What machine learning is

In ordinary programming, you write the rules: if marks >= 40: print("pass"). In machine learning (ML), the computer finds the rules itself from examples.

ordinary programming:   rules + data      →  answers
machine learning:       data + answers    →  rules (a "model")

That helps when the rules are too many or too subtle to write by hand: recognising a face, translating a sentence, guessing tomorrow's electricity demand, spotting a fraudulent card payment.

A first example

# Hours studied, and whether the student passed (1) or not (0)
hours  = [1, 2, 3, 4, 5, 6, 7, 8]
passed = [0, 0, 0, 1, 0, 1, 1, 1]

best_cutoff, best_correct = None, -1
for cutoff in range(1, 10):                   # training: try every rule "pass if hours >= cutoff"
    correct = sum(1 for h, p in zip(hours, passed) if (h >= cutoff) == (p == 1))
    if correct > best_correct:
        best_cutoff, best_correct = cutoff, correct

print(f"learned: pass if hours >= {best_cutoff} ({best_correct} of {len(hours)} right)")

for h in [2.5, 4.5, 9]:                       # prediction, for new students
    print(h, "hours ->", "pass" if h >= best_cutoff else "fail")

The program has the hours eight students studied and whether they passed. It does not know the rule. It tries every possible cut-off and keeps the one that gets most students right: that is training. Then it uses the learned cut-off on students it has never seen: that is prediction. Every ML method in this course is a smarter version of the same two steps.

The words

Three kinds of learning

• Supervised learning: every example has its label. If the label is a category (spam or not, which digit), it is classification; if it is a number (a price, a temperature), it is regression. Most of this course is supervised.
• Unsupervised learning: no labels. The model finds structure by itself, such as clusters of similar customers.
• Reinforcement learning: an agent acts, gets rewards or penalties, and learns which actions pay off: how game-playing programs and robots learn.

When not to use ML

ML needs data, lots of it and of good quality, and it gives answers that are usually right, not always. If you can write the rule (a tax calculation, a sort), write it. ML also learns whatever is in the data, including its mistakes and unfairness; the last lesson of this course comes back to that.

Mark lesson complete