CONTACT US

AI & Machine Learning

This glossary brings together the AI and Machine Learning terms you’ll encounter most often  from foundational concepts like algorithms and neural networks to advanced topics like generative AI and large language models. Each entry gives a clear, practical definition so you can quickly understand how a term is used in real-world AI/ML applications, without needing a technical background.

# A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

#

  • 3D Convolutional Neural Network A neural network variant applying convolution across three spatial dimensions — used for video analysis, medical imaging, and volumetric data processing.
  • 4-Layer AI Architecture A reference model organizing AI systems into data, model, application, and governance layers to guide enterprise AI design.
  • 4+1 View Model (AI Systems) A documentation approach for AI systems covering logical, development, process, physical, and scenario views.
  • 8-Bit Quantization A model compression technique reducing neural network weight precision from 32-bit floats to 8-bit integers — cutting model size and inference cost with minimal accuracy loss.

A

  • Agent (AI Agent) An autonomous software entity that perceives its environment, makes decisions, and takes actions to achieve a goal — often using LLMs as a reasoning backbone combined with tools and memory.
  • Agentic AI AI systems capable of autonomously planning, deciding, and executing multi-step tasks with minimal human intervention — using tools, memory, and reasoning loops to complete complex goals.
  • AI Bias Systematic errors in AI model outputs caused by skewed training data, flawed model design, or unrepresentative sampling.
  • AI Governance The framework of policies, standards, and controls ensuring AI systems are used responsibly, transparently, and in compliance with regulations.
  • AI Hallucination When a generative AI model produces confident-sounding but factually incorrect or fabricated outputs.
  • AI Inference The process of running a trained AI model on new input data to generate predictions or outputs in production.
  • AI Model A mathematical system trained on data to recognize patterns and make predictions or generate outputs for a defined task.
  • AI Orchestration Coordinating multiple AI models, tools, data sources, and agents into a unified workflow that accomplishes a complex task.
  • AI-Augmented Development AI-Augmented Development integrates AI tools to amplify human engineering productivity by automating routine tasks, allowing focus on high-value architectural work. This approach accelerates software delivery through agent-led engineering ecosystems that span the entire development lifecycle, from planning to deployment.
  • AI-Native Development A software development approach where AI capabilities — LLMs, embeddings, agents, and ML models — are designed into the product architecture from day one, rather than added as features after the fact.
  • Algorithm A defined sequence of computational rules or instructions that a system follows to solve a problem or make a decision.
  • Anomaly Detection An ML technique identifying patterns or behaviors that deviate significantly from expected norms — used in fraud detection, predictive maintenance, and monitoring.
  • Attention Mechanism A neural network component allowing a model to focus on the most relevant parts of an input — the core building block of transformer architecture.
  • AutoML Automated Machine Learning — automating model selection, training, and tuning so non-experts can build predictive solutions without deep data science expertise.

B

  • Backpropagation The algorithm used to train neural networks by computing gradients and adjusting weights layer by layer to minimize prediction error.
  • Batch Inference Running predictions on a large dataset all at once rather than in real time — used for high-volume prediction tasks such as churn scoring or demand forecasting.
  • Bayesian Inference A statistical method for updating probability estimates as new evidence arrives — used in ML for uncertainty quantification, A/B testing, and probabilistic forecasting.
  • BERT A pre-trained language model from Google reading text bidirectionally — enabling high-accuracy NLP tasks such as document classification and semantic search.
  • Bias (ML Bias) Systematic error in model predictions caused by flawed assumptions, unrepresentative training data, or poor feature engineering — leading to unfair or inaccurate outcomes.
  • Bias-Variance Tradeoff The tension between a model that overfits training data (high variance) and one too simple to capture patterns (high bias) — central to model tuning.

C

  • Chain-of-Thought Prompting A prompting technique guiding a language model to reason step-by-step before producing a final answer — improving accuracy on complex reasoning tasks.
  • Classification Classification is a supervised machine learning approach categorizing data into predefined classes. Learn how it automates enterprise decision-making.
  • Classification Model A supervised ML model assigning input data to predefined categories - such as spam/not-spam, risk levels, or product category labels.
  • Clustering An unsupervised ML technique grouping similar data points together without predefined labels - used for segmentation and pattern discovery.
  • Computer Vision A field of AI enabling machines to interpret visual information from images or video - including object detection, image classification, and OCR.
  • Confusion Matrix A quantitative performance evaluation table used in machine learning classification that compares a model's predicted categories against the actual ground-truth labels.
  • Contextual AI AI systems incorporating external context - user history, session data, business rules - into their reasoning for more relevant situation-aware outputs.
  • Convolutional Neural Network (CNN) A deep learning architecture specialized for processing grid - structured data like images, using convolutional filters to detect features.

