you have. The goal is fluency, not mastery. If you can load a messy CSV, fix a column, and plot a distribution without searching for every line, you are ready to move on.
Month two is statistics. Means, medians, variance, distributions, sampling, correlation versus causation, and what a confidence interval actually claims. Do this with data in hand rather than as theory — take a dataset you care about, maybe your own spending or local weather records, and describe it honestly in a notebook. Confusing a sample for a population is the most common mistake beginners make, and it starts here.
Month three is linear algebra and calculus intuition. Vectors, matrices, dot products, matrix multiplication, derivatives, and gradients. You do not need proofs. You need to know what a gradient tells you and why matrix shapes must match. By the end of month three, write a short piece explaining gradient descent to a friend who has never coded. If you can do that without hand-waving, the foundations are in place.
Months 4–6: Classical Machine Learning, End to End
Months four through six are the most valuable part of the year, and the part most self-taught learners skip in a rush toward neural networks. Use scikit-learn. Learn the loop properly:
- Split the data, and split it before you look at anything else.
- Build the simplest possible baseline — often predicting the mean or the most common class.
- Fit one model and read its errors, not just its score.
- Improve one thing at a time: features, then model, then hyperparameters.
- Write down what you tried and what happened. Notebooks lie by omission.
Cover linear and logistic regression, decision trees, random forests, and gradient boosting. Learn what overfitting looks like on a learning curve. Learn why accuracy is a poor metric on imbalanced data, and get comfortable with precision, recall, F1, and RMSE. Add k-means clustering and PCA so unsupervised methods are not a mystery. Three solid projects — a regression on tabular data, a classification with a real evaluation, and a clustering exploration — will teach you more than any course you watch at 1.5x speed.
Months 7–9: Neural Networks, Without the Ceremony
Pick one framework, PyTorch or TensorFlow, and stay with it. Week one: tensors, autograd, a training loop you type by hand rather than import. Train a small network on a standard image dataset and plot the loss curves. Then deliberately break it — set the learning rate too high, remove regularization, use a tiny batch — and watch what happens. Failure modes are the real curriculum.
From there: convolutional networks for images, embeddings for text, and the attention mechanism that sits inside modern sequence models. You do not need to train a large language model from scratch to understand how one works. Fine-tuning a small pretrained model on your own dataset, and explaining what changed, is a stronger signal of understanding than a benchmark score. Free notebook environments with modest GPU access are enough for everything described here.
Months 10–11: Choose a Direction and Ship Something
Pick one area based on the work you want, not the trend you saw last week: computer vision, natural language, tabular prediction, recommender systems, time series, or the engineering side of getting models into production. Spend month ten going deep in that one area — read three or four papers or long technical posts and reimplement the smallest one.
Month eleven is a single end-to-end project. Collect or clean the data yourself. Start with a baseline. Improve it. Then wrap it: a simple API or a small web interface, a README a stranger can follow, and a note on what you would do differently with more time. Deployment is where the unglamorous skills live, and they are the ones that get people hired.
Month 12: Consolidate, Publish, Choose Next Year's Edge
Spend the final month tidying rather than starting. Refactor your three best notebooks into clean scripts. Write two posts explaining something you found hard — teaching forces precision. Re-read your weekly log from January and note how many of those early confusions now feel trivial. Then plan next year around the gap you can still see: a competition, an open-source contribution, a reading group, or a formal course if structure suits you better than solitude.
Falling four days behind is not a reason to restart the year. It is a reason to study on day five.
Making It Stick
Put the sessions in your calendar as appointments, and treat them like ones you cannot move. Find one person to check in with weekly, even briefly; explaining a concept out loud exposes the gaps fast. When a month goes badly, shrink the goal rather than abandoning it — thirty minutes twice a week still moves you forward.
Most of all, keep the artifacts. A year from now you will not remember which tutorial you finished, but you will have twelve small things you built, and a clear picture of what you can do. That is a far better January than six open tabs.
Photo: K / Pexels
