How does machine learning differ from a fixed rule-based program when handling examples it has not seen?
A Follows only fixed steps
B Stores every input permanently
C Learns patterns from data
D Needs a human each time
Answer: C. Learns patterns from data. Machine-learning models use examples to learn patterns that can inform predictions on new inputs. Fixed rule-based programs follow rules written explicitly by people.
Within common AI terminology, machine learning is best understood as what?
A A subset of AI
B A replacement for AI
C A branch of robotics
D A type of database
Answer: A. A subset of AI. Machine learning is one approach within artificial intelligence. AI also includes other methods, such as search, planning, and rule-based reasoning.
What architectural feature is most associated with deep learning?
A One shallow decision tree
B A fixed lookup table
C A hand-coded rule list
D Many-layer neural networks
Answer: D. Many-layer neural networks. Deep learning uses neural networks with multiple layers to learn increasingly complex representations from data.
Speech recognition is commonly labelled narrow AI for which reason?
A It reasons like humans
B Designed for a limited task
C It requires a robot
D It learns every task
Answer: B. Designed for a limited task. Narrow AI systems are built for bounded tasks, such as speech recognition. That capability does not imply general human-like intelligence.
Once a trained model receives a fresh input and returns a prediction, which phase is underway?
A Training
B Inference
C Data labeling
D Model deployment
Answer: B. Inference. Inference is the use of a trained model to produce an output from input data. Training is the earlier process of fitting model parameters.
During supervised learning, what information accompanies each training example?
A Random rewards only
B Unlabeled clusters
C Predicted outputs only
D Known target outcomes
Answer: D. Known target outcomes. Supervised examples include input data paired with target labels or values, which provide the outcomes the model learns to predict.
Without target labels, what result is clustering designed to produce?
A Groups of similar data
B Sorted records by name
C Rules for each label
D Exact known outcomes
Answer: A. Groups of similar data. Clustering is an unsupervised learning method that groups examples by similarity without requiring known target labels.
Suppose an agent improves by trying actions and receiving rewards; which learning approach fits?
A Supervised learning
B Unsupervised learning
C Reinforcement learning
D Transfer learning
Answer: C. Reinforcement learning. Reinforcement learning trains an agent through interactions, using rewards or penalties to shape action choices.
When the desired output is a label such as spam or not spam, what task is being solved?
A Regression
B Clustering
C Dimensionality reduction
D Classification
Answer: D. Classification. Classification predicts a discrete category or class, such as whether a message is spam.
For predicting a continuous value such as delivery time, which modeling task is generally suited?
A Classification
B Clustering
C Regression
D Association mining
Answer: C. Regression. Regression models estimate numerical outcomes. Delivery time is a quantity, while classification selects among categories.
In a typical data table, what does one example usually correspond to?
A One model parameter
B One observation or record
C One evaluation metric
D One algorithm family
Answer: B. One observation or record. A dataset row commonly represents one observation, case, or record; its columns hold features and possibly a target.
Which term describes an input measurement supplied to a predictive model?
A Feature
B Label
C Epoch
D Reward
Answer: A. Feature. A feature is an input variable used by a model. A label is the target outcome in supervised learning.
For a supervised house-price model, which field is the target during training?
A Observed house price
B Number of bedrooms
C Building age
D Distance to transit
Answer: A. Observed house price. The target is the value the model is trained to predict. In this example, the observed house price is the target; the other fields can be features.
Why do practitioners separate training, validation, and test data?
A Make datasets larger
B Remove every noisy value
C Limit biased performance estimates
D Guarantee perfect predictions
Answer: C. Limit biased performance estimates. Separate splits support different roles: fitting, model choice, and final evaluation. Reusing evaluation data during development can make reported performance overly optimistic.
During model development, which split usually guides hyperparameter choices?
A Training set only
B Validation set
C Final test set
D Production data
Answer: B. Validation set. The validation set helps compare model settings during development. The held-out test set is kept for a final evaluation.
What should a properly held-out test set estimate?
A Training speed only
B Labeling cost only
C Memory use only
D Performance on unseen data
Answer: D. Performance on unseen data. A representative test set estimates how well a model may generalize to examples not used during training or tuning.
Strong performance on unseen cases signals which important capability?
A Memorization
B Data leakage
C Generalization
D Underfitting
Answer: C. Generalization. Generalization means a model performs usefully on new examples from the intended task, rather than only reproducing its training data.
Suppose a model memorizes training examples but fails on fresh cases; which problem does this suggest?
