AI Literacy Glossary

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

A

Ablation Study: An experiment where components of a model or pipeline are removed to assess their impact on performance.

Active Learning: A training strategy where a model identifies which data points would be most useful for humans to label next.

Adversarial Example: An input deliberately modified to cause a model to make an incorrect prediction.

Agentic AI: Systems that can plan, make decisions, and take actions across multiple steps, often using tools or other models.

Alignment: The extent to which a model’s behavior matches intended goals, values, or policies.

API (Application Programming Interface): A structured way for software to interact with an AI model or service programmatically.

Artificial Intelligence (AI): A broad field focused on creating systems that perform tasks normally requiring human intelligence, such as reasoning, learning, and decision-making.

Attention: A mechanism that allows models to focus on the most relevant parts of the input when processing information.

Autoencoder: A neural network designed to learn compact representations of data by reconstructing inputs from compressed versions.

B

Batch Job: A compute task submitted to a scheduler and run without interactive control.

Batch Processing: Running inference or training on many inputs simultaneously rather than one at a time.

Benchmark: A standardized dataset or task used to evaluate and compare model performance.

Bias: Systematic errors or skewed outputs arising from data, model design, or usage context.

C

Chain-of-Thought: A prompting approach that encourages models to generate intermediate reasoning steps.

Checkpointing: Saving model state during training so work can be resumed later.

Cloud Computing: On-demand access to remote computing resources over the internet.

Cluster: A group of connected computers that work together as a single computing resource.

Compute: Hardware resources, such as CPUs, GPUs, or TPUs, used to train or run AI models.

Computer Vision (CV): AI methods that allow machines to interpret visual data, including classification, detection, and segmentation.

Confidence Score: A numerical estimate of how certain a model is about a prediction.

Context Window: The maximum amount of information, measured in tokens, that a model can process at once.

CPU (Central Processing Unit): A general-purpose processor optimized for sequential tasks and system-level operations.

CUDA: A parallel computing platform and programming model used to run code on NVIDIA GPUs.

D

Data Leakage: When information outside the training set improperly influences model training or evaluation.

Data Provenance: Information describing where data originated, how it was collected, and how it has been processed.

Deep Learning (DL): A subset of machine learning using multi-layer neural networks to learn complex patterns.

Determinism: The degree to which a model produces the same output for the same input.

Diffusion Models: Generative models that create data by iteratively refining noise into structured outputs.

Distributed Training: Training models across multiple machines or GPUs to scale computation.

Domain Adaptation: Adjusting a model trained in one domain to perform well in another related domain.

Downstream Task: A specific application a pre-trained model is adapted or evaluated for.

E

Edge AI: Running AI models locally on devices rather than on cloud infrastructure.

Embeddings: Numeric vector representations of data that capture semantic or structural relationships.

Emergent Behavior: Capabilities that arise in large models without being explicitly programmed.

Evaluation Metrics: Quantitative measures used to assess model performance, such as accuracy or recall.

Explainability: The degree to which a model’s outputs and behavior can be understood by humans.

F

Federated Learning: A training approach where models learn from decentralized data without moving it to a central location.

Few-Shot Learning: A model’s ability to perform tasks with only a small number of examples.

Fine-Tuning: Adapting a pre-trained model to a specific task or dataset.

Foundation Models: Large, general-purpose models designed to be adapted across many tasks.

G

Generative AI (GenAI): Models that create new content, such as text, images, code, or audio.

Generalization: A model’s ability to perform well on unseen data.

GPU (Graphics Processing Unit): A processor optimized for parallel computation, widely used for training and running deep learning models.

Gradient: A signal used during training to update model parameters in order to reduce error.

Ground Truth: Trusted reference data used for training or evaluation.

Guardrails: Constraints or checks designed to limit unsafe or unintended model behavior.

H

Hallucination: When a model produces incorrect or fabricated information presented as factual.

Human-in-the-Loop: A workflow where human judgment is integrated into model training or decision-making.

Hyperparameters: Configuration settings chosen before training that influence how a model learns.

I

Inference: Using a trained model to generate outputs on new data.

Inference Cost: The computational or financial cost of running a model.

Interpretability: Methods for understanding how model inputs influence outputs.

J

K

Knowledge Distillation: Training a smaller model to mimic a larger model’s behavior.

L

Large Language Models (LLMs): Models trained on large text datasets to understand and generate natural language.

Latency: The delay between providing input to a model and receiving output.

M

Machine Learning (ML): A subset of AI where systems learn patterns from data rather than explicit rules.

Model Cards: Documentation describing a model’s data, intended use, limitations, and ethical considerations.

Model Collapse: Degradation in model quality caused by repeated training on synthetic or low-diversity data.

Model Drift: Performance degradation over time due to changing data or conditions.

Multimodal AI: Models that work across multiple data types, such as text and images.

Multitask Learning: Training a single model to perform multiple tasks simultaneously.

N

Natural Language Processing (NLP): AI methods for analyzing, understanding, and generating human language.

Neural Networks: Computational models composed of connected layers of processing units inspired by the brain.

Node: An individual machine within a computing cluster.

O

On-Device Learning: Updating models directly on local hardware rather than centralized servers.

Open-Source Models: Models whose code and weights are publicly available.

Out-of-Distribution Data: Inputs that differ significantly from the data a model was trained on.

Overfitting: When a model performs well on training data but poorly on new data.

P

Parameters: Internal values learned by a model during training.

Post-Training Evaluation: Assessing model performance after deployment.

Pretraining: Initial training of a model on large, general datasets.

Privacy-Preserving AI: Techniques that protect sensitive data during model training or inference.

Probabilistic Output: Predictions expressed as likelihoods rather than fixed answers.

Prompt Injection: A security risk where inputs manipulate a model’s instructions or behavior.

Prompting: Writing inputs that guide how a model responds.

Q

R

RAM: System memory used by the CPU to store active data and programs.

Reinforcement Learning (RL): A learning approach where agents learn by receiving rewards or penalties.

Regularization: Techniques used to reduce overfitting during training.

Reproducibility: The ability to replicate results using the same data and methods.

Responsible AI: Practices emphasizing fairness, transparency, accountability, and ethics.

Retrieval-Augmented Generation (RAG): A method where models retrieve relevant documents before generating responses.

S

Sampling: Methods that control how generative models select outputs.

Safety Filtering: Processes that detect and block harmful outputs.

Scaling Laws: Empirical relationships linking model size, data, compute, and performance.

Segmentation: A computer vision task that assigns labels to individual pixels or regions.

Supervised Learning: Training models using labeled examples.

Synthetic Data: Artificially generated data used for training or testing models.

System Prompt: Instructions that set a model’s overall behavior before user input.

T

Temperature: A parameter controlling randomness in generative outputs.

Token: A basic unit of text processed by language models.

Tool Use: A model’s ability to call external functions or software.

TPU (Tensor Processing Unit): A specialized accelerator designed for machine learning workloads.

Training Data: The examples used to teach a model.

Transfer Learning: Applying knowledge from one task or dataset to another.

Transformers: A neural network architecture based on attention mechanisms.

U

Unsupervised Learning: Learning patterns from data without labeled outputs.

V

Validation Data: Data used to tune and evaluate models during training.

Virtual Environment: An isolated software environment used to manage dependencies.

Vision Models: Models designed to interpret visual data.

VRAM: Dedicated memory on GPUs used to store model parameters and intermediate computations.

W

X

Y

Z

Zero-Shot Learning: Performing tasks without task-specific training examples.