Exam practice

• Supervised learning learns from labeled examples, to give the right label to new examples: the decision tree learning the districts’ belts from districts labeled with their belt; a model reading handwritten Devanagari digits.
• Unsupervised learning learns from examples without labels, finding structure in them: k-means finding groups of districts by height and rain (the dry high districts behind the Himalaya, the wet ones, the lowland, the middle hills); a shop finding kinds of customer.
• Reinforcement learning learns from rewards for its actions in a world: the walker learning the shortest route to the tea shop from points won and lost; AlphaGo learning Go from wins and losses.

The tree tries every question of the form “feature < cut?”, for every feature and every cut halfway between two neighboring values. For each question it splits the 77 districts into two piles and works out how mixed the belts are in each pile by Gini impurity (the chance that two districts picked from the pile have different belts), weighted by the piles’ sizes. It keeps the question with the lowest weighted impurity. For the districts that is “height < 385 m?”: one pile is 20 Terai districts (impurity 0), the other 57 mostly Hill and Mountain districts (0.428); weighted, 0.317, lower than any other question.

A depth-4 tree gets all 77 right because it adds thin strips round single odd districts (Dang, Gorkha). Those strips fit the particular districts it learned from, not a real pattern: when each district is left out in turn, depth 4 gets 69 right and depth 2 gets 70. Fitting the training examples too closely, so that the model does worse on new ones, is overfitting.

Steps: scale height and rain to between 0 and 1. Choose 4 districts at random as the first centers. Each district joins the nearest center. Each center moves to the mean of its group’s districts. Repeat the last two steps until no district changes group.

Reasons for care (any two):

• It finds as many groups as you ask for, whether or not the data really has that many natural groups.
• Different starts can end in different groups; one run may give a poor grouping (seed 3 gives a total squared distance of 2.018 instead of 1.644), so it should be run from many starts.
• It gives no names or reasons: what a group means, and whether it is useful, is a human judgment.
• It sees only the features it is given: with population density instead of rain, the groups would be different.

• a.A plain program. The rules are written down exactly; a program follows them exactly, costs nothing to run, can be checked line by line, and a wrong grade would harm a student. AI would only add mistakes.
• b.A neural network. The data is photos, which trees and rules can’t handle well; with thousands of labeled leaf photos a network can learn the signs. It can’t explain itself, but the risk is low if it advises (“this may be blight; ask the agriculture office”) rather than decides, and it can run on a phone with no internet. It must be tested on photos from Nepal’s farms, phones and light.
• c.A decision tree (or another small, readable model). The data is a table of a few features with a label (repaid or not). Applicants must be told why they were or weren’t chosen, and an officer must check, so the rules must be readable. It should be checked for each group (women and men, each ward), and the data checked for past unfairness that it would repeat; a person makes the final decision.

TP = 40, FN = 10, FP = 95, TN = 855.

• Accuracy = (40 + 855) ÷ 1,000 = 89.5%
• Precision = 40 ÷ (40 + 95) = 40 ÷ 135 = 29.6%
• Recall = 40 ÷ (40 + 10) = 80%
• F1 = 2 × 0.296 × 0.8 ÷ (0.296 + 0.8) = 0.474 ÷ 1.096 = 43.2%

Accuracy is not a good measure. Only 50 of the 1,000 have the illness, so a test saying “healthy” to everyone would be 95% accurate, better than this test, while finding nobody. Precision shows that fewer than a third of positive results are real, so positives need a second test; recall shows that 1 in 5 ill people is missed, which may matter more for a serious illness.

A parameter is learned by the model from the training data: a tree’s questions and cut-offs (385 m), a neural network’s weights, the centers k-means ends with, a Q-learner’s Q-values. A hyperparameter is chosen by the builder before training and controls how the model learns: a tree’s depth, k in k-means or nearest neighbors, the learning rate, the number of hidden neurons, how much a learner explores.

If the hyperparameter is chosen by the test set’s score, the test set has been used to build the model, so its score is no longer a fair measure of how the model will do on new data: it will be too hopeful. Hyperparameters are chosen on a separate validation set (or by leaving examples out in turn), and the test set is used once at the end.

How: the model learned from the messages the company had, which were mostly in English, so it had seen few Nepali scams and learned few of their signs. Its overall recall hid the gap, because English messages were the majority.

Better data: collect and label many more Nepali messages (scams and normal), with permission, and retrain. This fixes the cause: in the unit’s example the retrained model caught 13 of 15 scams in each language with few false alarms. It costs time and money.

A lower threshold for Nepali messages: quick, but the model’s Nepali scores are poor, so catching more scams brings many more false alarms (at 0.30, 9 of 15 scams caught but 39 normal messages flagged, so only 19% of its Nepali flags are real). Treating groups differently also needs a clear reason. It may help for a while, but better data is the real fix.

An AI agent is a system in which a model, given a goal, plans steps and carries them out by calling tools (programs it may use: search, payment, messages, reading pages), reading each result and deciding the next step until it thinks the goal is done.

Failures (any three):

• It does what was said, not what was meant: asked for the cheapest bus, it booked a night bus arriving at 4:30 am. Limit: it asks the user when a choice is unclear, or shows its plan before acting.
• Prompt injection: a web page told “AI assistants” to book 3 seats and text the tickets to a number, and it obeyed. Limit: a person approves every payment and message; the tool books one seat per request; it may text only approved numbers.
• Loops: the bus was full, and it kept retrying. Limit: a maximum number of tool calls (and of time and money) per run.
• Too much power: a coding agent deleted a live database during a code freeze. Limit: it gets no tool that can touch the live database; it works on a copy; a person approves changes.

A good plan covers ordinary use, the edges, both languages, groups of people and harms. For example:

Each case is run and the result recorded with the date and the chatbot’s version, and run again after each change.

Model card:

• Purpose: to suggest households for officers to visit for food support; not to refuse support by itself.
• Data: where the household data came from, when, who collected it, with permission, and which wards and groups are missing or few.
• How well it works: precision and recall on a test set not used in building it, with the threshold used.
• By group: the same numbers for each ward, for households headed by women and men, and for different castes or ethnic groups where that data was collected with permission.
• Limits: whom and what it may misjudge (new families, households who moved, seasonal work abroad).
• Values and choices: which mistake was treated as worse (missing a hungry household), and why the threshold was set where it was.

Safeguards (any three): officers visit and decide, the model only suggests; households are told how to ask for a visit if they were not suggested (appeal); the data is kept private and only what is needed is collected; the model is measured each season for each ward and retrained or stopped if it drifts; the old way of choosing households is kept ready.

Not necessarily. A written rule is explainable, which makes its unfairness easier to see and to challenge, but it can still be unfair, because it carries the values of the person who wrote it. For example, a cooperative rule “loans only to applicants with land in their own name” is exact and readable, but shuts out most women where land is usually registered in men’s names. A model learned from past decisions can repeat past unfairness too. Either way, a rule or a model must be checked for whom it fails, for each group it affects, and the values in its design must be stated and questioned.