Machine Learning MCQs – Set-1 | 50 questions on learning types, datasets, model training, evaluation, optimization, and deployment.
Mixed difficulty · Select one answer and open each explanation to review the concept.
What is the central idea of machine learning?
A Learn patterns from data
B Store files in folders
C Transmit packets between networks
D Render images on screen
Answer: A. Learn patterns from data. Machine learning fits models from data so they can make predictions or generate outputs.
What does a machine-learning model do?
A Installs hardware drivers
B Maps inputs to outputs
C Creates a network cable
D Formats a storage device
Answer: B. Maps inputs to outputs. A model applies learned patterns or parameters to inputs to produce outputs.
Which data does supervised learning use during training?
A Only unstructured files
B Unlabeled examples only
C Examples paired with labels
D Random network packets
Answer: C. Examples paired with labels. Supervised learning trains on labeled examples containing inputs and target labels.
What kind of output does a classification model predict?
A A continuous measurement
B A file location
C A natural-language paragraph only
D A category or class
Answer: D. A category or class. Classification predicts a discrete class or category.
Which task is a regression problem?
A Predicting a numerical value
B Choose from class labels
C Grouping unlabeled examples
D Finding nearest documents
Answer: A. Predicting a numerical value. Regression predicts a numerical target, often a continuous value.
What is a key feature of unsupervised learning?
A All examples labeled
B Training data lacks target labels
C The model receives explicit rewards
D The model uses no data
Answer: B. Training data lacks target labels. Unsupervised methods seek structure in data without labeled target values.
What does clustering attempt to find?
A Teacher-supplied class labels
B A numerical target value
C Groups of similar examples
D The fastest processor
Answer: C. Groups of similar examples. Clustering groups examples according to similarities in their features.
Why use dimensionality reduction?
A Add more target labels
B Encrypt a training dataset
C Increase the number of examples
D Represent data with fewer features
Answer: D. Represent data with fewer features. Dimensionality reduction represents data using fewer variables while retaining useful structure.
What guides an agent in reinforcement learning?
A Rewards from its actions
B A folder structure
C Fixed class labels
D A processor clock
Answer: A. Rewards from its actions. Reinforcement learning uses rewards or penalties to guide action selection.
What is a feature in a dataset?
A The predicted target only
B Input variable for an example
C A model’s license
D A storage device
Answer: B. Input variable for an example. Features are input variables used to describe examples for a model.
What is a label in supervised learning?
A A random file name
B A learning-rate setting
C The target value or class
D A validation metric
Answer: C. The target value or class. A label is the target value or class that the model is trained to predict.
Which dataset is used to fit model parameters?
A Test set
B Validation set
C Production logs
D Training set
Answer: D. Training set. The training set is used to fit a model’s parameters.
What is a validation set commonly used for?
A Tune choices during development
B Fit the final model only
C Replace all training data
D Store model predictions permanently
Answer: A. Tune choices during development. A validation set helps compare models and tune hyperparameters during development.
Why keep a separate test set?
A To select every training example
B Estimate performance on unseen data
C To choose the model’s name
D To compute training gradients
Answer: B. Estimate performance on unseen data. A held-out test set estimates how a finalized model performs on unseen examples.
What is data leakage in machine learning?
A Using too few model layers
B Training on random batches
C Evaluation influenced by leaked data
D Compressing input files
Answer: C. Evaluation influenced by leaked data. Data leakage occurs when information unavailable at prediction time improperly influences training or evaluation.
What can overfitting indicate?
A Excellent unseen performance
B No relationship in training data
C Oversized test set
D Fit to training-specific details
Answer: D. Fit to training-specific details. Overfitting occurs when a model fits training-specific details and performs worse on new data.
How does underfitting commonly appear?
A Poor fit on all data
B Perfect training accuracy
C A very large test set
D Too many model predictions
Answer: A. Poor fit on all data. An underfit model fails to capture useful patterns and performs poorly even on its training data.
What does generalization measure?
A How quickly data downloads
B Performance on new examples
C The number of model parameters
D The size of a database
Answer: B. Performance on new examples. Generalization is how well a model performs on examples it did not train on.
Why shuffle or split data carefully?
A To make labels more colorful
B To increase model size
C Prevent train-test contamination
D To reduce processor voltage
Answer: C. Prevent train-test contamination. Careful separation prevents information from the evaluation data leaking into training.
What is a hyperparameter?
A Predicted target
B A value measured from users
C A data row identifier
D Setting chosen before parameter fitting
Answer: D. Setting chosen before parameter fitting. Hyperparameters are settings chosen to control training or model structure rather than learned directly as ordinary model weights.
How are model parameters commonly obtained?
A Learned from training data
B Copied from test labels
C Chosen as file extensions
D Assigned by network routers
Answer: A. Learned from training data. Training algorithms adjust model parameters using examples and an objective.
What does gradient descent try to do?
A Increase every model weight
B Reduce loss through updates
C Select a test dataset
D Generate new feature labels
Answer: B. Reduce loss through updates. Gradient descent updates model parameters in a direction intended to reduce the loss.
What does the learning rate control?
A The number of input features
B Size of training files
C The size of parameter updates
D The test set’s class balance
Answer: C. The size of parameter updates. The learning rate scales parameter updates during optimization.
What is one epoch in model training?
A One prediction for each class
B One database transaction
C One test-set pass
D One pass through training examples
Answer: D. One pass through training examples. An epoch is one complete pass through the training dataset.
What does a training batch contain?
A A subset of training examples
B All possible model parameters
C Only the test labels
D A set of output classes
Answer: A. A subset of training examples. A batch is a subset of examples processed together during training.
What does a loss function quantify?
