I85PREP
⌘ K
AIAI study materials
AIAI study resources

AIAI practice questions

One question at a time. Build your AIAI understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI Fundamentals and Terminology

Which term describes a machine learning system that learns patterns from labeled training data?

  1. Supervised learning Correct answer.
  2. Unsupervised learning
  3. Generative modeling
  4. Reinforcement learning

Explanation

Supervised learning uses input-output pairs where the target label is known during training, so the model learns a mapping from features to labels. The other paradigms do not rely on labeled examples in the same way.

Question 2AI Fundamentals and Terminology

In machine learning, what is a 'feature'?

  1. The final predicted output of a model
  2. The hardware used to train the model
  3. The loss value computed after training
  4. A measurable input variable used to make predictions Correct answer.

Explanation

Features are the individual measurable properties or attributes of the data that a model uses as inputs. The target is what is predicted, and loss measures error.

Question 3AI Fundamentals and Terminology

What does the term 'inference' refer to in an AI system?

  1. Collecting and cleaning raw data
  2. Splitting data into training and test sets
  3. Using a trained model to produce outputs on new inputs Correct answer.
  4. Adjusting weights during backpropagation

Explanation

Inference is the deployment-time phase where a trained model computes predictions or generations for new inputs. Training, by contrast, updates model parameters.

Question 4AI Fundamentals and Terminology

Which statement best defines artificial general intelligence (AGI)?

  1. A system tuned to excel at one narrowly defined task
  2. A system that matches or exceeds humans across a broad range of cognitive tasks Correct answer.
  3. A system that only generates human-like text
  4. A system that stores a large curated knowledge base

Explanation

AGI denotes broad, general-purpose competence spanning many cognitive domains rather than specialization. Narrow AI, by contrast, is designed for a specific task.

Question 5AI Fundamentals and Terminology

What is the primary purpose of a validation set during model development?

  1. To generate additional labeled training examples
  2. To provide the final unbiased estimate of performance
  3. To store the model weights after training
  4. To tune hyperparameters and compare model variants Correct answer.

Explanation

The validation set is used for model selection and hyperparameter tuning, guiding decisions during development. The test set is reserved for the final unbiased estimate.

Question 6AI Fundamentals and Terminology

Which of these is a classic example of an unsupervised learning task?

  1. Estimating a person's age from a photo with labeled ages
  2. Clustering customers by purchasing behavior Correct answer.
  3. Classifying emails as spam or not spam
  4. Predicting house prices from labeled sales

Explanation

Clustering groups unlabeled data points by similarity, which is the hallmark of unsupervised learning. The other tasks rely on known target labels.

Question 7AI Fundamentals and Terminology

What is a 'token' in the context of large language models?

  1. A single neural network layer
  2. A unit of currency used to pay for hardware
  3. A unit of text, such as a word piece or character sequence, processed by the model Correct answer.
  4. A type of activation function

Explanation

Language models operate on tokens, which are subword or word-level chunks produced by a tokenizer. The model maps sequences of tokens to probabilities over the next token.

Question 8AI Fundamentals and Terminology

Which term describes an AI model trained on a specific, narrowly scoped task such as recognizing handwritten digits?

  1. Narrow AI Correct answer.
  2. Strong AI
  3. General AI
  4. Artificial consciousness

Explanation

Narrow AI is designed and trained for a limited task domain. General or strong AI would imply broad, human-like competence across many tasks.

Question 9AI Fundamentals and Terminology

A team wants to detect fraudulent transactions but has labels only for a small fraction of past transactions. Which learning setting best matches this situation?

  1. Supervised learning with no labels
  2. Reinforcement learning
  3. Semi-supervised learning Correct answer.
  4. Unsupervised clustering only

Explanation

Semi-supervised learning combines a small labeled set with a larger unlabeled set, which matches a scenario with sparse labels. Pure supervised learning would require labels for all examples.

Question 10AI Fundamentals and Terminology

What does 'overfitting' mean for a machine learning model?

  1. The model has too few parameters to learn the task
  2. The model performs poorly on both training and test data
  3. The model fits training data well but generalizes poorly to new data Correct answer.
  4. The model requires too much memory to deploy

Explanation

Overfitting occurs when a model captures noise or idiosyncrasies of the training set, producing low training error but high error on unseen data. Underfitting is the opposite pattern.

Question 11AI Fundamentals and Terminology

Which description best captures a 'foundation model'?

