The Machine Learning Landscape

Question 1

Géron classifies a spam filter as which combination?

Accepted Answer

Online, model-based and supervised — The three axes are independent, so every system gets one label from each.

Question 2

The difference between batch and online learning is whether the system:

Accepted Answer

Must be retrained on all the data, or can learn incrementally — A naming trap the book flags: mini-batches are often just called batches, so "batch learning" is a confusing name for the non-incremental kind.

Question 3

Instance-based learning generalises by:

Accepted Answer

Comparing new data to remembered instances by similarity — k-nearest neighbours is the pure case — it builds no model at all.

Question 4

The key parameter of an online learning system is the learning rate, which controls:

Accepted Answer

How fast the system adapts to new data, and how fast it forgets — A high rate adapts quickly, forgets old data quickly, and is more sensitive to nonrepresentative outliers.

Question 5

The main risk Géron identifies with online learning is that:

Accepted Answer

Bad data fed to the system will degrade performance — Out-of-core learning is actually one of online learning's advantages, not a limitation.

Question 6

Why does Géron argue self-supervised learning deserves its own category?

Accepted Answer

Because it targets the same tasks as supervised learning, unlike unsupervised learning — Unsupervised learning focuses on clustering, dimensionality reduction and anomaly detection. Self-supervised learning generates labels from unlabelled data to do supervised-style tasks.

Question 7

Géron notes that a model trained by self-supervised learning is:

Accepted Answer

More often than not, not the final goal — It is typically a step toward a downstream task — an image-repairing model as a route to something else.

Question 8

Sampling noise differs from sampling bias in that noise:

Accepted Answer

Shrinks as the sample grows, whereas bias does not — That is exactly why a very large sample can still be worthless: it fixes the part that was not the problem.

Question 9

The 1936 Literary Digest poll mailed about 10 million people and received:

Accepted Answer

2.4 million answers — A 24% response rate — an enormous sample by any standard, and still wrong.

Question 10

That poll predicted Landon at 57%. The actual result was:

Accepted Answer

Roosevelt won with 62% — A 19-point error from a 2.4-million-person sample, caused entirely by how the sample was drawn.

Three axes, one system

Ten million people, and still wrong

What kind of learning is this?

Test your knowledge

Géron classifies a spam filter as which combination?

Interview Questions

References & Further Reading

Love this visualization?