Module Check
This lesson consolidates everything in the module. You have traversed two major paradigms that extend machine learning beyond labeled data: unsupervised learning (clustering, dimensionality reduction, anomaly detection) and reinforcement learning (agents, MDPs, policies, exploration). The questions and tasks here are designed to test not just recall but genuine understanding — the ability to reason about new situations using the concepts you have built.
Core Concept Review
Flashcards — click each card to reveal the answer
Reasoning Questions
A researcher runs k-means with k=4 on a dataset and finds that two of the four clusters have only 3 points each, while the other two have thousands. What does this most likely indicate?
A PCA analysis of 500 features reveals that the first 8 components explain 91% of variance. A colleague says 'we can therefore ignore 492 features entirely.' What is wrong with this statement?
An anomaly detection system trained on 2023 financial transaction data is deployed in 2025. Its false-positive rate has risen sharply. What is the most likely cause?
An RL agent trained with γ = 0.99 and one trained with γ = 0.3 are both placed in a maze where the exit gives +100 reward but requires passing through a corridor that gives −1 per step. Which agent is more likely to find the exit?
A team uses RLHF to fine-tune a language model and finds that the model learned to write extremely long, verbose responses because human raters rated longer responses higher. What RL concept does this exemplify?
Why might dimensionality reduction improve clustering quality, even though it discards some information?
Supervised learning is powerful but limited by the cost and availability of labels. Unsupervised learning unlocks the structure latent in the world's vast stores of unlabeled data — finding groups, compressing representations, flagging the unusual. Reinforcement learning moves beyond pattern recognition entirely, asking how an agent can discover what to do through interaction with a world that only offers delayed, evaluative feedback. Together, these three paradigms cover nearly the full space of what learning from data can mean.
Capstone — Design and Defend a Complete ML System
- You are a machine learning engineer advising a city transit authority. The authority has:
- - 3 years of GPS traces from 500 buses (position every 30 seconds)
- - Ridership counts from sensors at each stop (but no records of which bus stop a rider boards)
- - Weather logs synchronized by timestamp
- - No labels defining 'on-time,' 'delayed,' or 'route inefficient'
- Part A — Unsupervised analysis:
- 1. Propose how you would use clustering to identify route segments that behave similarly across time. What features would you build from GPS traces? How would you validate the resulting clusters without ground-truth labels?
- 2. Propose how dimensionality reduction could help visualize variation across the 500 bus routes.
- 3. Propose how anomaly detection could flag unusual days or route segments for human review.
- Part B — Reinforcement learning proposal:
- 4. The authority wants an AI to dynamically adjust bus dispatch times to minimize average passenger wait time. Cast this as an MDP: define S, A, R, and γ. Justify each choice.
- 5. Identify the two biggest obstacles to deploying this RL system in real operations, and propose one mitigation for each.
- Part C — Synthesis:
- 6. A city council member asks: 'Could we just use one algorithm for all of this?' Write a 150-word response explaining why the problems above are best addressed by different paradigms, and why that is a feature of good system design, not a failure to find a unified solution.