D

  • Data Augmentation Techniques artificially expanding training datasets through transformations - flipping images, paraphrasing text - to improve model generalization.
  • Data Labeling The process of annotating raw data - images, text, audio - with tags that supervised ML models use as ground truth during training.
  • Deep Learning A subset of ML using multi - layered neural networks to learn hierarchical representations - enabling breakthroughs in image recognition, NLP, and generative AI.
  • Diffusion Model A generative AI architecture creating new data by learning to reverse a noise-adding process - the basis of image generation tools like Stable Diffusion and DALL-E.
  • Drift (Model Drift) The degradation of ML model performance over time as real-world data shifts away from training data - requiring monitoring and retraining strategies.

E

  • Embedding A dense numerical vector representation of data — text, images, entities — in a continuous space where semantic similarity corresponds to vector proximity.
  • Ensemble Learning A technique combining predictions from multiple models for more accurate results - e.g., Random Forest, Gradient Boosting.
  • Explainable AI (XAI) Methods making AI model decisions interpretable to humans - enabling auditors, regulators, and business users to understand why a model produced a given output.
  • eXplainable Gradient Boosting A variant of gradient boosting that adds SHAP values or local explanations - making tree-based ensemble predictions interpretable.

F

  • Feature Engineering The process of selecting, transforming, and creating input variables from raw data to improve ML model performance.
  • Feature Store A centralized repository for storing, managing, and serving ML features for model training and real-time inference.
  • Fine-Tuning Adapting a pre-trained AI model to a specific domain by continuing training on a smaller, task-specific dataset.
  • Foundation Model A large-scale AI model trained on vast data that can be adapted to a wide range of tasks — including GPT-4, Claude, Gemini, and Llama.

G

  • Generative AI (GenAI) A category of AI creating new content — text, images, code, audio, video — by learning patterns from training data and generating novel outputs from prompts.
  • GPT (Generative Pre-trained Transformer) A family of large language models from OpenAI generating coherent text based on prompts — powering chatbots, code generation, and document intelligence.
  • Gradient Descent An optimization algorithm training ML models by iteratively adjusting parameters in the direction that minimizes prediction error.
  • Graph Neural Network (GNN) A neural network operating on graph-structured data — nodes and edges — used for recommendation systems, fraud detection, and knowledge graphs.
  • Ground Truth The verified, accurate reference data used to train and evaluate ML models — determining how closely model outputs align with correct real-world answers.

H

  • Hallucination (AI) A phenomenon where a large language model generates plausible-sounding but factually incorrect information — a key risk in enterprise AI deployments.
  • Human-in-the-Loop (HITL) An AI design pattern incorporating human review or approval at key decision points — balancing automation efficiency with human oversight.
  • Hyperparameter Tuning Optimizing configuration settings of an ML algorithm — such as learning rate or tree depth — to improve model performance.

I

  • Image Classification A computer vision task assigning a label to an entire image based on its dominant content.
  • Image Recognition AI capability identifying objects, scenes, or attributes within images with high accuracy.
  • Inference Using a trained ML model to generate predictions on new, unseen data — as opposed to the training phase where the model learns.
  • Inference Engine The runtime system executing a trained AI model against new input data to produce predictions or outputs in production.
  • Instruction Tuning A fine-tuning technique training LLMs on pairs of instructions and ideal responses to improve task-following behavior.
  • Intelligent Automation Combining AI capabilities with RPA to automate complex, judgment-intensive business processes.
  • Intent Recognition An NLP capability identifying the purpose behind a user's input — fundamental to chatbots, virtual assistants, and conversational AI.

J

  • Jailbreak Detection Techniques identifying and blocking adversarial prompts attempting to bypass AI safety constraints or extract restricted information.
  • Jensen’s Inequality (ML) A mathematical property used in probabilistic ML models — particularly in variational inference and loss function design.
  • Joint Probability Model A statistical model capturing the likelihood of multiple variables occurring together — used in Bayesian networks and recommendation systems.
  • JSON Mode (LLM) A prompting or API configuration forcing an LLM to respond exclusively in valid, parseable JSON — enabling reliable downstream processing.
  • JSON Schema Validation (AI Outputs) Enforcing structured output formats from LLMs using JSON schemas to ensure reliability and downstream parsability.
  • Just-in-Time Inference A model serving approach loading AI models on demand rather than keeping them persistently in memory — reducing idle compute cost.
  • Just-in-Time Training An ML deployment strategy where models are retrained incrementally as new data arrives rather than periodically from scratch.

