Azure AI-900 glossary of AI concepts, vision, language, and generative AI terms.
Use this glossary when AI-900 terms start to blur together. Keep it next to the lessons and the cheat sheet, not in place of them.
| Term | Short meaning | Fast exam anchor |
|---|---|---|
| AI workload | the main kind of problem being solved | classify the task before naming the service |
| classification | predict a category | label output |
| regression | predict a number | numeric output |
| clustering | discover groups without labels | unlabeled grouping |
| supervised learning | train from labeled examples | labels already exist |
| deep learning | neural-network-based ML for complex patterns | common in image, audio, and language tasks |
| transformer architecture | model design used heavily in modern language and generative systems | relevant to language and generative AI |
| feature | input field used by a model | what the model reads |
| label | known target used to train supervised learning | what the model learns to predict |
| training data | examples used to fit the model | learning stage |
| validation data | held-out data used to check generalization | reality check against memorization |
| overfitting | model looks good on training data but weak on new data | memorized instead of generalized |
| inference | using a trained model or service on new input | production or scoring stage |
| OCR | extract text from images or documents | read visual text |
| object detection | identify objects and where they are | what plus where |
| sentiment analysis | detect tone in text | positive, negative, neutral |
| entity recognition | extract names, places, dates, organizations, or similar items | pull structure from text |
| grounding | provide trusted context to a generative model at runtime | answer from supplied content |
| hallucination | unsupported or fabricated output | sounds plausible but wrong |
| model catalog | collection of model choices in the Foundry platform surface | compare available model options |
| responsible AI | principles for trustworthy AI use | fairness, safety, privacy, inclusiveness, transparency, accountability |
| Pair | Keep this distinction clear |
|---|---|
| classification vs regression | category vs number |
| classification vs clustering | labeled prediction vs unlabeled grouping |
| training vs inference | learning vs using |
| training data vs validation data | fitting vs checking generalization |
| feature vs label | input vs target |
| computer vision vs document processing | general image interpretation vs document reading and structure extraction |
| NLP vs speech | text-first workload vs audio-first workload |
| Azure AI Language vs Azure AI Speech | text analysis vs spoken-language services |
| Azure Machine Learning vs prebuilt Azure AI services | custom-model lifecycle vs ready-made AI capability |
| grounding vs fine-tuning | runtime context vs changing model behavior through training |
| Azure AI Foundry vs Azure OpenAI Service | broader generative-AI platform surface vs model-service access |
| fairness vs inclusiveness | outcome bias vs usability across different users and conditions |
| If you keep mixing up… | Use this anchor |
|---|---|
| Vision, document processing, and generative AI | analyze images, extract existing document content, and create new content |
| classification, regression, and clustering | category, number, and unlabeled grouping |
| training, validation, and inference | learn, test generalization, and use |
| Language, Speech, and Machine Learning | text analysis, audio processing, and predictive modeling |
| Azure AI Foundry, Azure OpenAI Service, and model catalog | platform surface, model service, and model-selection shelf |
| If the term is really about… | Revisit this page |
|---|---|
| workload choice and service fit | Cheat Sheet |
| exam status, retirement, and naming changes | FAQ |
| pacing and remediation sequence | Study Plan |
| official Microsoft links and live scope | Resources |