📊 AI & Data Science Terminology Glossary: 55 Essential Terms Explained

Why These Terms Matter

Data science and applied AI have become part of everyday engineering practice — predictive maintenance models, energy forecasting tools, computer-vision inspection systems, and generative-AI copilots all rest on a shared vocabulary that most engineers never studied formally. Unlike a legal code with a single authoritative definitions article, data science terminology comes from statistics, computer science, and industry practice, and the same word is sometimes used loosely in marketing material and precisely in a technical paper. Knowing the precise meaning matters: confusing "accuracy" with "precision," or not understanding what "overfitting" actually means, leads to models that look good in a demo and fail in production.

This glossary covers 55 of the most important terms an engineer will encounter when reading a data science paper, evaluating a vendor's AI claims, or building their own predictive model — organized alphabetically with plain-language definitions and the practical context in which each term shows up.

A

Accuracy — Classification metric

The fraction of predictions a model gets correct: (correct predictions) ÷ (total predictions). Accuracy is intuitive but misleading on imbalanced datasets — a model that predicts "no fault" for every data point on a machine that fails 1% of the time is 99% accurate and completely useless. This is why precision, recall, and F1 score are used alongside accuracy for imbalanced problems like fault detection.

Activation Function — Neural network component

A mathematical function applied to a neuron's weighted input sum that introduces non-linearity, allowing a neural network to learn complex patterns rather than just linear relationships. Common activation functions include ReLU (rectified linear unit, which outputs zero for negative inputs and the input value otherwise), sigmoid (which squashes output between 0 and 1, useful for probabilities), and softmax (used in the output layer for multi-class classification).

AI Agent — System architecture term

A software system that uses an LLM (large language model) to decide which actions to take — calling tools, querying databases, writing code — in a loop, rather than producing a single one-shot response. Agents differ from simple chatbots because they can plan multi-step tasks, use external tools, and evaluate whether a step succeeded before proceeding. Agentic workflows are increasingly used for engineering tasks like automated code review or drafting calculation reports.

Anomaly Detection — Machine learning application

The task of identifying data points that deviate significantly from the expected pattern, without necessarily having labeled examples of every possible anomaly. In engineering, anomaly detection is widely used for predictive maintenance — flagging vibration signatures, temperature trends, or power-quality readings that differ from a piece of equipment's normal operating baseline before a hard failure occurs.

Artificial Intelligence (AI) — Field of study

The broad field of building systems that perform tasks normally requiring human intelligence — perception, reasoning, language understanding, decision-making. Machine learning is a subset of AI in which systems learn patterns from data rather than following explicitly programmed rules; deep learning is a subset of machine learning using multi-layer neural networks.

B

Backpropagation — Training algorithm

The algorithm used to train neural networks by computing the gradient of the loss function with respect to each weight, working backward from the output layer to the input layer using the chain rule of calculus. Backpropagation combined with gradient descent is how a neural network's weights are updated during training to reduce prediction error.

Bias-Variance Tradeoff — Model theory concept

The fundamental tension in model design between bias (error from overly simplistic assumptions, causing underfitting) and variance (error from being overly sensitive to training data noise, causing overfitting). A high-bias model is too rigid to capture real patterns; a high-variance model captures training-data noise as if it were signal. The goal is a model complex enough to capture the true pattern but not so complex that it memorizes the training set.

Big Data — Data characteristic

Datasets too large or complex for traditional data-processing tools, typically characterized by the "three Vs": volume (scale of data), velocity (speed of generation, e.g., sensor streams), and variety (structured, unstructured, and semi-structured formats combined). Building sensor networks and SCADA historians commonly generate big data that requires distributed processing frameworks rather than a single spreadsheet.

C

Classification — Supervised learning task

A supervised learning task where the model predicts a discrete category label — "pass/fail," "healthy/faulty," "high/medium/low risk" — rather than a continuous number. Binary classification has two classes; multiclass classification has three or more.

Clustering — Unsupervised learning task

An unsupervised learning technique that groups data points into clusters based on similarity, without any labeled "correct answer" provided during training. K-means is the most common clustering algorithm. Engineers use clustering to segment equipment into behavior groups or identify distinct operating regimes in sensor data without predefined categories.

Confusion Matrix — Evaluation tool

A table that breaks down a classifier's predictions into four categories: true positives (correctly predicted positive), true negatives (correctly predicted negative), false positives (incorrectly predicted positive — a "false alarm"), and false negatives (incorrectly predicted negative — a "miss"). The confusion matrix is the foundation from which precision, recall, and F1 score are calculated, and it reveals failure modes that a single accuracy number hides.

