OCI AI Foundations

Which statement best describes artificial intelligence?

• ASystems performing tasks associated with human intelligence
• BSystems that store and query business data in relational tables
• CSystems that always make error-free decisions from any input
• DSystems limited to controlling programmable physical robots

Which example is unstructured data commonly used in AI projects?

• AA customer table with fixed ID, date, and amount columns
• BAn audio recording captured from a customer support call
• CA spreadsheet with one typed field for each business attribute
• DA transaction CSV whose columns follow a documented schema

How do AI, machine learning, and deep learning relate to one another?

• AAI is a subset of machine learning, which is a subset of deep learning
• BDeep learning is a subset of machine learning, which is a subset of AI
• CAI and machine learning are identical terms
• DDeep learning is unrelated to AI

Which business scenario is a common AI personalization use case?

• ARecommending products from a customer's prior browsing behavior
• BEncrypting nightly backups with centrally managed security keys
• CAssigning private IP addresses across several application subnets
• DCompressing archived system logs before moving them to storage

What makes a system machine learning rather than purely rule-based automation?

• AIt learns patterns from examples and adapts its predictions
• BIt executes thresholds explicitly written by a domain expert
• CIt runs a fixed schedule configured by an administrator
• DIt retrieves a constant value from a static lookup table

Why is unstructured data often harder to use directly than structured data?

• AIt is always encrypted with an unknown cryptographic key
• BIt lacks a fixed schema and needs representation or feature extraction
• CIt can be stored only in a specialized relational database table
• DIt contains only numeric values with no contextual information

Which statement best distinguishes deep learning from many traditional machine learning methods?

• AIt applies one fixed linear formula to every possible input record
• BIt uses multilayer networks to learn hierarchical representations
• CIt depends entirely on manually written rules and decision thresholds
• DIt produces outputs without learning any parameters from training data

An assistant accepts a product photo together with a written question and uses both inputs to answer. What type of AI system is this?

• AA multimodal system combining visual and language inputs
• BA tabular regression system using one continuous target
• CA clustering system using unlabeled numeric rows
• DA rule engine using a fixed text-only decision tree

Which statement about AI and machine learning is accurate?

• AEvery artificial intelligence system must learn from labeled data
• BAI can include rule-based systems that never learn from examples
• CMachine learning includes all software that uses conditional logic
• DDeep learning is broader than both artificial intelligence and ML

Why is deep learning often applied to images and free text, while classic machine learning often starts with structured tables?

• AImages and text always contain fewer values than structured tables
• BDeep learning learns features from raw data; tables expose explicit fields
• CStructured tables never contain missing, noisy, or inconsistent values
• DDeep learning cannot process numeric values stored in table columns

In supervised learning, what is a label?

• AThe known target outcome paired with a supervised input example
• BAn input attribute supplied to the model when making a prediction
• CA learned weight adjusted by the optimizer during model training
• DA model output produced during inference for a previously unseen record

A model predicts monthly sales revenue for next month. What type of supervised task is this?

• AClassification
• BRegression
• CClustering
• DReinforcement learning

What does clustering do in unsupervised learning?

• APredicts future numeric values from labeled examples
• BChooses actions to maximize reward
• CGroups similar data points without predefined labels
• DEncrypts training data before use

Why might a team normalize numerical features before training a distance-based or gradient-based model?

• ATo reduce domination by features measured on much larger numeric scales
• BTo turn every regression target into a class label
• CTo replace missing labels with model predictions
• DTo make the test set identical to the training set

In reinforcement learning, what guides an agent toward better behavior over time?

• AA reward signal evaluating the consequences of selected actions
• BA class label paired with every example in supervised learning
• CA cluster centroid calculated before any observations are processed
• DA final test score repeatedly fed back into each training update

What is overfitting in machine learning?

• AA model too simple to capture patterns in its training data
• BA model fitting training noise but failing on new data
• CA model evaluated once on a held-out final test set
• DA model trained with both regularization and early stopping

What is an unsupervised anomaly-detection system intended to find?

• ARare observations that differ substantially from normal patterns
• BKnown target classes supplied with every training observation
• CThe numeric value paired with each supervised regression example
• DThe action yielding the greatest reward in a simulated environment

A help desk labels tickets as billing, account, or technical. What supervised task is this?

• AMulticlass classification with three possible target categories
• BRegression with one unrestricted continuous numeric target
• CClustering that discovers groups without using the ticket labels
• DReinforcement learning based on rewards from an environment

If missing a true positive is very costly, which metric usually deserves extra attention?

