Machine learning from scratch
Take this if you can call scikit-learn and want to know what each of those functions is actually doing.
What you’ll be able to do
You can choose an appropriate model family for a problem, implement the core ones yourself, evaluate them without fooling yourself, and engineer features that improve results more than swapping models does.
What it assumes
Calculus, linear algebra and probability at the level of the Foundations phase, plus fluent Python with NumPy and Pandas.
- Price
- ₹1,499
- Length
- 12 weeks
- Sessions
- 12
- Effort
- ~7 h/week · 84 h total
- For
- Class 11–12 · College · Working professional
Adds it to My courses on your dashboard so you can find it again. Payment is not open yet — nothing is charged.
What’s taught, in order
- 01
Regression, and what fitting means
Linear and logistic regression derived and implemented from the loss up, so that fitting is an optimisation you have written rather than a method you have called.
- 02
Evaluation before more models
Splits, cross-validation, metrics and model selection taught early on purpose, because a practitioner who can evaluate is more useful than one who knows more algorithms.
- 03
Trees, and then forests
Decision trees implemented with a real splitting criterion, then bagging and random forests. Watching variance drop as you average is the clearest lesson in the phase.
- 04
Boosting, and margins
Gradient boosting built up from the residual idea, plus support vector machines and kernels. These are still the models that win on tabular data, which is worth knowing before reaching for a network.
- 05
Unsupervised methods
Clustering, dimensionality reduction, naive Bayes, k-nearest neighbours and anomaly detection — the toolkit for when nobody has labelled anything for you.
- 06
Feature engineering and a capstone
The strand that produces most real-world gains, then a capstone that puts the whole phase against one problem end to end.
What you build
Take one substantial tabular problem and carry it from raw data to a defended final model: baseline first, then at least three model families compared under identical evaluation, a documented feature engineering pass with its measured effect, and a written recommendation including what you would monitor if this model were deployed.
Check it yourself against this
- A trivial baseline is reported before any real model, and beaten.
- Three or more model families compared under one identical evaluation protocol.
- At least one model implemented by you from scratch and validated against the library version.
- Feature engineering effects are measured separately from model choice.
- The recommendation names the metric it optimises and what it sacrifices.
- A monitoring plan states what would tell you the model had gone stale.
Then answer this
Across your experiments, did changing the model or changing the features move your metric more — and what does your answer imply about where to spend effort next time?
Two or three sentences, in your own words. If you cannot, the course is not finished — go back to the module it came from.
What this course does not do
No neural networks and no unstructured data: text, images and audio are the Deep Learning phase. Nothing here is deployed either, so serving, monitoring and drift belong to the Production phase rather than this one.
Where this leads
The Classical Machine Learning phase page lists the twelve weeks individually with the model implemented in each, so you can see exactly which algorithms you will have written by hand at the end.
FAQ
- Why implement models that libraries already provide?
- Because the implementation is where the assumptions live. Someone who has written a decision tree's splitting criterion knows why it overfits; someone who has only called one reaches for a deeper tree. You use the libraries too, immediately after, with a much better idea of which knob to turn.
- Is classical ML still worth twelve weeks in a deep learning world?
- For tabular data, which is most business data, gradient boosting still beats neural networks routinely while training in seconds and remaining inspectable. The evaluation discipline this phase builds also transfers directly to deep learning, where it is harder to acquire.
- Can I take this before the Foundations phase?
- Only if you already have the maths. The derivations in weeks one and four assume you can differentiate a composed function and read a matrix as a transformation; without that, you will be pattern-matching rather than following.
Other courses
This is a standalone course. It does not add weeks to the 78-week curriculum or change your roadmap. If you want the whole path instead, the full programme is ₹399/month · ₹2,000/year — see what it covers.