Foundations of AI Systems
Free IAPP AI Governance Professional (AIGP) practice — 6 questions on Foundations of AI Systems, with explanations. No sign-up. Full 12-question mixed test →
A financial services company is deploying a credit risk assessment system that combines a gradient boosting model for numerical features, a recurrent neural network for transaction sequences, and a rule-based system for regulatory compliance checks. The system's final decision is made by weighted voting among these three components. How should this architecture be classified?
This is a hybrid AI system because it explicitly combines different AI paradigms: machine learning (gradient boosting and neural networks) with symbolic AI (rule-based system). Hybrid systems integrate multiple distinct AI approaches to solve a problem, which is exactly what's described.
An organization needs to process medical imaging data where each scan must remain on hospital premises due to regulations, but they want a centralized model that learns patterns from all participating hospitals. The system must prevent any hospital from reconstructing another hospital's patient data from the shared model. Training occurs in rounds where each hospital updates local model weights before aggregating. Which approach is MOST appropriate?
Federated learning allows training on decentralized data without moving it from source locations (meets the 'remain on premises' requirement). Adding differential privacy ensures that individual data points (patient scans) cannot be reconstructed from the shared model weights, addressing the privacy constraint. The round-based local-then-aggregate training described is the classic federated learning pattern.
A recommendation system currently uses collaborative filtering based on user-item interaction matrices. The data science team wants to incorporate product descriptions, user demographic information, and temporal patterns of purchases. They also need the system to explain why specific products are recommended to users. What architectural change would BEST support these requirements?
Knowledge graphs can represent heterogeneous data types (user demographics, product attributes, temporal relationships) as entities and relationships. Graph neural networks can process this structure while attention mechanisms provide interpretability by showing which graph paths influenced recommendations. This addresses all requirements: multi-type data integration, relational reasoning, and explainability.
A manufacturing company deploys a predictive maintenance system where edge devices on factory equipment run inference locally using compressed models, while a cloud-based system periodically retrains the full model on aggregated data and redistributes updated compressed versions. Inference latency must be under 10ms with no internet dependency during operation. Which distributed AI architecture does this scenario represent?
This describes edge-cloud hybrid architecture: edge devices perform inference locally (meeting the 10ms and no-internet requirements) using compressed models, while the cloud handles resource-intensive retraining and distributes updates. The periodic synchronization pattern and model compression are key characteristics of edge-cloud hybrid systems.
An AI system for autonomous vehicles uses a convolutional neural network for object detection, a recurrent network for trajectory prediction, a reinforcement learning agent for path planning, and a separate safety validation module that can override decisions. During testing, the safety module overrides 3% of path planning decisions. From a system architecture perspective, what type of control mechanism does the safety module represent?
The safety module acts as a guardian/supervisory layer that monitors the primary system's outputs and intervenes when safety constraints are violated. This is a common architectural pattern in safety-critical AI systems where a separate, often simpler and more verifiable component can override AI decisions. The ability to override indicates supervisory control.
A global e-commerce platform needs to deploy personalized recommendation models where user data cannot leave specific geographic regions due to data sovereignty laws, but the company wants to leverage insights from user behavior across all regions to improve model performance globally. Model updates from each region must not reveal individual user behavior. Which combination of techniques is MOST appropriate?
Federated learning keeps data in each region while enabling global model improvement through aggregated updates (solves data sovereignty). Secure multi-party computation allows regions to collaboratively compute without revealing individual updates. Differential privacy ensures individual user behavior isn't exposed in the shared model updates. This combination addresses all three requirements: data locality, global learning, and privacy.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$109.99 $34.99 with code FREETEST33 — valid through August 22.
Get my $34.99 deal →