Streaming & Online ML Quiz

Streaming & Online ML Quiz

Check your grasp of learning on unbounded data — the one-pass model, sublinear sketches, regret and online gradient descent, drift detection, and the production pipeline that ties them together.

Exercise

Which set of constraints defines the classic data-stream model?

Need a hint?

Think about how many passes you get and how much memory relative to the stream length n.

Exercise

You need to estimate the number of DISTINCT users in a high-volume stream using only logarithmic memory. Which sketch is designed for this?

Need a hint?

Which algorithm tracks the maximum number of trailing zeros in hashed values?

Exercise

If every online loss is additionally λ-strongly convex, how does online gradient descent's regret bound improve over the general convex case?

Need a hint?

Curvature helps. The bound drops from a power of T to something much slower-growing.

Exercise

A Hoeffding tree splits a leaf once the best attribute's information gain beats the runner-up by more than ε = √( R²·ln(1/δ) / (2n) ). What does this bound let the tree do?

Need a hint?

n is the number of examples seen at the leaf. What does waiting for the gap to exceed ε guarantee?

Exercise

In a fraud system, the chargeback label for a transaction arrives 60 days after the prediction. What is the correct way to train an online model on such delayed labels?

Need a hint?

Which feature values should the eventual label be attached to — today's, or the ones the model actually saw?

Exercise

On a live stream, a drift detector watches for a change in the data-generating process and triggers the model to adapt. Which slot(s) of the ML project loop does this exercise?

Need a hint?

It's the loop's close-back-to-data arrow, running continuously.

Exercise

Transfer test — a technique this course did NOT teach: a residual (skip) connection that adds a layer's input to its output so gradients flow through deep networks. Which slot does it modify?

Need a hint?

It's an architectural building block that changes the model, easing optimization as a side effect.