Machine Learning

CT-AI · 53 questions

• A cheese-cave shop has three pitches: shelves already tagged “ready / wait,” a heap of unlabeled aging logs to be grouped by similarity, and a robot cart that is praised or fined after each aisle choice. How should those three be assigned?
• The same cave wants one model that stamps each wheel “sell / hold / discard” and another that estimates days-to-peak in a decimal. Both use labeled history. How should that supervised work be split?
• A test lead hears “we need regression on the mash-temperature predictor” and books a full re-run of last sprint’s functional suite. The data scientist meant a continuous-value supervised model. What distinction should the tester keep?
• A mountain-hut shop has unlabeled till rolls. One request is “put guests into groups that look alike.” The other is “which snack tends to appear with which fuel canister.” How should those unsupervised requests be assigned?
• A rope-tow start gate has no labeled “correct start” file. A cart learns by inching forward, receiving a plus when skiers load smoothly and a minus when the line bunches. Which ML form is that?
• The rope-tow cart is rewarded for “gate cycles per hour.” It learns to slam the gate whether or not a skier is ready. What RL challenge should a tester understand?
• A seed-bank desk has labeled germination outcomes, a drawer of unlabeled packet photos, and a robot arm that can only be scored after it places a packet. How should the tester pick the ML form?
• A paper mill wants a tear-class stamp (labeled), a grouping of unlabeled vibration weeks, and a reel-cart that is scored after each dock. What should the tester refuse?
• A beekeeper co-op argues for three weeks about whether the model exists to flag mite risk or to forecast honey weight, and only then writes the ML functional-performance numbers the model must meet. Which workflow activity is that?
• The co-op picks a development toolkit from the job and the acceptance bar, then picks an algorithm (from a library, not hand-coded) that matches the labeled mite sheets they actually have. Which two workflow steps are those?
• Before any fit, the mite sheets are acquired, cleaned, and featured, and the prep steps themselves are checked. A volunteer wants to “just train and see.” Which workflow activity is required?
• One setting is “how many layers / how deep is the tree” (structure passed into the algorithm). Another is “how many passes over the mite sheets” (how training itself is run). How should a tester name those knobs?
• The co-op fits a tree, a margin classifier, and a small net on the same training slice, then scores them on a validation slice against the agreed metrics and keeps the best pair of algorithm plus training data. Which workflow step is that comparison?
• After the validation scores arrive, the team nudges layer count and pass count, retrains, and scores again. They keep calling that loop “the test.” Where should a tester place that loop?
• Once a model looks good enough, a slice that never guided a nudge is used to check the agreed ML functional-performance criteria and is compared with the validation numbers. If the held-out score is much worse, the team returns to model generation or even to data prep. Which workflow step is that?
• The chosen mite model is re-engineered with its data pipe for a hut kiosk and a small cloud API. Acceptance is checked again on that packed form. Which workflow step is that?
• Once live, the model either scores a nightly batch of yard sheets or answers a ranger on demand. It sits inside a larger yard system, so conventional integration, system, and acceptance still happen. Over the season its scores may walk away from the agreed bar, so the team re-evaluates and may rebuild. Which closing workflow should the tester summarize?
• A herbarium cannot afford to train a leaf-net from empty weights. It starts from a network already fitted on a broad plant-image corpus, then spends a much smaller labeled set on its own sheets. What alternative should the tester summarize?
• The herbarium can extra-train the whole net, only the layers near the output, or a few added layers, then re-check ML functional performance and extra-train again until acceptance is met. What should a tester understand about those choices?
• Adapting a “maple-leaf vs oak-leaf” net to “beech vs birch” is likely to work. Asking that same net to score recorded bird-call accents is not. What should the tester summarize?
• A clock-tower help desk keeps a binder of winding procedures. Those pages are turned into a searchable store, the closest pages are stuffed into an enlarged prompt, and the language model’s weights are never changed. Which approach is that?
• The tower desk both extra-trains a small language model on last decade’s logbooks and still retrieves tonight’s weather bulletin into the prompt. What should the tester know?
• The bought leaf-net already favored glossy studio photos and failed on frost-bitten sheets. After fine-tuning on the herbarium’s own trays, those habits are still there. What risk should the tester summarize?
• A soap factory’s scent-batch project must first decide it needs numeric vat logs plus a few still photos, pull them from the historian and a bench camera, and have two chemists tag “in spec / off.” Which data-preparation activity is that?
• The vat log has smashed rows, the same batch twice, a one-off 900 °C spike, empty pH cells, and a column of worker names. Which preprocessing activities map to those defects?
• One vat reports temperature in a 0–1 scaled column; another still stores raw °C. A third dumps a free-text note where the model needs a fixed-width row. Which preprocessing work is that?