Cross-Validation — Model validation technique

A technique for estimating how well a model will generalize to new data by repeatedly splitting the dataset into training and validation subsets — most commonly "k-fold" cross-validation, which splits the data into k equal parts, trains on k-1 of them, and validates on the remaining part, rotating through all k combinations. Cross-validation gives a more reliable performance estimate than a single train/test split, especially with limited data.

D

Data Drift — Model monitoring concept

The phenomenon where the statistical properties of production input data change over time relative to the data a model was trained on, causing prediction quality to degrade even though the model itself hasn't changed. A predictive maintenance model trained on one fleet of pumps may drift in accuracy when deployed on a fleet with different operating conditions or after equipment ages. Monitoring for data drift is a core MLOps practice.

Deep Learning — Machine learning subfield

A subset of machine learning using neural networks with many layers ("deep" networks) capable of automatically learning hierarchical feature representations from raw data — pixels, audio waveforms, or text — without manual feature engineering. Deep learning powers computer vision, speech recognition, and large language models.

Dimensionality Reduction — Preprocessing technique

The process of reducing the number of input variables (features) in a dataset while preserving as much meaningful information as possible. Principal Component Analysis (PCA) is the most common technique. Dimensionality reduction speeds up training, reduces overfitting risk, and helps visualize high-dimensional sensor data in two or three dimensions.

E

Embedding — Representation technique

A numerical vector representation of a piece of data (a word, sentence, image, or document) in a continuous multi-dimensional space, positioned so that semantically similar items end up close together in that space. Embeddings are the foundation of modern search, recommendation, and retrieval-augmented generation (RAG) systems — a technical manual query is converted to an embedding and compared against embeddings of manual passages to find the most relevant match.

Epoch — Training term

One complete pass of the training algorithm through the entire training dataset. Neural networks are typically trained over many epochs, with model weights updated incrementally after each batch of data within an epoch. Training for too many epochs on a fixed dataset risks overfitting; too few risks underfitting.

Ensemble Method — Modeling technique

A technique that combines predictions from multiple models to produce a result more accurate and robust than any single model alone. Random forests (an ensemble of decision trees) and gradient boosting (which builds trees sequentially, each correcting the previous one's errors) are the most widely used ensemble methods in applied engineering data science.

F

F1 Score — Classification metric

The harmonic mean of precision and recall, giving a single number that balances both concerns: F1 = 2 × (precision × recall) ÷ (precision + recall). F1 score is preferred over raw accuracy for imbalanced classification problems (like rare-fault detection) because it penalizes models that achieve high accuracy purely by predicting the majority class.

Feature — Data science term

An individual measurable input variable used by a model to make a prediction — for example, motor current, ambient temperature, and vibration RMS could each be a feature in a predictive maintenance model. The complete set of features for one observation is often called a feature vector.

Feature Engineering — Modeling practice

The process of creating, transforming, or selecting input features to improve a model's predictive performance — for example, converting a raw timestamp into "hour of day" and "day of week," or computing a rolling average of sensor readings instead of using instantaneous values. Feature engineering often has more impact on model performance than the choice of algorithm itself, especially in engineering domains where domain expertise reveals which derived quantities actually matter physically.

Fine-Tuning — LLM adaptation technique

The process of further training a pre-trained model (typically a large language model or deep neural network) on a smaller, domain-specific dataset to adapt its behavior for a specialized task, rather than training a model from scratch. Fine-tuning a general LLM on a company's engineering standards and past project reports can make it more accurate for domain-specific question answering.

G

Generative AI — AI category

AI systems that create new content — text, images, code, audio — rather than only classifying or predicting from existing categories. Large language models (LLMs) like GPT and Claude, and image generators like diffusion models, are generative AI. Generative AI is increasingly used in engineering for drafting calculation narratives, summarizing RFIs, and generating boilerplate code or documentation.

Gradient Descent — Optimization algorithm

An iterative optimization algorithm that adjusts a model's parameters step by step in the direction that most reduces the loss function, using the gradient (slope) of the loss with respect to each parameter. The "learning rate" controls how large each step is — too large and training becomes unstable; too small and training takes too long to converge.

H

Hallucination — LLM failure mode

When a generative AI model produces output that is fluent and confident-sounding but factually incorrect or fabricated — for example, an LLM citing a nonexistent code section or inventing a plausible-sounding but wrong equation. Hallucination is a known limitation of LLMs and is a primary reason human review of AI-generated engineering content is required before it is used in a deliverable.

