[reference] — from the roadmap.sh machine-learning roadmap. Orientation for the whole modeling half.

What machine learning is

Machine learning is getting a computer to learn a task from data instead of from hand-written rules. Rather than a programmer specifying “if the email contains X and Y, mark it spam,” you show the system thousands of labeled emails and it infers the rule — a function mapping input to output — by adjusting internal parameters (what a model is). Reach for ML when the pattern is too complex, variable, or poorly-understood to hand-specify; a plain rule beats a model whenever the logic is actually known (choosing the right tool).

The whole modeling half of this track is variations on one loop: guess → measure error (loss) → adjust parameters to reduce it → repeat (optimization), with the real goal being generalization — performing well on data never seen during training, not memorizing the training set.

The types of learning

The single most useful first classification of any ML problem — it determines which algorithms even apply:

Supervised learning — learning from labeled examples

You have inputs and their correct outputs (labels), and the model learns to predict the output for new inputs. The dominant paradigm in practice. Two sub-types by output shape:

  • Classification — predict a category (spam/not-spam, which digit, which disease). Output is discrete.
  • Regression — predict a continuous number (house price, temperature, demand). Output is a quantity.

This is the world of regression, trees, and most of deep learning.

Unsupervised learning — finding structure without labels

You have inputs but no labels; the model finds structure on its own:

  • Clustering — group similar points (customer segments, topic discovery) — k-means.
  • Dimensionality reduction — compress many features into fewer while keeping the signal (PCA).

Self-supervised learning — labels from the data itself

A hugely important modern middle ground: the model generates its own labels from unlabeled data by hiding part of the input and predicting it. This is how LLMs are pretrained — “predict the next token” is a label the text provides for free (how LLMs work). It unlocked training on internet-scale unlabeled data. Semi-supervised learning (a little labeled data + a lot of unlabeled) is the related cousin.

Reinforcement learning — learning from reward

No fixed dataset of answers at all — an agent takes actions in an environment and receives a reward signal, adjusting behavior to maximize cumulative reward over time. Fundamentally different from the above. Game-playing (AlphaGo), robotics, and — notably — the RLHF step that makes LLMs follow instructions. Covered in reinforcement learning.

The ML workflow

Whatever the type, real ML work follows a loop that’s mostly not modeling:

frame the problem → get & clean data → engineer features → split data
   → train a model → evaluate honestly → tune → deploy → monitor → (retrain)

The unglamorous truth: most of the effort is data (working with data) and honest evaluation (model evaluation), not picking a clever algorithm. And the last two steps — deploy and monitor — are the MLOps half that distinguishes an ML engineer from someone training models in a notebook.

↑ Back to 01 — Foundations of ML