The three types of machine learning — question bank
35 standalone questions on supervised, unsupervised and reinforcement learning — and which situation you are actually in.
Audience: absolute beginners. No maths, no code, no prior exposure assumed.
Source: the session-one whiteboard — supervised (classification, regression), unsupervised (clustering, anomaly detection), and reinforcement learning.
How to use this: every question stands alone. Pick an option, then read the answer. The sections follow the five boxes on the board, then the decision path that tells you which one you are in.
A · The one question that splits all three
Q1What single question separates supervised, unsupervised and reinforcement learning?
Answer: (2)Supervised: a teacher with an answer key — every example comes with the right answer attached. Unsupervised: no teacher at all — just a pile of data. Reinforcement: no answer key but a scoreboard — nobody says what is right; you find out afterwards whether it went well. That is the whole taxonomy.
Q2Are these three different technologies?
Answer: (2)And the consequence is worth saying out loud: the situation picks the approach — you do not. If nobody recorded the right answers, no amount of preference makes it a supervised problem.
Q3Three ways to learn to cook: a chef stands beside you and corrects every dish; someone hands you 500 recipes with the titles torn off; you open a restaurant and find out weeks later from whether people come back. Which is which?
Answer: (2)The chef supplies the right answer every time. The recipe box supplies no answers at all — but you would still notice that a lot of these use tomato and basil, and that pile is all baked. The restaurant gives you only till receipts, weeks later, attached to everything you did.
B · Supervised learning and the label problem
Q4What is the most important word in supervised learning?
Answer: (2)Forty cars and what each sold for. Ten customers and whether each one cancelled. A million photos and what is in each one. Without the second half of each of those, there is nothing to learn from.
Q5Your company has twenty years of invoices. Is that enough to build a supervised model predicting problem customers?
Answer: (2)Having data is not the requirement. Having data with the answers attached is the requirement. Companies routinely discover at exactly this point that nobody ever recorded what happened next — which is the single most common project-killer in the field.
Q6Which source of labels explains why the largest platforms have such an advantage?
Answer: (3)Free, enormous, and continuous. The other two sources are real but limited: reality provides some labels free if you remember to record them, and paying humans to label is expensive, slow, and the bottleneck on most real projects.
Q7What is the test that decides whether a supervised problem is classification or regression?
Answer: (2)And the jargon on the board — "continuous" — decodes to one question: is there anything in between two answers? Between ₹10 lakh and ₹11 lakh there is ₹10.5 lakh and ₹10.51 and infinitely more. Between "spam" and "not spam" there is nothing; you cannot be 40% of the way from one to the other.
Q8"Which of five departments should this complaint go to?" What type is it?
Answer: (2)A short, countable list of possible answers. Note that clustering would be the answer to a different question — what natural groups exist among these complaints? — where nobody supplies the categories in advance.
Q9A classifier is often described as outputting "Yes" or "No." What does it actually output?
Answer: (2)And that cut-off is not a technical setting. It is a business decision — move it one way and you catch more leavers while annoying more happy customers; move it the other way and you annoy nobody while quietly losing people.
Q10Why is the word "regression" unhelpful?
Answer: (2)Telling a room that plainly buys more credibility than pretending the name makes sense. There is nothing to decode; the word simply does not describe what it names.
C · The trap: the same question can be either
Q11One customer, one dataset. "How much will they spend next year?" gives ₹47,300. "Will they spend more than ₹50,000?" gives No. What has changed between the two?
Answer: (3)Same customer, same data, same information, different type. The type is chosen by the question, not by the data. Anyone who says "our data is a classification problem" has skipped a decision that was theirs to make.
Q12Turning ₹47,300 into "under ₹50,000" throws information away. Is that a mistake?
Answer: (2)Information is genuinely lost, and that can be worth it. Ask which version you would hand a sales team on a Monday morning (most say the band), then which you would want for the annual budget (most switch to the number). The right type depends on what the answer is for.
Q13"Predict whether a loan will be repaid" and "predict how much of the loan will be repaid." Same data, same customer. What types are they?
Answer: (3)Repaid / not repaid is two answers on a menu. "How much" lands anywhere on a scale. If a room takes one thing from this topic, make it this pair — identical data, and the type changed because the question changed.
D · Unsupervised learning
Q14What makes unsupervised learning genuinely a different activity?
Answer: (2)On a supervised sheet, reality told you the answer and you learned to match it. Here nothing does. Supervised learning answers a question you asked. Unsupervised learning tells you something you did not know to ask.
Q15You cluster your customers and ask for two groups. Then you ask for five. The machine obliges both times. What does that tell you?
Answer: (2)Like sorting laundry: nobody gives you the categories, piles emerge, and different people would make slightly different piles — none of them wrong. The machine will always give you the number of groups you ask for.
Q16Clustering a retailer's shoppers reveals a group that visits twice a year and spends a fortune. Why is that valuable?
Answer: (2)The data had it all along — clustering just made somebody look. That is the business case for the whole technique: it surfaces structure that no one thought to specify.