Hyperparameter — Model configuration term

A configuration setting chosen before training begins that controls how a model learns, as opposed to a parameter (like a neural network weight) that the model learns from data. Examples include learning rate, number of decision trees in a random forest, and the number of layers in a neural network. Hyperparameter tuning — searching for the best combination — is a standard part of the model development process.

I

Imbalanced Dataset — Data characteristic

A dataset where one class vastly outnumbers another — for example, a predictive maintenance dataset with 10,000 "normal operation" records and 50 "failure" records. Imbalanced datasets require special handling (resampling, class weighting, or metrics like F1 score and precision-recall curves instead of accuracy) because standard training tends to ignore the rare, often more important, minority class.

Inference — Deployment term

The process of using an already-trained model to make predictions on new, previously unseen data, as opposed to "training," which is the process of learning the model's parameters from historical data. A model running in production evaluating live sensor data is performing inference.

L

Label — Supervised learning term

The known, correct output value associated with a training example in supervised learning — for example, "failed within 30 days" or "did not fail" attached to historical equipment records. Labels are what the model learns to predict; unlabeled data (used in unsupervised learning) has no such ground truth attached.

Large Language Model (LLM) — Model architecture

A deep learning model, typically built on the transformer architecture, trained on massive amounts of text to predict the next word (token) in a sequence, which gives it the ability to generate coherent text, answer questions, summarize documents, and write code. GPT, Claude, and Gemini are examples of LLMs.

Loss Function — Training component

A mathematical function that quantifies how far a model's predictions are from the true values — the number that gradient descent works to minimize during training. Mean squared error is a common loss function for regression; cross-entropy loss is common for classification.

M

Machine Learning (ML) — Field of study

A subset of AI in which systems learn patterns and make predictions from data rather than being explicitly programmed with rules. ML encompasses supervised learning (learning from labeled examples), unsupervised learning (finding patterns in unlabeled data), and reinforcement learning (learning through trial-and-error feedback from an environment).

MLOps — Practice/discipline

Short for "machine learning operations" — the set of practices for deploying, monitoring, and maintaining machine learning models reliably in production, analogous to DevOps for traditional software. MLOps covers model versioning, automated retraining pipelines, data drift monitoring, and rollback procedures when a deployed model's performance degrades.

Model — Core artifact

The trained mathematical artifact — a set of learned parameters plus an algorithm structure — that takes input data and produces a prediction or output. "Training a model" means using data and an optimization algorithm to find the parameter values that minimize prediction error.

Model Drift — Monitoring term

Also called concept drift — the degradation of a deployed model's predictive performance over time because the real-world relationship between inputs and outputs has changed, distinct from data drift where only the input distribution shifts. A model predicting energy demand may experience concept drift after a building's occupancy pattern changes structurally, even if sensor readings look statistically similar.

N

Natural Language Processing (NLP) — AI subfield

The subfield of AI concerned with enabling computers to understand, interpret, and generate human language, covering tasks like text classification, sentiment analysis, machine translation, and the language modeling underlying modern LLMs.

Neural Network — Model architecture

A machine learning model loosely inspired by the structure of biological neurons, composed of layers of interconnected nodes ("neurons") that each apply a weighted sum and activation function to their inputs. Data flows from an input layer through one or more hidden layers to an output layer; the network learns by adjusting connection weights through backpropagation.

O

Overfitting — Model failure mode

When a model learns the training data too well — including its noise and idiosyncrasies — such that it performs excellently on training data but poorly on new, unseen data. Overfitting is the single most common failure mode in applied machine learning and is why a train/validation/test split, cross-validation, and regularization techniques are standard practice.

Outlier — Data characteristic

A data point that differs significantly from the rest of a dataset, either due to genuine rare events, sensor error, or data entry mistakes. Outliers can badly skew a model's training if not identified and handled deliberately — whether by removal, capping, or explicit modeling as an anomaly-detection target.

P

Precision — Classification metric

Of all the instances a model predicted as positive, the fraction that were actually positive: true positives ÷ (true positives + false positives). High precision means few false alarms. In a fault-detection system, high precision matters when false alarms are costly (e.g., unnecessary shutdowns for inspection).

Predictive Maintenance — Application area

The use of sensor data and machine learning models to predict when equipment is likely to fail, allowing maintenance to be scheduled proactively rather than performed on a fixed calendar (preventive) or only after breakdown (reactive). Predictive maintenance typically relies on anomaly detection and regression models trained on vibration, temperature, current, or acoustic sensor data.

