Most people who try to learn machine learning quit within a week. Rarely because the subject is too hard — usually because they start in the wrong place: a four-hour lecture on linear algebra, a course that assumes you already write code, a textbook that spends forty pages on notation before showing a single result. Seven days is enough to build a working mental model, train a few models, and know exactly what to study next. This plan assumes nothing except roughly two hours a day and the willingness to type code you do not fully understand yet.

Before Day One: A Workspace and a Notes File

You do not need an expensive laptop or a paid course. Everything here runs in a free browser-based notebook — the kind that gives you a blank cell, a run button, and the output underneath. Create an account, open a new notebook, print one line of text. If setup takes longer than half an hour, you are over-engineering it.

Then create a plain text file called notes.md. Every time something clicks, write it down in your own words, one sentence, no jargon. By day seven that file will be worth more than any cheat sheet you could download, because it is written in the only vocabulary you actually think in.

Days 1 and 2: Just Enough Python to Be Useful

Two days, not two months. You need five things and nothing else:

• Variables and types — storing a value like 0.85 under a name, and knowing the difference between the number 0.85 and the text "0.85".
• Lists and dictionaries — holding many values in order, and looking one up by name.
• Loops — applying the same small operation to every item in a collection.
• Functions — giving a block of code a name so you can reuse it instead of pasting it twice.
• Imports — pulling in someone else's library and calling their functions with your data.

Skip classes, decorators, generators and virtual environments for now. Type every example by hand; copying and pasting teaches your fingers nothing. When something breaks, read the error message out loud — it usually names the problem outright. A message complaining that "scoer" is not defined means you misspelled a variable three lines up.

Day 3: Understand What a Model Actually Is

Strip away the vocabulary and a model is a function: inputs go in, an output comes out. The difference from ordinary code is who writes the rules. You do not hand it a list of conditions; you hand it examples, and it adjusts numbers inside itself until its outputs line up with the answers you gave it.

Take spam filtering. The inputs might be counts of certain words; the output is a yes or a no. The model has no idea what spam means. It has found a set of weights that separates the examples you showed it, and it will happily misjudge anything that looks unlike them.

Three words worth knowing this week

• Features — the inputs you feed the model, one column per feature.
• Labels — the answers you want it to predict.
• Training — the process of adjusting those internal numbers.

Regularization, gradient descent and cross-validation can wait. This week you need one sentence: data goes in, numbers get adjusted, predictions come out.

Day 4: Train Your First Model End to End

Use a small built-in dataset — the iris flowers one is a classic for a reason. Work in this order:

1. Load the data into a table and print the first five rows.
2. Ask three questions of it: how many rows, what type is each column, and are any values missing?
3. Split the columns into features and the one column you want to predict.
4. Split the rows into a training portion and a testing portion. An 80/20 split is a fine default.
5. Create a simple model — logistic regression or a decision tree is plenty.
6. Fit it on the training portion only, predict on the testing portion, and score the result.

Whatever number you get, write it down. Accuracy of 0.92 and 0.61 are equally useful right now: they are your baseline, the thing every future change gets compared against.

Day 5: Measure Honestly

Testing a model on the data it trained on is like handing students the exam they already saw. The scores look great and mean nothing. The testing portion exists so you measure how the model behaves on rows it has never seen, which is the only situation that matters once it leaves your notebook.

When the training score is high and the testing score is low, the model memorized instead of learning. You will hear this called overfitting, and the usual fixes are a simpler model, fewer cluttered features, and more examples.

Accuracy alone can also mislead. If only a small fraction of cases belong to the class you care about, a model that always says "no" can score well and still be useless. Precision and recall exist to catch that. You do not need to master them this week — just know that one number is never the whole story.

Day 6: Break It, Then Fix It

Day six is for poking at your model instead of polishing it. Change one thing at a time and record what happens:

• Train on half the data. Does the score move?
• Swap in a different model type. Better, worse, or the same?
• Delete the column you suspect matters least. Any change?
• Feed it one made-up row and read the prediction. Does it make sense to you?

That last test is the most revealing, and it is close to what the work actually looks like. Most of the job is not writing clever code. It is noticing that something is off and tracing it back to a bad column, a missing value or a label you misread.

Run the code, change one thing, run it again. Repeat that loop a few hundred times and you have the core skill.

Day 7: Build One Small Thing of Your Own

Pick data you actually care about: your own spreadsheet, a public dataset about your city, a list of films you have rated. Small and personal beats big and famous, because you will spot nonsense in the results immediately.

Finish with one notebook that loads data, trains a model, scores it, and opens with three plain sentences: what it does, how well it works, and what it cannot do. That third sentence separates people who are learning from people who are collecting certificates.

How to Keep Going After Day Seven

Week one buys you vocabulary and one finished project. That is a real win, and it is also a beginning rather than an end. The distance from here to a job or a serious side project is mostly volume, plus honesty about the gaps.

• Redo the plan on a new dataset without opening your

  Photo: 1139623 / Pixabay