The 212AY Library
AI
Glossary.
200+ artificial intelligence terms explained in plain language — from machine learning and LLMs to AI agents, RAG and prompt engineering. Every definition comes with a concrete example, in English, French and Arabic.
200 terms
Model Compression
Model compression encompasses techniques that reduce the size and computational requirements of a trained neural network while preserving as much accuracy as possible. Methods include quantization (reducing numerical precision), pruning (removing unnecessary weights), knowledge distillation (training a smaller model to mimic a larger one), and low-rank factorization. Compression makes it feasible to deploy capable AI models on smartphones, edge devices, and cost-effective servers.
Model Optimization
Model optimization encompasses techniques that improve a trained model's efficiency — reducing latency, memory usage, and computational cost — without significantly degrading prediction quality. Methods include quantization, pruning, knowledge distillation, operator fusion, and graph optimization. Optimization is critical for deploying models at scale: a model that takes 2 seconds per prediction may be too slow for a real-time customer service bot, but optimized to 50 milliseconds it becomes production-ready.
Model Registry
A model registry is a centralized system for tracking and managing AI models throughout their lifecycle — from development and experimentation through staging to production deployment and retirement. It stores model versions, training metadata, performance metrics, approval status, and deployment history. MLflow Model Registry, AWS SageMaker Model Registry, and Vertex AI Model Registry are popular implementations. A model registry is critical for governance and compliance, ensuring organizations know exactly which model version is serving which business process.
Model Serving
Model serving is the process of deploying a trained AI model behind an API endpoint so that applications can send requests and receive predictions in real time. It involves batching incoming requests for GPU efficiency, managing multiple model versions, scaling based on traffic, and monitoring latency. Platforms like NVIDIA Triton, TensorFlow Serving, and tools like BentoML and vLLM provide production-grade serving infrastructure.
Multimodal AI
Multimodal AI refers to models that can process and generate multiple types of data — text, images, audio, and sometimes video — within a single system. GPT-4V can analyze images and answer questions about them; Gemini processes text, images, and audio together. This breaks down silos between data types: a factory can show a camera image to a multimodal model and ask in text what defect it shows, getting a text answer referencing the visual.
Named Entity Recognition (NER)
Named entity recognition is an NLP task that automatically identifies and classifies key entities in text — names of people, organizations, locations, dates, monetary amounts — into predefined categories. It is the first step in many information extraction pipelines: a legal assistant can use NER to pull company names and contract dates from thousands of documents, and a healthcare system can extract patient names and drug mentions from clinical notes.
Natural Language Processing (NLP)
Natural language processing is the branch of AI focused on enabling computers to understand, interpret, and generate human language — both text and speech. NLP powers translation services, sentiment analysis, chatbots, email filtering, voice assistants, and document summarization. Recent advances in transformer models have dramatically improved NLP capabilities, making systems that can hold nuanced conversations, write code, and summarize complex legal documents in multiple languages.
Neural Architecture Search (NAS)
Neural architecture search automates the design of neural network architectures by using algorithms — often itself powered by machine learning — to explore a vast space of possible architectures and find ones that perform best on a given task. Instead of a human researcher deciding the number of layers, filter sizes, and connection patterns, NAS discovers them automatically. It has produced architectures that outperform human-designed ones, though the search process itself can be computationally expensive.
Neural Network
A neural network is a computational model inspired by the human brain, composed of layers of interconnected nodes (neurons) that process information by passing signals through weighted connections. Each layer extracts increasingly abstract features from the input. Neural networks are the foundation of deep learning, powering image recognition, speech synthesis, language translation, and game playing. The key insight is that with enough data and layers, neural networks can learn virtually any pattern.