Download Oracle Cloud Infrastructure OCI AI Foundations Associate Ultimate Exam and more Exams Technology in PDF only on Docsity!
Associate Ultimate Exam
Question 1. Which OCI service provides pre‑trained vision models accessible via a simple API? A) OCI Data Flow B) OCI Vision C) OCI Container Engine D) OCI Resource Manager Answer: B Explanation: OCI Vision offers ready‑to‑use image classification, object detection, and OCR models. Question 2. In OCI AI Infrastructure, what is the primary advantage of using Bare Metal over Virtual Machines for large‑scale model training? A) Lower cost per hour B) Direct access to GPU hardware without hypervisor overhead C) Automatic scaling of compute resources D) Built‑in data encryption Answer: B Explanation: Bare Metal provides exclusive GPU access, eliminating virtualization latency, which is critical for high‑throughput training. Question 3. Which term best describes the process of converting raw text into high‑dimensional vectors for similarity search in Oracle Database 23ai? A) Tokenization B) Embedding C) Normalization D) Indexing Answer: B
Associate Ultimate Exam
Explanation: Embedding maps text to dense vectors that can be stored and queried using AI Vector Search. Question 4. In the ML lifecycle, which step directly follows feature engineering? A) Data collection B) Model training C) Model evaluation D) Data preprocessing Answer: B Explanation: After features are engineered, the model is trained on those features. Question 5. Which activation function is most commonly used in the hidden layers of deep neural networks to mitigate vanishing gradients? A) Sigmoid B) ReLU C) Tanh D) Softmax Answer: B Explanation: ReLU (Rectified Linear Unit) preserves gradients for positive inputs, reducing vanishing‑gradient problems. Question 6. What does the “self‑attention” mechanism in a Transformer compute? A) The gradient of loss with respect to weights B. The similarity between each token and every other token in the sequence C) The convolutional filter response
Associate Ultimate Exam
B) To incorporate external knowledge bases during inference C) To fine‑tune the model on synthetic data D) To perform gradient clipping Answer: B Explanation: RAG fetches relevant documents at inference time, grounding the LLM’s responses. Question 10. Which OCI infrastructure component is specifically optimized for AI workloads that require massive parallelism? A) OCI Block Volume B) OCI Autonomous Data Warehouse C) OCI AI Supercluster D) OCI File Storage Service Answer: C Explanation: The AI Supercluster offers large numbers of GPUs and high‑speed interconnects for distributed training. Question 11. In reinforcement learning, what term describes the signal that encourages the agent to repeat a behavior? A) Loss B) Reward C) Penalty D) Gradient Answer: B Explanation: Rewards reinforce actions that lead to desirable outcomes.
Associate Ultimate Exam
Question 12. Which of the following best differentiates generative AI from discriminative AI? A) Generative models predict class labels; discriminative models generate data. B) Generative models create new content; discriminative models estimate conditional probabilities. C) Generative models require labeled data; discriminative models do not. D) Generative models are always unsupervised; discriminative models are always supervised. Answer: B Explanation: Generative AI synthesizes data (e.g., text, images), while discriminative AI focuses on classification or regression. Question 13. Which Oracle Cloud service enables developers to write natural‑language prompts that are translated into SQL queries? A) OCI Data Flow B) Select AI C) OCI Functions D) OCI API Gateway Answer: B Explanation: Select AI interprets conversational language and produces executable SQL. Question 14. In a Convolutional Neural Network, what is the primary role of a pooling layer? A) Increase the number of parameters B) Reduce spatial dimensions while retaining important features C) Apply non‑linear activation D) Perform batch normalization Answer: B
Associate Ultimate Exam
D) OCI Data Integration Answer: B Explanation: OCI Document Understanding applies OCR and entity extraction to business documents. Question 18. In the context of AI ethics, which principle focuses on ensuring that model outputs do not systematically disadvantage protected groups? A) Transparency B) Fairness C) Accountability D) Privacy Answer: B Explanation: Fairness aims to mitigate bias and prevent disparate impact. Question 19. Which type of data is most suitable for training a Recurrent Neural Network? A) Tabular sales records B) Static images C) Time‑series sensor readings D) One‑hot encoded categorical variables only Answer: C Explanation: RNNs capture temporal dependencies, making them ideal for sequential data. Question 20. What is the main benefit of Parameter‑Efficient Fine‑Tuning (PEFT) compared to full model fine‑tuning? A) It requires more GPU memory.
