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
Noise Reduction
Noise reduction is the process of removing unwanted random variations (noise) from data to reveal the underlying meaningful patterns. In images, noise appears as grain or random pixel variations; in audio, as static or hiss; in text, as typos or irrelevant information. AI-powered noise reduction uses learned models to distinguish signal from noise, enabling clearer speech recognition, sharper image processing, and more accurate data analysis. Denoising autoencoders are a popular deep learning approach.
Object Detection
Object detection identifies and locates multiple objects within an image by drawing bounding boxes around them and assigning class labels with confidence scores. Unlike simple image classification that labels an entire image, detection pinpoints where each object is. Popular architectures include YOLO, Faster R-CNN, and DETR. Applications include autonomous driving, surveillance, quality inspection in manufacturing, and inventory management in retail.
On-Device AI
On-device AI refers to running AI models directly on end-user devices — smartphones, tablets, wearables — rather than in the cloud. This eliminates network latency, works offline, and keeps sensitive data on the device, improving privacy. Apple's on-device models, Google's Gemini Nano, and Samsung's Galaxy AI demonstrate this trend. Quantization and model compression make on-device inference practical, though models are necessarily smaller than their cloud counterparts.
Open-Source AI
Open-source AI refers to AI models, tools, and frameworks whose source code and/or model weights are publicly available, allowing anyone to use, modify, and deploy them without paying licensing fees. Llama, Mistral, and Stable Diffusion are prominent examples. Open-source AI democratizes access: a Moroccan startup can build products on top of a state-of-the-art model without depending on a single provider, though it must handle hosting, fine-tuning, and security itself.
Optical Character Recognition (OCR)
Optical character recognition converts text from images, scanned documents, or photographs into machine-readable text. Modern OCR powered by deep learning handles handwritten text, multiple languages, and poor-quality scans. A Moroccan government office processing thousands of handwritten identity forms daily can use OCR to digitize them automatically, dramatically reducing manual data entry.
Orchestration
Orchestration in AI refers to the coordinated management of multiple AI agents, tools, data sources, and human approval steps to execute a complex workflow. An orchestrator determines the order of operations, routes data between components, handles errors, and decides when human intervention is needed. Frameworks like LangGraph, CrewAI, and AutoGen provide orchestration capabilities. It is the glue that turns individual AI capabilities into end-to-end automated business processes.
Overfitting
Overfitting occurs when a model learns the training data too well — including its noise and quirks — and fails to generalize to new, unseen data. It memorizes rather than learns, acing training examples but failing on real-world inputs. Like a student who memorizes past exams word-for-word but cannot handle rephrased questions. Remedies include regularization, dropout, more training data, data augmentation, and early stopping.
Parameter-Efficient Fine-Tuning (PEFT)
Parameter-efficient fine-tuning (PEFT) is a family of techniques that adapt large pre-trained models to new tasks by updating only a small fraction of parameters, rather than all of them. Methods include LoRA, QLoRA, adapters, prefix tuning, and IA3. PEFT dramatically reduces the memory and compute needed for fine-tuning, making it possible for a startup to customize a billion-parameter model on a single GPU. It is now the default approach for adapting LLMs in production.
Perplexity
Perplexity is a measurement of how well a language model predicts a sample of text. Lower perplexity means the model is less «surprised» by the text, indicating it has learned the underlying patterns better. It is calculated as the exponent of the average negative log-likelihood per token. Perplexity is commonly used to compare language models during development, though it does not directly measure task performance or human preference. A model with perplexity of 10 is better calibrated than one with perplexity of 100.