A Underfitting
B Data normalization
C Class balancing
D Overfitting
Answer: D. Overfitting. Overfitting occurs when a model fits details or noise in its training data too closely and performs poorly on new cases.
Which pattern points to underfitting rather than overfitting?
A Poor training and test results
B Great training, poor test
C Great results on both
D Perfect test, weak training
Answer: A. Poor training and test results. An underfit model has not captured the important structure, so it often performs poorly on both training and evaluation data.
If future-only information leaks into training features, what evaluation problem can result?
A Feature scaling
B Data leakage
C Label smoothing
D Model compression
Answer: B. Data leakage. Data leakage occurs when training or evaluation includes information that would not legitimately be available at prediction time, creating misleading results.
Why can class imbalance make raw accuracy misleading?
A Accuracy ignores all predictions
B Minority labels disappear
C Models cannot use labels
D Majority guesses inflate accuracy
Answer: D. Majority guesses inflate accuracy. When one class dominates, always predicting that class may yield high accuracy while missing most examples of a rare but important class.
Among cases predicted positive, which metric measures the share that is truly positive?
A Recall
B Precision
C Specificity
D Accuracy
Answer: B. Precision. Precision is true positives divided by all predicted positives. It describes how many positive predictions are correct.
Of all actual positive cases, which metric tracks the share correctly found?
A Precision
B Specificity
C Recall
D Mean squared error
Answer: C. Recall. Recall is true positives divided by actual positives. It measures how much of the positive class the model identifies.
Which evaluation table separates true positives, false positives, true negatives, and false negatives?
A Confusion matrix
B Feature matrix
C Training log
D Learning curve
Answer: A. Confusion matrix. A confusion matrix compares predicted and actual classes, showing counts for true and false positives and negatives.
How is simple classification accuracy calculated?
A False positives divided by total
B Correct predictions divided by total
C Positive predictions divided by negatives
D Recall multiplied by precision
Answer: B. Correct predictions divided by total. Accuracy is the proportion of all evaluated predictions that match the true labels: correct predictions divided by total predictions.
What are learned numerical values inside many predictive models called?
A Parameters or weights
B Categories or labels
C Epochs or batches
D Features or records
Answer: A. Parameters or weights. Parameters, often called weights, are values learned during training. They are distinct from features and hyperparameters.
Setting a learning rate before training makes it which kind of value?
A Model label
B Training example
C Evaluation score
D Hyperparameter
Answer: D. Hyperparameter. A learning rate is selected as part of the training configuration rather than learned directly as an ordinary model weight.
In neural-network training, what is one complete pass through training examples called?
A Feature
B Token
C Epoch
D Inference
Answer: C. Epoch. An epoch is one full pass through the training dataset. Training commonly uses multiple epochs.
While updating model parameters, what does gradient descent aim to reduce?
A A loss function
B Number of input features
C Dataset size
D Class count
Answer: A. A loss function. Gradient descent adjusts parameters in directions intended to reduce a defined loss function, which measures prediction error.
Text tokenization breaks an input into what for a language model?
A Complete sentences only
B Database columns
C Image pixels
D Model-specific text units
Answer: D. Model-specific text units. Tokenizers split text into units chosen by a model's vocabulary. These units may be words, word pieces, characters, or other segments.
Many autoregressive language models generate a response by estimating which quantity next?
A Future user intentions
B Database row counts
C Next-token probabilities
D Image sensor values
Answer: C. Next-token probabilities. Autoregressive language models repeatedly estimate a distribution over possible next tokens, then select or sample a token to continue the sequence.
The phrase context window refers to which limit in a language model?
A Number of model developers
B Input tokens model can consider
C Number of training datasets
D Files stored by the browser
Answer: B. Input tokens model can consider. A context window is the amount of tokenized input and conversation a model can process at one time; the exact limit depends on the model.
In a request to a generative model, what term covers instructions and supplied context?
A Parameter
B Epoch
C Prompt
D Confusion matrix
Answer: C. Prompt. A prompt is the input provided to a model, which may include instructions, questions, examples, and source material.
Raising sampling temperature generally causes what change in generated answers?
A Guaranteed factual accuracy
B More output variability
C Shorter context windows
D Fewer model parameters
Answer: B. More output variability. For sampling-based generation, a higher temperature generally makes token selection less concentrated and outputs more variable. It does not ensure correctness.
When a chatbot confidently states an invented fact, what failure is this often called?