• AAccuracy
• BRecall
• CStorage capacity
• DInference region count

Why is using the test set repeatedly during model tuning a problem?

• AIt makes the training loop consume more memory than necessary
• BIt turns the test set into tuning data and biases evaluation
• CIt prevents the model from learning useful numeric features
• DIt forces every candidate model to use identical hyperparameters

In neural-network training, what is a mini-batch?

• AA subset of training examples used to compute one parameter update
• BA group of layers removed after each training epoch
• CA validation set that is reused as training data
• DA compressed model artifact exported for inference

Which model family is especially suited to image analysis tasks?

• AConvolutional neural networks using shared spatial filters
• BRecurrent neural networks maintaining an ordered hidden state
• CLinear regression models fitting one continuous numeric target
• DCentroid clustering algorithms grouping unlabeled table rows

Which type of model is designed for ordered data such as text or time series?

• ASequence models such as recurrent networks and LSTMs
• BImage models based only on two-dimensional convolution filters
• CRegression models that ignore order among independent table rows
• DClustering models that require no sequence position information

Why are convolutional neural networks effective for image tasks?

• ALocal filters detect patterns and share weights across image locations
• BRecurrent states retain ordered context across many language tokens over time
• CVector indexes retrieve database rows by semantic similarity scores
• DReward functions evaluate actions selected by an autonomous agent

Why does deep learning often benefit from large datasets and GPU acceleration?

• AMany parameters require matrix operations that GPUs execute in parallel
• BGPU acceleration guarantees accuracy without representative examples
• CLarge datasets entirely remove the need for model validation and final testing
• DDeep networks can train without calculating gradient-based updates

In a neural network, what is mainly learned during training?

• AThe weights and biases controlling transformations between layers
• BThe fixed activation-function definitions selected by the architect
• CThe original training records stored outside the neural network
• DThe ground-truth labels supplied with supervised training examples

A manufacturer wants to identify scratches on product photos from an assembly line. Which deep learning approach is most appropriate?

• AA convolutional neural network
• BA sequence model for time steps
• CA relational database index
• DA text summarization pipeline

Why are LSTMs commonly used instead of simple feed-forward networks for sentence modeling?

• ATheir gated state can retain ordered context across a sentence
• BThey discard token order and treat a sentence as an unordered set
• CThey replace learned parameters with a fixed vocabulary lookup table
• DThey require every sentence to contain exactly the same words

What limitation of simple bag-of-words features do sequence models address?

• AA word-count vector preserves exact grammar and token positions
• BAn unordered word-count vector loses token order and longer context
• CA bag-of-words vector includes learned positional embeddings by default
• DBag-of-words requires a reward signal for every vocabulary item

What problem do LSTMs help mitigate compared with basic RNNs?

• ADifficulty retaining useful gradients across many sequence steps
• BDifficulty assigning rows to centroids in an unlabeled table
• CDifficulty applying convolution filters near an image boundary
• DDifficulty converting a generated SQL statement into a query plan

What is generative AI designed to do?

• ACreate new content by sampling patterns learned during training
• BAssign only fixed class labels to every possible business record
• CStore relational rows without learning relationships among them
• DGuarantee that every generated statement is factually supported

Large language models are primarily trained to do what?

• APredict the next likely token from the preceding sequence context
• BAssign every document to a fixed cluster without token ordering
• CCalculate a vector index directly from relational primary keys
• DDetect object locations in images using bounding-box coordinates

What key idea makes transformers effective for language tasks?

• ARecurrent processing that handles only one token at a time
• BSelf-attention relating each token to relevant context tokens
• CSpatial pooling that reduces the dimensions of an input image
• DA vector index that performs similarity search inside a database

What is prompt engineering?

• AChanging model weights through another supervised training run
• BDesigning prompt instructions, context, and demonstrations
• CCreating a database vector index before collecting documents
• DAttaching a ground-truth label to every raw training example

What is instruction tuning?

• ATraining on instruction-response pairs to improve instruction following
• BAdding clearer words to one prompt without changing model parameters
• CRetrieving current documents and placing their text into the context
• DReplacing the tokenizer while leaving training examples unchanged

When is fine-tuning more appropriate than prompt engineering alone?

• AWhen repeated domain behavior is needed and curated examples exist
• BWhen factual content changes weekly and retrieval can supply updates
• CWhen one request can be solved by clearer instructions and examples
• DWhen no relevant training examples or evaluation data are available