Prompt Engineering — LLM interaction practice

The practice of crafting the input text (the "prompt") given to a large language model to reliably produce useful, accurate output — including providing context, examples, explicit constraints, and step-by-step instructions. Effective prompt engineering is increasingly treated as a core engineering skill for using AI copilots productively and safely.

R

RAG (Retrieval-Augmented Generation) — Architecture pattern

An architecture that improves LLM accuracy by first retrieving relevant documents or passages from a knowledge base (using vector embeddings and similarity search) and then supplying that retrieved content to the LLM as context before it generates its answer. RAG reduces hallucination and allows an LLM to answer questions about proprietary or up-to-date content — such as a company's own engineering standards — that it was never trained on directly.

Recall — Classification metric

Of all the instances that were actually positive, the fraction the model correctly identified: true positives ÷ (true positives + false negatives). Also called sensitivity or true positive rate. High recall matters when missing a positive case is costly — for example, failing to flag a genuine equipment fault.

Regression — Supervised learning task

A supervised learning task where the model predicts a continuous numeric value — remaining useful life in days, expected energy demand in kW, or predicted settlement in inches — rather than a discrete category. Linear regression is the simplest form; more complex models (random forests, gradient boosting, neural networks) can capture nonlinear relationships.

Regularization — Overfitting-prevention technique

A set of techniques that discourage a model from becoming overly complex or overly reliant on any single feature, in order to reduce overfitting — including L1/L2 penalty terms added to the loss function, dropout in neural networks, and early stopping of training. Regularization trades a small amount of training-set fit for better generalization to new data.

Reinforcement Learning — Learning paradigm

A machine learning paradigm in which an agent learns to make decisions by taking actions in an environment and receiving rewards or penalties as feedback, gradually learning a policy that maximizes cumulative reward — used in robotics control, game-playing AI, and increasingly in tuning industrial control policies.

S

Supervised Learning — Learning paradigm

A machine learning approach where the model is trained on labeled data — input-output pairs where the correct answer is known — to learn a mapping from inputs to outputs. Most classification and regression models used in engineering (fault prediction, cost estimation, load forecasting) are trained via supervised learning.

Supervised Fine-Tuning (SFT) — LLM training step

A stage of LLM training where a pre-trained base model is further trained on curated example prompt-response pairs to teach it to follow instructions and produce helpful responses, typically followed by reinforcement learning from human feedback (RLHF) to further align the model's behavior with human preferences.

T

Test Set — Data split term

A portion of the dataset held out entirely from training and hyperparameter tuning, used only once at the very end to give an unbiased estimate of how the final model will perform on genuinely new data. Using the test set to make modeling decisions defeats its purpose and gives an overly optimistic performance estimate.

Time Series — Data structure

A sequence of data points indexed in time order — sensor readings taken every minute, monthly energy bills, daily equipment vibration measurements. Time series data has special structure (trend, seasonality, autocorrelation) that requires specialized forecasting methods like ARIMA, exponential smoothing, or recurrent neural networks rather than standard i.i.d.-assuming models.

Token — LLM processing unit

The basic unit of text an LLM processes — roughly a word, part of a word, or punctuation mark (not a whole word or single character in most cases). LLM pricing, context window limits, and processing speed are typically measured in tokens rather than words or characters.

Training Set — Data split term

The portion of a dataset actually used to fit a model's parameters — typically 60-80% of available data, with the remainder split between a validation set (used to tune hyperparameters) and a test set (used for final unbiased evaluation).

Transfer Learning — Modeling technique

The technique of taking a model already trained on one large, general dataset and adapting it to a new, related, often smaller task — rather than training a new model from scratch. Transfer learning is why a computer-vision model pre-trained on millions of general images can be adapted with relatively little additional data to detect specific defects on a manufacturing line.

Transformer — Model architecture

A neural network architecture, introduced in 2017, built around a mechanism called "self-attention" that allows the model to weigh the relevance of every other element in an input sequence when processing each element — enabling much better handling of long-range context than earlier recurrent architectures. Transformers are the architecture underlying essentially all modern large language models.

U

Underfitting — Model failure mode

When a model is too simple to capture the true underlying pattern in the data, resulting in poor performance on both training data and new data. Underfitting is the opposite failure mode from overfitting — the fix is usually a more expressive model, better features, or reduced regularization.