Associate Ultimate Exam
B) It modifies only a small subset of weights, reducing compute and storage. C) It changes the model architecture. D) It eliminates the need for a validation set. Answer: B Explanation: PEFT updates a lightweight adapter or prefix, keeping the base model frozen. Question 21. Which OCI feature allows you to store embeddings and perform nearest‑neighbor search without external vector databases? A) OCI Object Storage B) AI Vector Search in Oracle Database 23ai C) OCI Data Safe D) OCI File Storage Service Answer: B Explanation: AI Vector Search natively indexes high‑dimensional vectors for similarity queries. Question 22. In a Transformer encoder, what is the purpose of positional encoding? A) Encode token semantics B) Provide order information absent from self‑attention C) Reduce model size D) Perform dimensionality reduction Answer: B Explanation: Positional encodings inject sequence order because self‑attention treats inputs as a set. Question 23. Which machine‑learning algorithm is inherently unsupervised?
Associate Ultimate Exam
Question 26. What does “zero‑shot prompting” enable an LLM to do? A) Generate outputs without any task description B) Perform a task it has never seen by relying solely on its pre‑training knowledge C. Fine‑tune on a single example D. Reduce model size to zero parameters Answer: B Explanation: Zero‑shot prompts instruct the model to apply its general knowledge to a new task. Question 27. Which OCI service would you choose to run a custom TensorFlow model that you have containerized? A) OCI Functions B) OCI Container Engine for Kubernetes (OKE) C) OCI Data Flow D) OCI Streaming Answer: B Explanation: OKE orchestrates Docker containers, ideal for hosting custom AI workloads. Question 28. In the context of AI model evaluation, what does AUC‑ROC measure? A) Accuracy on the training set B) Area under the precision‑recall curve C) Ability of a classifier to rank positive instances higher than negative ones D) Computational cost of inference Answer: C
Associate Ultimate Exam
Explanation: AUC‑ROC quantifies the trade‑off between true‑positive and false‑positive rates across thresholds. Question 29. Which of the following best describes “embedding dimensionality reduction” in vector search? A) Converting vectors to one‑hot encoding B) Using PCA or UMAP to shrink vector size while preserving similarity C) Adding more dimensions to improve accuracy D) Normalizing vectors to unit length only Answer: B Explanation: Dimensionality reduction techniques compress vectors while retaining relative distances. Question 30. What is the primary role of batch normalization in deep networks? A) Increase model depth B) Stabilize and accelerate training by normalizing layer inputs C) Perform feature selection D) Convert categorical data to numeric form Answer: B Explanation: Batch norm reduces internal covariate shift, leading to faster convergence. Question 31. Which OCI AI service can automatically tag key entities (e.g., person, organization) in multilingual text? A) OCI Vision B) OCI Language C) OCI Speech
Associate Ultimate Exam
C) Autonomous Linux D. OCI Edge Services Answer: B Explanation: Selecting a specific OCI region keeps data and compute within that jurisdiction. Question 35. Which model architecture is most suitable for image segmentation tasks? A) RNN B) CNN with encoder‑decoder (e.g., U‑Net) C) Transformer encoder only D) Linear regression Answer: B Explanation: Encoder‑decoder CNNs capture spatial context and produce pixel‑wise predictions. Question 36. What does “gradient clipping” address during training? A) Vanishing gradients B) Exploding gradients that destabilize learning C) Data leakage D) Model bias Answer: B Explanation: Clipping caps gradient magnitude, preventing extreme updates. Question 37. Which OCI service provides a managed environment for hosting and versioning machine‑learning models for inference? A) OCI Data Catalog
Associate Ultimate Exam
B) OCI Model Deployment C) OCI Data Integration D) OCI Resource Manager Answer: B Explanation: OCI Model Deployment streamlines model serving with scaling and monitoring. Question 38. In the context of AI fairness, what is “counterfactual fairness”? A) Ensuring model predictions are identical for all inputs B) Comparing outcomes when a protected attribute is altered while everything else stays constant C) Using synthetic data to balance classes D) Removing all demographic features from the dataset Answer: B Explanation: Counterfactual fairness checks if changing a sensitive attribute would change the prediction. Question 39. Which of the following is a characteristic of a “generative adversarial network” (GAN)? A) Two models trained jointly: a generator and a discriminator B) Uses reinforcement learning for policy gradients C) Relies on a single autoencoder architecture D) Operates only on tabular data Answer: A Explanation: GANs pit a generator against a discriminator to produce realistic samples.