  1. A model that has been fully fine-tuned and cannot be reused
  2. A large model pretrained on broad data that can be adapted to many downstream tasks Correct answer.
  3. A small model trained only on a single company's spreadsheet data
  4. A rule-based system with hand-written decision logic

Explanation

Foundation models are pretrained at scale on broad data and then adapted via prompting or fine-tuning to many tasks. Their defining property is broad pretraining plus adaptability.

Question 12AI Fundamentals and Terminology

In reinforcement learning, what is the role of the 'reward signal'?

  1. It labels each training example with a class
  2. It initializes the neural network weights
  3. It stores the environment's transition probabilities
  4. It defines the objective the agent learns to maximize Correct answer.

Explanation

The reward signal provides scalar feedback that specifies what the agent should try to maximize over time. The agent adjusts its policy to increase cumulative reward.

Question 13AI Fundamentals and Terminology

A model achieves 99% accuracy on its training set but only 62% on a held-out test set. What is the most likely diagnosis?

  1. Overfitting to the training data Correct answer.
  2. A data leakage-free but biased test set
  3. Underfitting to the training data
  4. Insufficient model capacity

Explanation

A large gap between high training accuracy and low test accuracy is the classic signature of overfitting. The model has memorized training-specific patterns that do not transfer.

Question 14AI Fundamentals and Terminology

What is 'prompt engineering'?

  1. Designing input text to steer a model toward desired outputs Correct answer.
  2. Compiling model weights into a binary format
  3. Building the physical hardware for model training
  4. Auditing datasets for labeling errors

Explanation

Prompt engineering is the practice of crafting and iterating on input instructions and examples to elicit better model behavior. It operates at the input level rather than changing the model architecture.

Question 15AI Fundamentals and Terminology

Which statement distinguishes parametric knowledge from retrieval-based knowledge in an AI system?

  1. Parametric knowledge is fetched from a database, while retrieval is stored in weights
  2. Both terms describe the same mechanism with different names
  3. Parametric knowledge refers only to image data, while retrieval refers only to text
  4. Parametric knowledge is stored in model weights, while retrieval fetches external documents at query time Correct answer.

Explanation

Parametric knowledge is baked into the model's parameters during training, whereas retrieval augments generation by looking up external sources at inference time. This distinction underpins retrieval-augmented generation.

Question 16AI Fundamentals and Terminology

What does 'hallucination' mean for a generative AI model?

  1. Running inference more slowly than expected
  2. Generating output in the wrong language
  3. Failing to produce any output at all
  4. Producing fluent output that is factually incorrect or unsupported Correct answer.

Explanation

Hallucination refers to confident, fluent generation of content that is not grounded in fact or source material. It is a correctness problem rather than a formatting or latency problem.

Question 17AI Fundamentals and Terminology

A model is described as 'multimodal'. What does this mean?

  1. It can process or generate more than one data type, such as text and images Correct answer.
  2. It was trained by multiple organizations
  3. It has multiple hidden layers in its network
  4. It can run on multiple hardware vendors

Explanation

Modality refers to the type of data, such as text, image, or audio. A multimodal model handles more than one such type.

Question 18AI Fundamentals and Terminology

Which statement best describes 'fine-tuning'?

  1. Removing all biases from a dataset
  2. Training a model from scratch on random initialization
  3. Compressing a model for faster inference
  4. Further training a pretrained model on task-specific data to adapt its behavior Correct answer.

Explanation

Fine-tuning starts from pretrained weights and continues training on a smaller, targeted dataset so the model adapts to a specific task or style. It is distinct from training from scratch.

Question 19AI Fundamentals and Terminology

In the standard machine learning workflow, what is the main purpose of the training set?

  1. To report final production performance
  2. To validate that the pipeline runs without errors
  3. To update model parameters so the model learns patterns Correct answer.
  4. To choose between competing model families

Explanation

The training set drives parameter updates through an optimization procedure. Model selection uses validation data and final reporting uses test data.

Question 20AI Fundamentals and Terminology

What is 'data leakage' in machine learning?

  1. Loss of data during storage migration
  2. Information from the test set influencing training, inflating measured performance Correct answer.
  3. Unauthorized access to a model's weights
  4. Using too many features in a model

Explanation

Data leakage occurs when information that should be unavailable at prediction time contaminates training, producing overly optimistic evaluation results. Careful splitting and feature construction prevent it.

Question 21AI Fundamentals and Terminology

A team builds a system that summarizes support tickets. Which capability category does this primarily represent?

