AI Bot
AI-103: AI Models, Data, and Responsible AI Flashcards
Microsoft Azure AI App and Agent Developer Associate AI-103 Flashcards
Study our AI-103: AI Models, Data, and Responsible AI flashcards for the Microsoft Azure AI App and Agent Developer Associate AI-103 exam with 36+ flashcards. View as flashcards, a searchable table, or as a fun matching game.
| Front | Back |
| Define bias in machine learning | Systematic errors that advantage or disadvantage groups often due to data or modeling choices |
| Explain synthetic data and when to use it | Artificial data generated to augment scarce or sensitive datasets for training while reducing privacy risk |
| Give three strategies to mitigate bias | Preprocessing reweighting or balancing in training and postprocessing adjustments and model auditing |
| How can you handle unlabeled data for training | Use semi supervised learning active learning weak supervision or human in the loop labeling |
| How do you choose an evaluation dataset | Representative of production inputs balanced across subgroups and including edge cases |
| How do you ensure data governance and compliance in Azure | Use policies auditing access retention labels and follow regulatory standards like GDPR and HIPAA |
| How do you evaluate generative model quality | Use human evaluation task specific metrics likelihood and diversity measures |
| How do you monitor deployed models in Azure | Track performance metrics data drift alerts logs and use Azure Monitor and Application Insights |
| How do you perform A B tests for models | Randomly route users compare metrics test for statistical significance and monitor for unintended effects |
| How do you reduce leakage in training data | Remove future information duplicate records and ensure clean temporal splits |
| List key steps in data preparation for ML in Azure | Ingest cleanse normalize label balance split and feature engineer |
| Name Azure tools for Responsible AI | Azure Machine Learning Responsible AI dashboard interpretability features fairness assessment and model explainers |
| Name common evaluation metrics for classification | Accuracy precision recall f1 score and ROC AUC |
| Name common evaluation metrics for regression | Mean squared error mean absolute error and R squared |
| Name three data ingestion challenges in Azure | Data heterogeneity volume and latency |
| What are best practices for storing sensitive training data | Encrypt at rest and in transit use access controls minimize copies and apply retention policies |
| What are common privacy controls in Azure for ML | Data encryption access controls private endpoints and customer managed keys |
| What are key factors when selecting an AI model for deployment in Azure | Performance compute cost latency accuracy data availability and regulatory constraints |
| What are prompt engineering best practices | Use clear instructions few shot examples control temperature and iterate with evaluation |
| What are the main responsible AI pillars to include in Azure projects | Fairness privacy reliability safety transparency and accountability |
| What checks should you perform before fine tuning | Assess data quality labeling size privacy requirements and choose a suitable base model |
| What does model provenance mean | Record of model artifacts data versions training code parameters and lineage for reproducibility and audit |
| What is a common fairness metric for classification | Demographic parity equal opportunity or equalized odds depending on context |
| What is a method for concept drift detection | Statistical tests feature distribution monitoring and performance degradation alerts |
| What is a model card and what should it include | A document with model purpose limitations performance metrics datasets and ethical considerations |
| What is adversarial robustness and why does it matter | Resistance to inputs crafted to fool models important for security and reliability |
| What is calibration and why is it important | Calibration means predicted probabilities match true outcomes important for decision making and trust |
| What is data drift and why does it matter | When input distribution changes over time causing model performance to degrade |
| What is differential privacy in simple terms | A technique that adds noise to computations to protect individual data while preserving aggregate statistics |
| What is model interpretability | Techniques to explain model predictions such as feature importance SHAP LIME and rule extraction |
| What is the difference between fine tuning and prompt tuning | Fine tuning updates model weights prompt tuning modifies inputs or uses soft prompts to steer model without changing weights |
| What is the role of MLOps in the model lifecycle | Automates deployment testing monitoring retraining and governance for production ML systems |
| When should you choose fine tuning over prompt tuning | When you need task specific behavior high accuracy and can afford compute data and model access |
| When should you choose prompt tuning over fine tuning | When model access is limited rapid iteration is needed or you want lower cost and lower risk |
| When should you use post hoc explainability methods | When model is complex or opaque and you need to explain individual predictions to stakeholders |
| Which Azure services are used for data ingestion and storage | Azure Data Factory Azure Blob Storage Azure Data Lake and Event Hubs |
About the Flashcards
Flashcards for the Microsoft Azure AI App and Agent Developer Associate exam help students review practical machine learning and AI concepts used in Azure environments. This deck focuses on model selection, fine tuning, prompt tuning, data ingestion, preparation, evaluation metrics, and deployment considerations.
Students can also reinforce key ideas in responsible AI, including fairness, privacy, interpretability, governance, bias mitigation, and model monitoring. These cards are useful for reviewing terminology and understanding how production ML systems are evaluated, secured, explained, and maintained over time.
Topics covered in this flashcard deck:
- Fine tuning
- Prompt tuning
- Azure ML deployment
- Model evaluation
- Responsible AI
- Data drift monitoring