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
Deep Learning
Deep learning is a subset of machine learning that uses neural networks with many layers to automatically learn hierarchical representations of data. Instead of hand-crafting features, deep learning models discover patterns directly from raw inputs — pixels for images, words for text, waveforms for audio. It powers breakthroughs in image recognition, speech synthesis, language translation, and autonomous driving. The trade-off: deep learning models require large datasets and significant computing power.
Diffusion Model
A diffusion model is a generative AI that learns to create data — typically images — by adding noise to training examples until they become pure static, then learning to reverse the process and remove noise step by step. This produces stunningly realistic images and is the foundation of Stable Diffusion, DALL-E, and Midjourney. Diffusion models have largely replaced GANs for image generation due to training stability and output quality.
Dimensionality Reduction
Dimensionality reduction compresses high-dimensional data into fewer features while preserving the most important information, making it easier to visualize, process, and model. PCA (Principal Component Analysis) is the classic linear method; t-SNE and UMAP are popular non-linear alternatives for visualization. A data scientist at a Moroccan telecom might reduce hundreds of customer-behavior features to a handful of components before clustering or building a churn model.
Distributed Training
Distributed training splits the computation for training a large AI model across multiple GPUs, machines, or data centers, dramatically reducing training time. Hundreds of GPUs work in parallel, each handling a portion of data or model layers. Frameworks like PyTorch's DistributedDataParallel and DeepSpeed manage coordination. Training GPT-4-class models would be impossible without distributed techniques.
Dropout
Dropout is a regularization technique for neural networks where randomly selected neurons are temporarily ignored during each training step, forcing the network to learn redundant representations rather than relying on any single neuron. It is like a team where random members sit out each practice session, making the remaining players more versatile. Dropout significantly reduces overfitting and is one of the most widely used tricks in deep learning, applied in vision, language, and recommendation models.
Edge AI
Edge AI refers to running AI models directly on local devices — smartphones, sensors, cameras, or IoT gateways — instead of sending data to a distant cloud server. This reduces latency, saves bandwidth, and preserves privacy because data never leaves the device. A Moroccan factory can run defect-detection models on cameras at the production line, getting instant results without relying on an internet connection. Quantized and distilled models make edge deployment practical.
Embedding Model
An embedding model converts inputs — words, sentences, images, or audio — into dense numerical vectors where semantically similar items cluster together. Popular models include OpenAI's text-embedding-ada-002, Sentence-BERT, and CLIP. These models power semantic search, recommendation systems, and RAG pipelines. The quality of embeddings directly determines downstream AI application performance.
Embeddings
Embeddings are dense numerical representations of words, sentences, images, or other data points, where similar items are mapped close together in a high-dimensional vector space. They capture semantic meaning: the embeddings for «king» and «queen» are closer to each other than to «car». Embeddings power semantic search, recommendation systems, and RAG pipelines. A Moroccan e-commerce site uses text embeddings to match customer queries to the most relevant product descriptions.
Ensemble Learning
Ensemble learning combines multiple models to produce predictions that are more accurate and robust than any single model alone. Random forests aggregate hundreds of decision trees; gradient boosting machines sequentially correct errors from previous models. The wisdom-of-crowds principle: each model has different weaknesses, so averaging or voting cancels out individual errors. Ensembles dominate Kaggle competitions and are widely used in credit scoring, fraud detection, and demand forecasting.