  1. Natural language processing Correct answer.
  2. Robotic control
  3. Speech synthesis
  4. Computer vision

Explanation

Summarizing text is a natural language processing task because it involves understanding and generating human language. Vision, speech synthesis, and robotics concern other modalities.

Question 22AI Fundamentals and Terminology

Which statement best captures the difference between a 'model' and an 'algorithm' in machine learning?

  1. An algorithm is a learning procedure, while a model is the fitted artifact produced by that procedure Correct answer.
  2. The two terms are interchangeable in all contexts
  3. A model is a learning procedure, while an algorithm is the fitted artifact
  4. An algorithm is always a neural network, while a model is always a decision tree

Explanation

An algorithm specifies how learning proceeds, such as gradient descent on a defined objective. The model is the resulting artifact with learned parameters that can be used for inference.

Question 23AI Fundamentals and Terminology

Why is accuracy often a misleading metric for a highly imbalanced binary classification problem?

  1. Accuracy cannot be computed for binary problems
  2. A model predicting the majority class always can score high accuracy while missing the minority class Correct answer.
  3. Accuracy requires labels that are never available in practice
  4. Accuracy always underestimates the majority class

Explanation

With severe class imbalance, predicting the majority class for every example yields high accuracy yet zero recall on the minority class. Metrics such as precision, recall, or F1 better reflect performance on the minority.

Question 24AI Fundamentals and Terminology

An AI assistant must answer questions about a company's internal policy documents that change weekly. Which architecture best addresses freshness and grounding?

  1. A rule-based keyword matcher with fixed responses
  2. A model fine-tuned once on last year's policies
  3. A larger context window with no external sources
  4. Retrieval-augmented generation over an updated document index Correct answer.

Explanation

RAG retrieves current documents at query time, so answers stay grounded in the latest content without retraining. Fine-tuning on stale data would not reflect weekly changes.

Question 25AI Fundamentals and Terminology

Which scenario best illustrates the 'alignment' problem in AI?

  1. A model is deployed on hardware it was not benchmarked on
  2. A model is trained on a dataset that is too small to converge
  3. A model optimizes a measurable proxy objective in ways that diverge from the designers' true intent Correct answer.
  4. A model's inference latency exceeds the service-level target

Explanation

Alignment concerns ensuring that a system's optimized objective matches human intent; optimizing a proxy can produce unintended behavior. Convergence, latency, and hardware issues are engineering concerns rather than alignment.

Question 26AI Fundamentals and Terminology

Which term describes the process of dividing a dataset into subsets for training, validation, and testing?

  1. Feature engineering
  2. Data augmentation
  3. Data splitting Correct answer.
  4. Model selection

Explanation

Data splitting is the practice of partitioning available data into separate subsets so that model training and evaluation use different data points. This helps estimate how well a model generalizes to unseen data.

Question 27AI Fundamentals and Terminology

A data scientist trains a model on a dataset where 95% of examples belong to class A. The model predicts class A for every input and achieves 95% accuracy. Which action would best address this issue?

  1. Increase the size of the test set
  2. Collect more features for each example
  3. Use precision, recall, or F1 score instead of accuracy Correct answer.
  4. Switch to a more complex model architecture

Explanation

With severe class imbalance, accuracy can be misleading because a trivial majority-class predictor scores high. Metrics like precision, recall, or F1 focus on the minority class and reveal whether the model actually detects it.

Question 28AI Fundamentals and Terminology

A team wants to reduce the dimensionality of a large dataset while preserving as much variance as possible. Which technique is most appropriate?

  1. One-hot encoding
  2. Random forest
  3. Principal Component Analysis (PCA) Correct answer.
  4. Gradient descent

Explanation

PCA is a dimensionality-reduction method that projects data onto a lower-dimensional space defined by principal components, which capture the directions of maximum variance. It is specifically designed to preserve variance while reducing the number of features.

Question 29AI Fundamentals and Terminology

In a neural network, what is the role of an activation function?

  1. To update the weights during backpropagation
  2. To compute the loss between predictions and labels
  3. To introduce non-linearity into the network's computations Correct answer.
  4. To normalize the input features before training

Explanation

Activation functions are applied to the output of each neuron to introduce non-linear transformations, enabling the network to learn complex patterns. Without non-linear activation, a multi-layer network would collapse to an equivalent linear model.

Question 30AI Fundamentals and Terminology

In supervised learning, what does the term 'label' refer to?