A Overfitting
B Data labeling
C Gradient descent
D Hallucination
Answer: D. Hallucination. A hallucination is generated content that is unsupported or false but may be presented fluently or confidently. Important claims still need verification.
Generative models differ from simple classifiers because they are built to do what?
A Create new content
B Only assign fixed labels
C Only sort stored records
D Only calculate accuracy
Answer: A. Create new content. Generative models produce content, such as text, images, or audio. A classifier instead assigns an input to one or more categories.
Why might a model's token count differ from its word count?
A Every word equals one token
B Words may split into pieces
C Tokens count only sentences
D Punctuation is always ignored
Answer: B. Words may split into pieces. A tokenizer may represent one word with several tokens, or group punctuation and word pieces in its own way. Tokens are not universally equal to words.
Within a transformer, what does self-attention help the model do?
A Increase dataset labels
B Store files permanently
C Replace all training data
D Relate tokens within context
Answer: D. Relate tokens within context. Self-attention lets a transformer weigh relationships among tokens in the current sequence, helping its representations reflect context.
Before trusting an important factual claim from generated text, how should a reader check it?
A Verify reliable primary sources
B Accept fluent wording
C Check only the headline
D Assume confidence means accuracy
Answer: A. Verify reliable primary sources. Generated language can be fluent yet wrong. Checking authoritative primary sources helps confirm consequential factual claims.
Retrieval-augmented generation supplies which aid when producing an answer?
A Extra model parameters
B A new neural network
C Retrieved external information
D A longer training epoch
Answer: C. Retrieved external information. Retrieval-augmented generation searches a knowledge source and places relevant retrieved material into the model's context at answer time.
What is the name for adapting a pretrained model with task examples while updating its weights?
A Tokenization
B Clustering
C Prompt formatting
D Fine-tuning
Answer: D. Fine-tuning. Fine-tuning continues training a pretrained model on selected examples, updating model parameters to adapt behavior to a task or domain.
Unlike machine learning, rule-based automation depends mainly on what?
A Explicit programmed rules
B Learned patterns from examples
C Unlabeled data groups
D Reward-based exploration
Answer: A. Explicit programmed rules. Rule-based automation follows logic written explicitly by people. Machine-learning systems infer patterns from data, although real systems may combine both.
Unequal error rates across groups can arise when a model learns from what?
A A balanced test process
B Unrepresentative training data
C Clear evaluation metrics
D Reviewed decision rules
Answer: B. Unrepresentative training data. Unrepresentative data can encode historical gaps or imbalances, contributing to different errors across groups. Evaluation and context also matter.
Before entering personal information into an AI service, what should a user review?
A Model color settings
B Screen resolution
C Data handling terms
D Keyboard layout
Answer: C. Data handling terms. Users should check how a service collects, stores, uses, and shares submitted data, especially when it includes personal or sensitive information.
Who should review AI assistance for a high-stakes medical or financial decision?
A A qualified human
B An unrelated chatbot
C An anonymous dataset
D A browser extension
Answer: A. A qualified human. Human expertise and accountability remain important for consequential decisions. AI output can inform review but should not replace qualified judgment.
Explainability aims to make which part of a model's behavior clearer?
A How to erase all data
B Which brand built hardware
C Why it produced an output
D How to add more labels
Answer: C. Why it produced an output. Explainability methods help people understand factors or processes behind model outputs, though an explanation may be limited or approximate.
Beyond average accuracy, what should a responsible safety evaluation examine?
A Font choice and layout
B Only training duration
C The model's file name
D Failure cases and harms
Answer: D. Failure cases and harms. Safety evaluation should examine failures, foreseeable misuse, and potential harms in relevant contexts, not only average benchmark performance.
Robustness checks reveal how performance changes after what?
A A longer product name
B Changed or noisy inputs
C More website colors
D A different file folder
Answer: B. Changed or noisy inputs. Robustness checks assess whether model behavior remains dependable when inputs vary, contain noise, or differ from ideal test examples.
Which system can accept text and image inputs in one task?
A Text-only parser
B Single-label classifier
C Rule-only calculator
D Multimodal model
Answer: D. Multimodal model. A multimodal system can process or combine more than one data modality, such as text and images, depending on its design.
Does an AI system have to be a physical robot to qualify as AI?
A Yes, it needs wheels
B Yes, it needs sensors
C No, software can qualify
D No, it must be human
Answer: C. No, software can qualify. AI can be implemented entirely in software, such as a recommendation system or language model. A physical robot is not required.