The Machines That Learn

Rules vs. examples — the central shift

This is the most important idea in the entire book. If you get it, everything else follows. If you don't, nothing else makes sense.

Imagine you're trying to teach a friend to recognize a cat.

The old way (rules-based programming): you'd write a list. "Has fur. Has four legs. Has whiskers. Says 'meow.' Tail under 14 inches. Eyes round and forward-facing…" But every rule has an exception. Hairless cats. Three-legged cats. Kittens that haven't learned to meow. You'd be writing rules forever and you'd still get it wrong.

The new way (machine learning): you show your friend ten thousand pictures of cats and ten thousand pictures of dogs, and just say "cat" or "dog" each time. After a while, they can tell new cats from new dogs without you ever having written down a single rule. They've learned the pattern from examples.

That second way is machine learning. The whole field is just figuring out clever ways to do that "show examples and let it figure out the pattern" trick on harder and harder problems.

Where you've already seen ML

ML isn't some future thing. It's already running your life:

You've been training ML systems with your behavior for years. Every time you skipped a song, watched a video to the end, or scrolled past a post, that was a labeled example feeding a model.

Three flavors of ML

You'll see these names again and again. For now, just plant the flag — we unpack each one in the next chapter.

• Supervised learning. Learning from labeled examples ("this is a cat, this is a dog").
• Unsupervised learning. Finding patterns in unlabeled data.
• Reinforcement learning. Learning by trial and error with rewards.

Your first code — hello, Colab

Open colab.research.google.com. Sign in with a Google account. Click "New notebook."

In the first cell, type:

print("Hello, ML.")

Hit Shift+Enter. The output appears below the cell.

Add a new cell. Now you'll do something slightly more ML-shaped: load a real dataset and look at it.

# Import pandas. Pandas is the standard library for working with
# spreadsheet-like data in Python. We rename it to "pd" because
# that's what every Python ML tutorial does.
import pandas as pd

# Load the famous iris dataset. This is a built-in dataset of flower
# measurements, often used as the "hello world" of machine learning.
from sklearn.datasets import load_iris
iris = load_iris(as_frame=True)
df = iris.frame

# Show the first 5 rows.
df.head()

Run it. You'll see a small table: 5 rows, 5 columns. Sepal length, sepal width, petal length, petal width, target (the species).

That table is a dataset. Every row is one example (one flower). Every column is one feature (one measurement). The target column is the label — which species the flower is. By the end of Phase 3, you will train a model on this exact dataset that can identify a flower's species from its measurements with 95+% accuracy.

Run one more cell to see why the dataset is so well-loved:

import matplotlib.pyplot as plt

# Scatter plot: petal length vs petal width, colored by species.
plt.scatter(df['petal length (cm)'], df['petal width (cm)'],
            c=df['target'], cmap='viridis')
plt.xlabel('petal length (cm)')
plt.ylabel('petal width (cm)')
plt.title('Iris dataset: three species are visibly separable')
plt.show()

The output shows three colored clusters in a 2D plot. Three different species form three visible groups, just from two measurements. Your job by the end of the year is to turn that visible separation into an algorithm a computer can use.

Vocabulary

Questions you might have