Unsupervised Learning — Learning paradigm

A machine learning approach that finds structure or patterns in data without any labeled outputs — clustering similar data points, detecting anomalies, or reducing dimensionality. Unsupervised learning is useful when labeled data (like confirmed failure events) is scarce or expensive to obtain.

V

Validation Set — Data split term

A portion of the dataset set aside during training — separate from both the training set and the test set — used to tune hyperparameters and make model-selection decisions during development, without contaminating the final, unbiased test-set evaluation.

Vector Database — Infrastructure component

A specialized database optimized for storing and efficiently searching high-dimensional embedding vectors by similarity (nearest-neighbor search) rather than exact match. Vector databases are the retrieval backbone of RAG systems, enabling fast lookup of the most semantically relevant document chunks for a given query.

Quick Reference Table — 55 AI & Data Science Terms

Term	Category	Key Point
Accuracy	Metric	Correct ÷ total; misleading on imbalanced data
Activation Function	Neural network	Adds non-linearity (ReLU, sigmoid, softmax)
AI Agent	Architecture	LLM that plans and calls tools in a loop
Anomaly Detection	Application	Flags deviations without full labeled dataset
Artificial Intelligence	Field	Broad field; ML and deep learning are subsets
Backpropagation	Algorithm	Computes gradients layer by layer
Bias-Variance Tradeoff	Theory	Underfitting vs overfitting tension
Big Data	Data characteristic	Volume, velocity, variety
Classification	Task	Predicts discrete category label
Clustering	Task	Groups data without labels
Confusion Matrix	Evaluation	TP/TN/FP/FN breakdown
Cross-Validation	Validation	Repeated train/validate splits (k-fold)
Data Drift	Monitoring	Input distribution shifts over time
Deep Learning	Subfield	Multi-layer neural networks
Dimensionality Reduction	Preprocessing	Fewer features, less noise (PCA)
Embedding	Representation	Numeric vector capturing meaning
Epoch	Training	One full pass over training data
Ensemble Method	Modeling	Combines multiple models (random forest, boosting)
F1 Score	Metric	Harmonic mean of precision and recall
Feature	Data term	One measurable input variable
Feature Engineering	Practice	Creating/transforming inputs to improve models
Fine-Tuning	Adaptation	Further training on domain-specific data
Generative AI	Category	Creates new content (LLMs, image models)
Gradient Descent	Optimization	Iteratively reduces loss via learning rate steps
Hallucination	Failure mode	Confident but fabricated LLM output
Hyperparameter	Configuration	Set before training (learning rate, tree count)
Imbalanced Dataset	Data characteristic	One class vastly outnumbers another
Inference	Deployment	Using trained model on new data
Label	Supervised term	Known correct output for training example
Large Language Model	Architecture	Transformer trained on text (GPT, Claude)
Loss Function	Training	Quantifies prediction error to minimize
Machine Learning	Field	Learns patterns from data, not hard rules
MLOps	Discipline	Deploy/monitor/maintain models in production
Model	Artifact	Trained parameters + algorithm structure
Model Drift	Monitoring	Input-output relationship changes over time
NLP	Subfield	Language understanding and generation
Neural Network	Architecture	Layers of weighted, activated nodes
Overfitting	Failure mode	Memorizes training noise; fails on new data
Outlier	Data characteristic	Point significantly different from the rest
Precision	Metric	TP ÷ (TP + FP); false-alarm rate inverse
Predictive Maintenance	Application	Forecasts failure from sensor data
Prompt Engineering	Practice	Crafting LLM input for reliable output
RAG	Architecture	Retrieves docs before LLM generates answer
Recall	Metric	TP ÷ (TP + FN); miss rate inverse
Regression	Task	Predicts continuous numeric value
Regularization	Technique	Reduces overfitting (L1/L2, dropout)
Reinforcement Learning	Paradigm	Learns via reward/penalty feedback
Supervised Learning	Paradigm	Learns from labeled input-output pairs
Supervised Fine-Tuning	LLM training	Trains on curated instruction examples
Test Set	Data split	Held out for final unbiased evaluation
Time Series	Data structure	Sequential, time-indexed data
Token	LLM unit	Basic text unit LLMs process
Training Set	Data split	Data used to fit model parameters
Transfer Learning	Technique	Adapts pre-trained model to new task
Transformer	Architecture	Self-attention; basis of modern LLMs
Underfitting	Failure mode	Model too simple to capture pattern
Unsupervised Learning	Paradigm	Finds structure without labels
Validation Set	Data split	Tunes hyperparameters during development
Vector Database	Infrastructure	Stores/searches embeddings by similarity