K

  • K-Means Clustering An unsupervised ML algorithm grouping data points into K clusters based on feature similarity — used for customer segmentation and anomaly grouping.
  • K-Nearest Neighbor (KNN) A simple ML algorithm classifying inputs based on the labels of their closest data points in feature space.
  • Knowledge Base (AI) A structured repository of domain-specific information that an AI system retrieves from to answer questions accurately — the backbone of RAG-based enterprise AI.
  • Knowledge Distillation A model compression technique where a smaller student model is trained to replicate a larger teacher model — reducing inference cost while preserving accuracy.
  • Knowledge Graph A structured representation of entities and their relationships, enabling AI systems to reason across connected information.

L

  • LangChain An open-source framework for building LLM-powered applications with tools, memory, agents, and retrieval components.
  • Large Language Model (LLM) A neural network trained on massive text datasets capable of understanding, generating, and reasoning about natural language — including GPT-4, Claude, Gemini, and Llama.
  • Latent Space A compressed, lower-dimensional representation of data learned by a neural network where semantically similar inputs are positioned close together.
  • Linear Regression A foundational statistical ML algorithm modeling the linear relationship between input variables and a continuous output.
  • LLM Fine-Tuning Adapting a large language model to a specific domain or task using supervised training on curated datasets.
  • Loss Function A mathematical function measuring how far model predictions are from ground truth during training — guiding the optimization process.

M

  • Machine Learning (ML) A branch of AI where systems learn to perform tasks by detecting patterns in data rather than being explicitly programmed with rules.
  • MLOps The discipline applying DevOps principles to machine learning — automating model training, deployment, monitoring, and retraining at scale.
  • Model Registry A centralized repository for storing, versioning, and managing trained ML models — enabling reproducibility, auditing, and safe production deployment.
  • Multi-Modal AI An AI system processing and integrating multiple input types — text, images, audio, video — to produce richer, context-aware outputs.
  • Named Entity Recognition (NER) An NLP task identifying and classifying specific entities — names, dates, locations, organizations — within unstructured text.

N

  • Named Entity Recognition (NER) An NLP task identifying and classifying specific entities — names, dates, locations, organizations — within unstructured text.
  • Natural Language Processing (NLP) A field of AI enabling computers to understand, interpret, and generate human language — powering chatbots, translation, summarization, and sentiment analysis.
  • Neural Network A computational model consisting of interconnected layers of nodes that learn to recognize patterns through training.

O

  • Object Detection Computer vision task locating and classifying multiple objects within an image or video frame simultaneously.
  • Online Learning An ML paradigm where models are updated continuously as new data arrives in real time rather than being retrained periodically.
  • ONNX (Open Neural Network Exchange) An open format for representing ML models — enabling models trained in one framework to be deployed in another.
  • Orchestration (AI) The coordination of multiple AI agents, models, or tools within a workflow — managing sequencing, state, error handling, and output routing.
  • Output Validation Automated checking of AI-generated outputs against defined quality, safety, and format criteria before delivery to end users.
  • Overfitting A condition where a model learns training data too precisely including its noise — performing poorly on new, unseen data.
  • Overfitting Prevention Techniques reducing the tendency of ML models to memorize training data rather than learn generalizable patterns.

P

  • Pre-trained Model An ML model trained on a large dataset before adaptation to a specific task — allowing organizations to leverage powerful AI without building from scratch.
  • Predictive Analytics Using statistical models and ML to forecast future events or behaviors — such as churn prediction, demand forecasting, or risk scoring.
  • Prompt Engineering The practice of designing and optimizing input prompts to guide LLMs toward accurate, reliable, and contextually appropriate outputs.
  • PyTorch An open-source deep learning framework widely used for model research, training, and production deployment.

Q

  • Q-Learning A model-free reinforcement learning algorithm where an agent learns to maximize cumulative reward by building a Q-value table through experience.
  • Quantization (ML) A model optimization technique reducing numerical precision of model weights — making models smaller and faster, especially on edge devices.
  • Quantum Machine Learning An emerging field applying quantum computing principles to accelerate ML algorithms beyond classical compute limits.
  • Query Expansion Automatically augmenting a user's search query with related terms to improve recall in semantic search systems.
  • Query Routing Directing AI queries to the most appropriate model or retrieval source based on intent, complexity, or domain.
  • Query Understanding (AI) The process by which an AI system interprets intent, entities, and context behind a user query to return the most relevant response.