Associate Ultimate Exam
Explanation: Masking ensures autoregressive generation respects causal order. Question 43. Which metric is most appropriate for evaluating an imbalanced binary classification model where false negatives are critical? A) Accuracy B) ROC‑AUC C) Recall (Sensitivity) D) F1‑Score Answer: C Explanation: Recall emphasizes correctly identifying positive cases, reducing false negatives. Question 44. What does “model drift” refer to in production AI systems? A) Gradual loss of GPU performance B) Changes in data distribution causing degradation of model accuracy over time C. Increase in model size due to incremental learning D) Shift from supervised to unsupervised learning Answer: B Explanation: Drift occurs when the input data evolves away from the training distribution. Question 45. Which OCI resource type is used to securely store API keys and credentials for AI services? A) OCI Object Storage B) OCI Vault C) OCI Data Flow
Associate Ultimate Exam
D) OCI Streaming Answer: B Explanation: OCI Vault manages secrets, keys, and certificates for secure access. Question 46. In the context of LLM scaling laws, what happens when you increase model parameters while keeping data constant? A) Perplexity improves indefinitely B) Overfitting risk rises, yielding diminishing returns C) Training time decreases D) Model becomes more interpretable Answer: B Explanation: Without more data, larger models can overfit, offering limited performance gains. Question 47. Which technique reduces the dimensionality of word embeddings while preserving cosine similarity? A) One‑hot encoding B) Quantization with product quantizers C) Adding noise to vectors D. One‑dimensional scaling Answer: B Explanation: Product quantization compresses embeddings yet maintains similarity relationships. Question 48. What does “prompt chaining” enable when interacting with an LLM? A) Parallel inference across multiple GPUs
Associate Ultimate Exam
B) Re‑using a pre‑trained model’s weights as a starting point for a related task C) Converting a model to a different programming language D) Applying reinforcement learning to a supervised model Answer: B Explanation: Transfer learning leverages learned representations to accelerate downstream training. Question 52. What is the main benefit of using “early stopping” during model training? A) Guarantees zero training error B) Prevents overfitting by halting training when validation loss stops improving C) Increases model size automatically D) Reduces the need for data preprocessing Answer: B Explanation: Early stopping monitors validation performance and stops before over‑training. Question 53. Which OCI service provides managed streaming of data that can be consumed by AI pipelines in real time? A) OCI Data Flow B) OCI Streaming C) OCI Data Catalog D) OCI Functions Answer: B Explanation: OCI Streaming offers durable, ordered message streams for low‑latency ingestion. Question 54. In a Transformer, what does the term “multi‑head attention” refer to?
Associate Ultimate Exam
A) Using several attention layers in parallel to capture different representation subspaces B) Attending to multiple datasets simultaneously C) Applying attention after each convolutional layer D) Splitting the model into multiple GPUs Answer: A Explanation: Multi‑head attention projects inputs into multiple sub‑spaces, enriching context capture. Question 55. Which of the following is a key characteristic of “responsible AI” as defined by Oracle? A) Maximizing model size B) Ensuring transparency, fairness, privacy, and accountability C) Using only open‑source tools D) Deploying models without testing Answer: B Explanation: Oracle’s responsible AI framework emphasizes ethical considerations across the lifecycle. Question 56. What does “gradient descent with momentum” aim to improve? A) Learning rate scheduling B) Convergence speed by accumulating past gradients C) Data augmentation D) Model interpretability Answer: B Explanation: Momentum adds a velocity term, smoothing updates and accelerating progress across ravines.