Q17Anomaly detection versus clustering — what is the difference in what you are looking for?
Answer: (2)The red sock in the white wash. You are not sorting — you are hunting the odd one out: a transaction that does not look like you, a machine vibrating slightly wrong before it fails, a login from a country you have never visited.
E · Why fraud cannot just be classification
Q18Fraud is Yes/No. Why not take past frauds, label them, and learn what fraud looks like?
Answer: (2)Which makes it permanently one step behind — exactly the spam-filter problem. This is the first of two reasons, and both matter.
Q19If 1 in 10,000 transactions is fraud, a model that says "not fraud" every single time is how accurate — and how useful?
Answer: (2)Accuracy is a broken measure when one answer is overwhelmingly common. This is the second reason classification fails here, and it is the line that lands the whole section: a headline accuracy figure can be near-perfect and describe a model that has never once done its job.
Q20How does anomaly detection flip the question?
Answer: (2)Classification learns what the thing looks like. Anomaly detection learns what normal looks like. You need no examples of fraud at all — only lots of examples of normal, which you have, by the million, for free.
Q21What is the honest cost of that flip?
Answer: (2)Which is why anomaly systems generate so many false alarms, and why a human sits at the end of almost all of them. Be honest about this rather than overselling the technique.
F · Reinforcement learning
Q22What defines reinforcement learning?
Answer: (2)The board's parenthesis — reward and punishment — is exactly right, and it is the only vocabulary needed. A dog sits and gets a treat; nobody ever describes the rule, the dog cannot state it, and within weeks the behaviour is what you wanted.
Q23You win a chess game on move 60. Which of your sixty moves won it?
Answer: (2)The business version is instantly recognisable: a large deal closes after eight meetings, three emails and a dinner. Which one closed it? That is not a failure of record-keeping — the credit genuinely cannot be assigned from a single outcome.
Q24Friday night: your favourite restaurant, which you know is good, or the new place, which might be better and might be terrible. What is this called, and who faces it?
Answer: (2)Go to the favourite every time and you never find anything better. Try somewhere new every time and you eat a lot of bad dinners. Keep running the ad that works, or test a new one. Sell to the segment you understand, or try the one you do not.
Q25Asked what share of the budget goes to exploring, most organisations answer honestly: "almost none, and we call that discipline." What does reinforcement learning contribute here?
Answer: (2)The value is the framing, not the algorithm. It is a board-level question long before it is a technical one.
Q26Why do you rarely see reinforcement learning in business?
Answer: (2)A simulated robot can fall over a million times before lunch. A pricing algorithm cannot destroy a million real transactions to find out what works. It thrives where mistakes are cheap, fast and reversible — games, robotics, ad selection, route optimisation, data-centre cooling.
Q27A raw language model trained on a mountain of text was good at predicting the next word and not naturally good at being helpful, safe or willing to follow instructions. What fixed that?
Answer: (3)Reward and punishment, exactly as the board says, with humans holding the treats. Predicting the next word made it knowledgeable. Reinforcement learning made it useful.
Q28What is the hidden human cost inside that process?
Answer: (2)And there is a second lesson attached: reward what you actually want. A model rewarded for answers that look good learns to produce answers that look good — one honest explanation of why these systems sometimes sound confident about things they have invented.
G · Which one is your problem?
Q29Working down the decision path, which question stops most real projects?
Answer: (2)Not "do we have data" but do we have the answers, written down, for past cases? Companies routinely discover at this exact point that nobody ever recorded what happened next.
Q30"Group 50,000 customers into segments nobody has defined." Which type?
Answer: (3)The giveaway is "nobody has defined." If somebody had defined the segments and labelled past customers, it would be classification.
Q31"Find machines on the factory floor behaving unusually before they break." Which type?
Answer: (2)Same shape as spotting the one supplier invoice unlike all the others. You have millions of examples of a machine running fine and very few of it failing in the specific new way it is about to.
Q32"Work out how much to bid, in real time, for an ad slot, learning from what wins." Which type?
Answer: (3)And notice it satisfies the conditions: the decisions are cheap, fast, reversible and enormously numerous. That is precisely the terrain where reinforcement learning genuinely earns its place.
Q33"Read a customer's star rating out of five from their written review." Which type?
Answer: (3)Treat it as five labels and you cannot express 4.5. Treat it as a number and the model may output 3.7 stars, which nobody ever gave. Both are defensible and both lose something. Real problems sit on this boundary constantly, and admitting it buys more trust than a clean answer would.
Q34You reach the bottom of the decision path: no answer key, and you cannot afford to experiment. What follows?
Answer: (3)No labels, no scoreboard, no safe place to fail. That is not a machine learning problem waiting for a better algorithm; it is a signal to revisit whether a rule or a lookup already answers the question.
Q35The three-line summary. Which is correct?
Answer: (2)Answer key, nothing, scoreboard. What does the teacher look like — that single question carries the whole taxonomy, and a non-technical room can hold it for the rest of their lives.