A The model’s network speed
B Prediction error relative to targets
C The number of input files
D The dataset’s storage format
Answer: B. Prediction error relative to targets. A loss function assigns a value that measures prediction error for an optimization objective.
What does an optimizer do during training?
A Creates new labels automatically
B Chooses the model’s license
C Updates parameters to reduce loss
D Stores evaluation reports
Answer: C. Updates parameters to reduce loss. An optimizer updates model parameters using gradients or another optimization rule.
Which term names the error measure used to guide learning?
A Feature
B Epoch
C Batch
D Loss
Answer: D. Loss. Loss quantifies model error relative to the selected objective.
What does a confusion matrix summarize?
A Predicted versus actual counts
B The order of training files
C Model parameters by layer
D The training set’s storage size
Answer: A. Predicted versus actual counts. A confusion matrix counts combinations of predicted and actual classes.
How is classification accuracy calculated?
A True positives over predicted positives
B Correct predictions over total
C False positives over negatives
D Recall divided by precision
Answer: B. Correct predictions over total. Accuracy is the fraction of all predictions that are correct.
What does precision measure for a positive class?
A Actual positives recovered
B All correct predictions
C Correct positives among predicted positives
D True negatives among actual negatives
Answer: C. Correct positives among predicted positives. Precision is true positives divided by all positive predictions.
What does recall measure for a positive class?
A Correct positives among predicted positives
B Correct predictions among all examples
C False positives among actual negatives
D Correct positives among actual positives
Answer: D. Correct positives among actual positives. Recall is true positives divided by all actual positives.
What does the F1 score balance?
A Precision and recall
B Accuracy and training speed
C Features and labels
D Loss and model size
Answer: A. Precision and recall. The F1 score is the harmonic mean of precision and recall.
Why can class imbalance make accuracy misleading?
A It increases every class equally
B Majority-only model may score highly
C Removes need for labels
D It guarantees overfitting
Answer: B. Majority-only model may score highly. When one class dominates, predicting that class can yield high accuracy while missing the rare class.
What does a classification threshold control?
A Which features enter training
B The size of the dataset
C Converts scores into classes
D How many model layers exist
Answer: C. Converts scores into classes. A threshold converts a model score or probability into a class decision.
What is cross-validation used to estimate?
A A model’s purchase price
B The dataset’s file format
C The processor’s learning rate
D Performance across data splits
Answer: D. Performance across data splits. Cross-validation repeats evaluation across data splits to estimate how performance varies.
Why scale numerical features before some models?
A Put values on comparable ranges
B Increase the number of labels
C Remove the test set
D Convert regression into classification
Answer: A. Put values on comparable ranges. Scaling can help optimization and distance-based methods when features use very different ranges.
What is feature engineering?
A Changing test labels
B Creating useful model inputs
C Increasing network bandwidth
D Deleting model parameters
Answer: B. Creating useful model inputs. Feature engineering creates or transforms inputs to make useful patterns easier for a model to learn.
How can regularization reduce overfitting?
A By adding more test examples
B By removing all labels
C By penalizing overly complex solutions
D By increasing every weight
Answer: C. By penalizing overly complex solutions. Regularization penalizes complexity or large parameter values to encourage simpler fits.
What does early stopping use to choose when training ends?
A Only the model’s file size
B The test labels repeatedly
C The number of feature names
D Validation performance
Answer: D. Validation performance. Early stopping halts training when validation performance stops improving or begins to worsen.
What is a neural network layer made of?
A Units that transform incoming values
B A set of training labels
C A database of test rows
D Network routers
Answer: A. Units that transform incoming values. Neural network layers contain units that transform inputs through weights and activation functions.
What does an activation function add to a neural network?
A A larger training dataset
B Nonlinear transformations
C A class label
D A validation split
Answer: B. Nonlinear transformations. Activation functions introduce nonlinear transformations that let networks represent complex relationships.
What does backpropagation calculate?
A A dataset’s class balance
B A model’s market price
C Gradients for parameter updates
D The number of test examples
Answer: C. Gradients for parameter updates. Backpropagation computes gradients of the loss with respect to model parameters.
What distinguishes deep learning from many simpler models?
A It uses no data
B It has no parameters
C It never needs evaluation
D Neural networks with multiple layers
Answer: D. Neural networks with multiple layers. Deep learning commonly uses neural networks with multiple layers of learned representations.
Which architecture is commonly used to learn spatial image patterns?
A Convolutional neural network
B Decision list
C Linear regression only
D Key-value database
Answer: A. Convolutional neural network. Convolutional neural networks use local filters that can learn spatial patterns in image-like data.
What is transfer learning?
A Training without any source data
B Reuse knowledge from pretrained models
C Converting labels into features
D Measuring network latency
Answer: B. Reuse knowledge from pretrained models. Transfer learning adapts knowledge from a pretrained model to a related task.
What can an embedding represent for a machine-learning system?
A A physical server location
B A model’s license key
C Item represented as vector
D A training epoch
Answer: C. Item represented as vector. Embeddings map items such as words or images into numerical vectors that capture useful relationships.
What is model drift?
A Database schema change
B Smaller training batch
C A feature scaling method
D Performance shift as data changes
Answer: D. Performance shift as data changes. Model drift occurs when changing data or conditions reduce a model’s performance over time.
Why monitor a deployed machine-learning model?
A Detect performance changes and failures
B Increase its training labels
C Prevent every data update
D Remove all prediction errors
Answer: A. Detect performance changes and failures. Monitoring can reveal drift, failures, or changes in prediction quality after deployment.
How can data augmentation support model training?
A Varied examples from existing data
B Removes all training labels
C Replaces the loss function
D Shuts down model inference
Answer: A. Varied examples from existing data. Data augmentation applies transformations to existing examples to increase training variation.