Complete AI Terminology: A to Z Glossary & Definitions

Artificial Intelligence (AI) uses specialized vocabulary spanning computer science, statistics, and hardware engineering. This reference glossary defines key terminology across machine learning, neural architectures, generative models, and hardware acceleration to help developers, creators, and enthusiasts navigate modern AI concepts.

Direct Definition (Quick Answer Box) Artificial Intelligence Terminology refers to the standardized lexicon of concepts, mathematical algorithms, neural architectures, and hardware components used to design, train, and deploy machine learning models and autonomous systems.

Core AI & Machine Learning Foundations

• Artificial Intelligence (AI): The broad field of computer science focused on building machines capable of performing tasks that typically require human cognition.

• Machine Learning (ML): A subfield of AI where algorithms discover patterns from data to improve task accuracy without being explicitly coded.

• Deep Learning (DL): A subset of ML utilizing multi-layered artificial neural networks to process complex, unstructured data (text, images, audio).

• Artificial Narrow Intelligence (ANI): AI specialized for a single dedicated domain (e.g., chess engines, spam filtering, computer vision).

• Artificial General Intelligence (AGI): Hypothetical AI capable of understanding, learning, and executing any cognitive task at human capability.

• Artificial Superintelligence (ASI): Theoretical AI that exceeds human cognitive abilities across all domains.

• Agentic AI: Autonomous software systems designed to execute multi-step plans, use tools, and make decisions to complete defined objectives.

Neural Architectures & Modeling

• Artificial Neural Network (ANN): Computing systems inspired by biological brains, made of interconnected nodes (neurons) structured in layers.

• Transformer: A neural network architecture utilizing self-attention mechanisms to process sequence data simultaneously rather than sequentially.

• Large Language Model (LLM): High-parameter deep learning models trained on massive text corpora to understand and generate natural language.

• Diffusion Model: Generative models that synthesize data (images, audio) by iteratively removing noise from a random signal.

• Weights & Biases: Learnable mathematical parameters inside neural networks that adjust during training to map inputs to accurate outputs.

• Activation Function: Mathematical functions (e.g., ReLU, Sigmoid, GELU) determining whether a neuron activates, introducing non-linearity to the network.

Training Paradigms & Data Processing

• Supervised Learning: Training models using labeled input-output dataset pairs.

• Unsupervised Learning: Training models on unlabeled data to discover hidden patterns, clusters, or distributions independently.

• Reinforcement Learning (RL): Training an agent via trial-and-error interactions with an environment using reward and penalty signals.

• RLHF (Reinforcement Learning from Human Feedback): Aligning language model behavior using human evaluations to score and refine outputs.

• Fine-Tuning: Training a pre-trained model further on a domain-specific dataset to adapt it for specialized tasks.

• Overfitting / Underfitting: Overfitting occurs when a model memorizes training noise and fails on unseen data; underfitting occurs when a model is too simple to learn underlying patterns.

• Gradient Descent & Backpropagation: Core optimization algorithms used to calculate error gradients and update neural network weights during training.

Generative AI, Prompting & Inference

• Generative AI: Models capable of synthesizing new, original content (text, imagery, synthetic voice, code).

• Token / Tokenization: The atomic units (words, characters, subwords) into which text is broken down for LLM processing.

• Context Window: The maximum number of tokens a language model can process and retain across a single input-output sequence.

• Embeddings & Vector Space: Numerical vector arrays that represent the semantic meaning of text, images, or audio in high-dimensional space.

• Retrieval-Augmented Generation (RAG): Connecting an LLM to external vector databases or search indexes to retrieve real-time facts before generating an answer.

• Hallucination: When an AI model generates factually incorrect, fabricated, or nonsensical output with high confidence.

• Temperature / Top-P: Sampling hyperparameters that control the randomness, creativity, and predictability of model generations.

AI Hardware & Deployment Terms

• Inference: The operational phase where a fully trained model takes new input and computes a live prediction or generation.

• Quantization: Compressing model weights (e.g., FP16 to INT8 or INT4) to lower memory footprint and enable local hardware execution.

• NPU (Neural Processing Unit): Specialized silicon optimized for low-power matrix multiplication and on-device AI acceleration.

• TPU (Tensor Processing Unit): Custom ASIC hardware designed specifically for high-throughput tensor operations and machine learning workloads.