R

  • RAG (Retrieval-Augmented Generation) An AI architecture combining a retrieval system with a generative model — fetching relevant documents before generating a response to reduce hallucination.
  • Random Forest An ensemble ML algorithm building multiple decision trees and combining their outputs for improved accuracy and robustness.
  • Regression Model A supervised ML model predicting a continuous numeric output — such as revenue, customer lifetime value, or failure probability.
  • Reinforcement Learning (RL) An ML paradigm where an agent learns by interacting with an environment, receiving rewards for correct actions and penalties for incorrect ones.

S

  • Semantic Search A search approach understanding meaning and intent behind a query — using embeddings and vector similarity to return contextually relevant results.
  • Sentiment Analysis An NLP technique classifying the emotional tone of text — positive, negative, or neutral — enabling systematic analysis of customer feedback at scale.
  • Supervised Learning An ML approach where models are trained on labeled input-output pairs — learning to map inputs to correct outputs for prediction tasks.
  • Synthetic Data Artificially generated data mimicking real-world statistical properties — used to augment training datasets, protect privacy, and simulate rare events.

T

  • TensorFlow An open - source ML framework by Google widely used for training and deploying deep learning models at scale.
  • Time Series Forecasting ML techniques predicting future values based on historical sequential data - demand forecasting, anomaly detection, financial projection.
  • Token (LLM) The basic unit of text processed by an LLM — roughly 3/4 of a word in English — used to measure input/output length and calculate API costs.
  • Transfer Learning A technique where a model trained on one task is adapted for a related task - reducing training data and compute requirements.
  • Transformer The neural network architecture underpinning modern LLMs - using self-attention to process sequences in parallel, enabling unprecedented language understanding.
  • Transformer Architecture A neural network design based on self-attention mechanisms - the foundational architecture behind GPT, BERT, and most modern LLMs.

U

  • Uncertainty Quantification Methods measuring and communicating how confident an AI model is in its predictions - critical for high-stakes decision support.
  • Underfitting A condition where an ML model is too simple to capture underlying patterns - resulting in poor performance on both training and test data.
  • Unsupervised Learning An ML paradigm where models find patterns in data without labeled examples - used for segmentation, anomaly detection, and dimensionality reduction.
  • Unsupervised Pre-Training Training a model on unlabeled data to learn general representations before fine-tuning on a specific task.
  • User Behavior Modeling Using ML to represent and predict how users interact with a system - including click patterns, purchase sequences, and engagement signals.
  • User Intent Classification An NLP task automatically identifying the goal behind a user's message - routing it to the correct handler in AI systems.

V

  • Validation Set A subset of training data held out to tune hyperparameters and evaluate model performance before final testing — preventing overfitting to the test set.
  • Variational Autoencoder (VAE) A generative AI model learning a compressed latent representation of data and generating new samples by sampling from that space.
  • Vectorization Converting non-numeric data — text, images, categories — into numerical vectors that ML models can process.

W

  • Weight Initialization The strategy for setting neural network weights at the start of training — critically affecting convergence speed and final performance.
  • Word2Vec An embedding technique representing words as dense vectors based on co-occurrence — enabling semantic relationships to be captured mathematically.

X

  • XGBoost An optimized gradient boosting algorithm known for high accuracy and speed on structured data — one of the most widely used algorithms in enterprise ML.
  • XML-Based Model Exchange (PMML) The use of Predictive Model Markup Language to serialize and exchange trained ML models between different tools and platforms.

Y

  • YAML Config (MLOps) Using YAML files to define ML pipeline configurations, hyperparameters, and experiment settings in a version-controlled, reproducible format.
  • YAML-Defined ML Pipelines Defining ML training, evaluation, and deployment workflows in YAML configuration files — enabling reproducible, version-controlled MLOps pipelines.
  • Yield Optimization (ML) Applying ML to maximize output, efficiency, or revenue across variable conditions — used in pricing, manufacturing, and digital advertising.
  • Yield Prediction (ML) Using ML models to forecast output volumes in manufacturing, agriculture, or financial instruments.

Z

  • Z-Score Normalization A data preprocessing technique standardizing features to zero mean and unit variance — ensuring no single feature dominates model training.

Need help implementing , not just understanding them?

Kyanon Digital helps enterprises design and build scalable customer experience platforms, from Omnichannel strategy to personalized AI-delivery.

  • contact@kyanon.digital

How can we help you?

    Drop us a line! We are here to answer your questions 24/7.


    Create project brief with AICreate project brief with AI