  1. The learning rate used during optimization
  2. The number of layers in a neural network
  3. The raw input features fed into the model
  4. The target output associated with a training example Correct answer.

Explanation

A label is the ground-truth target that a supervised model is trained to predict for a given input example. Inputs are the features, while labels supply the supervision signal that defines the learning objective.

Question 31Machine Learning Concepts and Model Lifecycle

A model scores 0.98 on the training set but only 0.62 on a held-out test set. Which term best describes this pattern?

  1. Overfitting Correct answer.
  2. Data leakage
  3. Class imbalance
  4. Underfitting

Explanation

A large gap between strong training performance and weak held-out performance indicates the model has memorized training-specific noise rather than generalizable structure. Underfitting would show poor performance on both sets.

Question 32Machine Learning Concepts and Model Lifecycle

Which split is used to tune hyperparameters while keeping a final unbiased estimate of performance?

  1. Training set
  2. Test set
  3. Validation set Correct answer.
  4. Production sample

Explanation

Hyperparameters are selected using a validation set, which is separate from training data. The test set is held out until the end to provide an unbiased final estimate.

Question 33Machine Learning Concepts and Model Lifecycle

In k-fold cross-validation, what does the value k represent?

  1. The number of hyperparameters tuned
  2. The number of features retained
  3. The number of partitions the data is divided into Correct answer.
  4. The number of boosting rounds

Explanation

Cross-validation partitions the dataset into k folds; each fold serves once as validation while the others train the model. Performance is averaged across the k runs.

Question 34Machine Learning Concepts and Model Lifecycle

A team wants to reduce variance in their performance estimate without collecting more data. Which technique directly addresses this?

  1. Removing the validation set
  2. Training longer on the full dataset
  3. Cross-validation with averaged scores Correct answer.
  4. Increasing the learning rate

Explanation

Averaging results over multiple folds reduces the variance of the estimated performance compared with a single train-validation split. It reuses the same data more efficiently.

Question 35Machine Learning Concepts and Model Lifecycle

Which metric is most appropriate for a highly imbalanced binary classification problem where missing a positive case is costly?

  1. Accuracy
  2. Recall Correct answer.
  3. R-squared
  4. Mean squared error

Explanation

Recall measures the fraction of actual positives that are correctly identified, so it directly reflects the cost of missed positives. Accuracy can look high while ignoring minority-class failures.

Question 36Machine Learning Concepts and Model Lifecycle

What is the primary purpose of a baseline model in the ML lifecycle?

  1. To replace the need for a test set
  2. To eliminate feature engineering
  3. To serve as the final deployed model
  4. To provide a reference performance level for comparing new models Correct answer.

Explanation

A baseline, such as a majority-class or simple heuristic predictor, establishes a minimum performance bar. Newer models are judged against it to confirm they add value.

Question 37Machine Learning Concepts and Model Lifecycle

Which activity belongs to feature engineering rather than model training?

  1. Encoding categorical variables as numeric vectors Correct answer.
  2. Updating weights via backpropagation
  3. Selecting the number of epochs
  4. Minimizing a loss function with gradient descent

Explanation

Feature engineering transforms raw inputs into representations the model can use, such as encoding categories. Loss minimization, weight updates, and epoch selection are training procedures.

Question 38Machine Learning Concepts and Model Lifecycle

A model's performance degrades sharply months after deployment as user behavior shifts. What is this phenomenon called?

  1. Data drift Correct answer.
  2. Underfitting
  3. Regularization
  4. Overfitting

Explanation

Data drift occurs when the distribution of input data changes over time relative to what the model learned from. This degrades performance even if the model itself is unchanged.

Question 39Machine Learning Concepts and Model Lifecycle

Which practice most directly prevents information from the test set leaking into model development?

  1. Reporting test accuracy after every experiment
  2. Fitting preprocessing statistics only on training data Correct answer.
  3. Tuning hyperparameters on the test set
  4. Reusing the test set for feature selection

Explanation

Preprocessing such as scaling or imputation must be fit on training data only, then applied to validation and test data. Otherwise test information contaminates the training pipeline.

Question 40Machine Learning Concepts and Model Lifecycle

What does L2 regularization add to the training objective?

  1. A penalty proportional to the absolute value of the weights
  2. Random noise injected into the inputs
  3. A hard constraint that weights equal zero
  4. A penalty proportional to the squared magnitude of the weights Correct answer.