• The factory has too few off-spec batches, so it adds rotated stills, a few synthetic off-spec rows, and some deliberately noisy frames so the later model is less brittle. Which data-prep activities are those?
• The historian holds ten years of vats. Compute and calendar only allow a representative subset this sprint. Which preprocessing activity is that?
• Twenty sensor columns exist; three of them actually move the “in spec” call, and two others are near-duplicates of those three. The team keeps a smaller, informative set. Which activity is that?
• Beside the cleaning work, a chemist sketches weekly pH clouds and spots a week when the probe stuck high. Which parallel activity is that?
• The soap team reorders cleaning and feature steps twice, skips anonymization on a lab-only trial, then insists the live vats must show the same distributions and ranges as the training vats — even if the live pipe is simpler for speed. What should the tester require?
• A yarn-dye house randomly draws three equivalent slices from one representative lot: one to fit the shade model, one to score and nudge it, one held back as the final judge. How should those slices be contrasted?
• The dye-house slices were all drawn from winter lots. Spring lots are more pastel. What must be true of the training, validation, and test slices?
• With plenty of lots, the dye house still must decide how much data each slice gets. Which factors should drive those volumes rather than a single recited percentage?
• Only a short season of lots exists. Cutting it three ways leaves too little to fit, and the shade model underfits. What risk and response should the tester recognize?
• The remaining pool is cut into k parts (often 5 or 10). Each part is the validation slice once while the others train; scores are averaged; the winning hyperparameters then refit on the whole pool and are judged once on the holdout. Folds are usually random and often stratified when classes are thin. How does that differ from a single static validation slice?
• A second dye house is so short on lots that it skips the holdout and quotes the average fold score as the official number. What should the tester know?
• A kiln-crack report prints a 2×2 of predicted versus actual “crack / sound.” One vendor swaps which axis is predicted. What must the tester still recover from either layout?
• A kiln-crack matrix counts 42 true cracks called crack, 48 true sound tiles called sound, 7 sound tiles called crack, and 3 true cracks called sound. What is accuracy as (TP + TN) / all × 100%?
• A candle-wick inspector’s false “defect” calls stop the line. The matrix is 21 true defects called defect, 7 sound wicks called defect, 2 missed defects, and 70 sound wicks called sound. What is precision, and what does it mean?
• A cave-bat-count model that misses an actual roost night is the expensive error. The matrix is 15 true roosts called roost, 5 missed roosts, 4 false roost calls, and 76 true quiet nights called quiet. What is recall (sensitivity), and what does it mean?
• A tea-blend grader posts precision of 90% and recall of 60%. What is F1 as 2 × (precision × recall) / (precision + recall), and why is it not the arithmetic mean?
• A glass-studio “bubble / no-bubble” set is 190 no-bubble tiles and 10 bubble tiles. A model calls every tile no-bubble: 0 true bubbles found, 10 bubbles missed, 0 false bubble calls, 190 true no-bubble calls. Why does 95% accuracy mislead, and which named metric should the tester also compute?
• A print-shop registration model never hits 100% on accuracy, precision, recall, or F1. Stakeholders treat that as a failed build. What should the tester explain?
• Two bicycle-share “needs-service” models each score 100 bikes. Model A: 20 true positives, 15 false alarms, 5 missed bad bikes, 60 true negatives. Model B: 20 true positives, 5 false alarms, 15 missed bad bikes, 60 true negatives. Missed bad bikes staying in service is the costly error. Which model matches that cost story, and why?
• A soapbox-derby “legal / illegal ballast” demo uses a single-layer perceptron that can separate the two classes only if a straight line in feature space can do it. How should a tester place that structure?
• A weather-balloon photo hits an input layer of pixel-like values, passes through one or more hidden layers of nodes, and leaves an output layer as an “icing-risk” score. In a fully connected net, each node talks to every node in the next layer, and layer widths may differ. What structure should the tester name?
• Each hidden or output node forms a weighted sum of the previous layer’s activations, adds its own bias term, and pushes the total through a non-linear activation function. Different functions give different activations. What per-neuron computation should the tester recount?
• Weights start small and random (biases often at zero). A balloon photo goes forward, the output is compared with the known icing label, the error (loss) is sent back to nudge weights and biases, and one full pass through the training set is an epoch. Training stops when the output is good enough. What working cycle is that?
• A fairness reviewer hears “bias” on a slide about hidden nodes and opens the Chapter 5 bias-test checklist. What distinction should the tester keep?
• A tile-factory net has many nodes that never produced an activation above the tester’s chosen cut during the test set. Which coverage measure is that, and what does a silent region suggest?
• Each node’s possible output range is cut into k sections; testers ask what fraction of those sections fired on the test set. Which coverage measure is that?
• During test, some nodes fire higher than any training maximum or lower than any training minimum. Which coverage measure is that, and what caution should the tester remember?