Explanation

L2 regularization adds the sum of squared weights, scaled by a coefficient, to the loss. This shrinks weights toward zero and reduces overfitting.

Question 41Machine Learning Concepts and Model Lifecycle

A dataset has 1,000 rows and a model uses 300 features. Which concern is most relevant?

  1. Inability to compute a loss function
  2. Loss of the ability to split the data
  3. Guaranteed underfitting from too few features
  4. Risk of overfitting due to high dimensionality relative to sample size Correct answer.

Explanation

When the number of features is large relative to the number of samples, models can fit noise easily. Regularization, feature selection, or more data help mitigate this.

Question 42Machine Learning Concepts and Model Lifecycle

Which step should occur first when building a supervised learning pipeline?

  1. Pruning the neural network
  2. Tuning the learning rate
  3. Deploying the model
  4. Defining the prediction target and success metric Correct answer.

Explanation

Clarifying the target variable and the metric that defines success guides all later choices, including data collection, model selection, and evaluation. Tuning and deployment come later.

Question 43Machine Learning Concepts and Model Lifecycle

A confusion matrix shows 40 true positives, 10 false negatives, 5 false positives, and 45 true negatives. What is the recall?

  1. 0.89
  2. 0.90
  3. 0.40
  4. 0.80 Correct answer.

Explanation

Recall equals true positives divided by the sum of true positives and false negatives: 40 / (40 + 10) = 0.80. It measures coverage of actual positives.

Question 44Machine Learning Concepts and Model Lifecycle

Why is stratified sampling often used when creating train-test splits for classification?

  1. To remove outliers automatically
  2. To increase the learning rate
  3. To reduce the number of features
  4. To preserve the class proportions in each split Correct answer.

Explanation

Stratification keeps the class distribution consistent across splits, which matters most for rare classes. This yields more reliable and comparable evaluation results.

Question 45Machine Learning Concepts and Model Lifecycle

Which statement best describes the bias-variance tradeoff?

  1. Increasing model complexity typically lowers bias but raises variance Correct answer.
  2. Variance is caused only by insufficient training data
  3. Bias and variance are unrelated to model capacity
  4. Increasing model complexity lowers both bias and variance indefinitely

Explanation

More flexible models can fit training data more closely, reducing bias, but they become more sensitive to sampling noise, raising variance. The goal is to balance the two.

Question 46Machine Learning Concepts and Model Lifecycle

What is the role of a learning curve that plots training and validation error against training set size?

  1. To diagnose whether more data would likely improve performance Correct answer.
  2. To compute the final test accuracy directly
  3. To select the optimizer automatically
  4. To guarantee the model will not overfit

Explanation

If validation error is still decreasing as training size grows and remains well above training error, additional data may help. A plateau suggests other remedies are needed.

Question 47Machine Learning Concepts and Model Lifecycle

A model is retrained monthly, but a feature's upstream source changes units without notice. What is the most likely consequence?

  1. Automatic regularization of the model
  2. Improved generalization from new units
  3. Reduced need for monitoring
  4. Training-serving skew causing degraded predictions Correct answer.

Explanation

When the data seen at serving time differs in form from training data, the model receives inputs outside its learned distribution. This is training-serving skew and typically harms predictions.

Question 48Machine Learning Concepts and Model Lifecycle

Which is the best reason to version both datasets and model artifacts in the ML lifecycle?

  1. To reduce storage costs
  2. To reproduce and audit any past model result Correct answer.
  3. To eliminate the need for testing
  4. To increase training speed

Explanation

Versioning data and models lets teams recreate the exact conditions behind a past result, supporting debugging, audits, and rollback. Reproducibility is the core benefit.

Question 49Machine Learning Concepts and Model Lifecycle

A team evaluates a model on data collected after the training period. What does this assess?

  1. Hyperparameter sensitivity
  2. Feature importance
  3. Training loss convergence
  4. Out-of-time generalization Correct answer.

Explanation

Evaluating on data from a later time period tests whether the model generalizes forward in time, which is often the realistic deployment condition. It captures temporal robustness.

Question 50Machine Learning Concepts and Model Lifecycle

Which scenario most clearly indicates underfitting?

  1. Validation error decreases steadily with more data
  2. Training error equals zero
  3. Training error is low while validation error is high
  4. Both training and validation error are high and similar Correct answer.

Explanation

Underfitting occurs when the model is too simple to capture the underlying pattern, so it performs poorly on both training and validation data. The two errors being high and close together is the hallmark.