Glossary

Glossary

Plain definitions of the terms used in AI and in Frugal AI work, each with an example. Terms marked Frugal AI framework come from the Hub’s work and the book Frugal AI.

Frugal AI framework

Absorption

An institution’s capacity to sustain an AI system after adopting it: to fund it over its whole life, fit it into real work, govern it and keep it running.

Example: A hospital has absorbed a documentation tool when it is costed, governed and in routine use across its wards.

Frugal AI framework

Absorption Gap

The distance between what an AI system can do and what an institution can responsibly sustain.

Example: A pilot that impresses in demonstration but stalls in production shows an Absorption Gap.

Frugal AI framework

Adoption

The decision to start using an AI system. It marks the start of absorption, which is the harder part.

Example: Signing a contract and launching a pilot are acts of adoption.

Algorithm

A set of rules or steps used for solving a problem or performing a task, especially by a computer.

Example: A recommendation system uses a collaborative filtering algorithm to suggest movies.

API (Application Programming Interface)

A set of tools and definitions that allow software programs to communicate with each other.

Example: OpenAI provides an API that developers use to integrate ChatGPT into apps.

Artificial Intelligence (AI)

The simulation of human intelligence in machines that are programmed to think and learn.

Example: AI powers autonomous vehicles by processing sensor data and making decisions.

ASR (Automatic Speech Recognition)

A technology that converts spoken language into text using specialised audio processing models.

Example: Google Assistant uses ASR to convert your spoken query into text.

Bias (in AI)

Systematic errors in AI systems that result from flawed data, design, or assumptions, often leading to unfair outcomes.

Example: A facial recognition model performs poorly on darker-skinned individuals due to biased training data.

Big Data

Extremely large data sets that may be analysed computationally to reveal patterns, trends, and associations.

Example: Retailers analyse big data from loyalty cards to personalise offers.

Chatbot

A computer program designed to simulate conversation with human users.

Example: A bank’s chatbot answers FAQs about account services.

Cloud Computing

Delivery of computing services (like servers, storage, databases) over the internet.

Example: AWS cloud services host large-scale AI training models.

Compute Power

The processing capacity of a computer, often required in large quantities for AI tasks.

Example: Training GPT-4 requires thousands of GPUs running in parallel.

Dataset

A structured collection of data, often used to train or test machine learning models.

Example: ImageNet is a large dataset used to train computer vision models.

Deep Learning

A subset of machine learning involving neural networks with many layers.

Example: Deep learning models detect tumors from medical images.

Frugal AI framework

Deployment Lab

A Frugal AI Hub partnership with a public institution on a live AI challenge, measured from launch, to learn what deployment and absorption require.

Example: A Lab with a state department testing a voice advisory service under real connectivity and budget limits.

Diffusion Models

A type of generative model used in AI to produce images, audio, and other content.

Example: DALL·E uses diffusion models to generate images from text prompts.

Frugal AI framework

Edge AI

Running AI models on local devices such as phones, sensors or small computers, close to where data is created.

Example: A speech model running offline on a low-cost device in a village with patchy connectivity.

Efficient AI

AI systems designed to minimise resource usage (compute, memory, energy) while delivering effective results.

Example: An optimised image classifier that runs on low-power edge devices.

Embedding

A numerical representation of data (e.g., words, sentences) that captures semantic meaning.

Example: Word2Vec converts words into embeddings used for similarity detection.

Energy-Efficient AI

AI systems optimised to consume less energy, helping reduce environmental impact.

Example: A vision system in a smart thermostat using lightweight AI to reduce home energy use.

Ethical AI

The practice of designing AI technologies in ways that prioritize fairness, accountability, and transparency.

Example: An AI tool is audited to ensure it doesn’t discriminate against job applicants.

Frugal AI framework

Federated Learning

Training a shared model across many devices or sites while the raw data stays where it was collected.

Example: Frugal Voice trains on community devices and shares only model updates.

Fine-tuning

Adjusting a pre-trained AI model on a new, often smaller, dataset to perform specific tasks better.

Example: GPT-3 is fine-tuned on legal documents to draft contracts.

Foundation Model

A large AI model trained on broad data that can be adapted to many different tasks.

Example: GPT-4 is a foundation model that can be adapted for many tasks.

Frugal AI framework

Frugal AI

An outcomes-first design discipline for AI: start from the result an institution needs, work within its real constraints, and use what is sufficient to deliver it, across compute, models, data, orchestration and governance.

Example: Using a small, locally hosted language model on a Raspberry Pi for offline healthcare in rural clinics.

Frugal AI framework

Frugal Stack

The five layers an institution needs to align to deploy AI and keep it working: compute, model, data, orchestration and governance.

Example: A cheap, accurate model still fails if its data pipeline cannot be maintained.

Generative AI

A type of AI that can create new content (text, images, music, etc.) based on training data.

Example: Midjourney generates artwork based on short descriptions.

GPU (Graphics Processing Unit)

Hardware used to accelerate AI computations, especially for training models.

Example: NVIDIA’s GPUs are commonly used to train deep learning models.

Hallucination (in AI)

When an AI model generates plausible but incorrect or fabricated information.

Example: ChatGPT might say ‘Albert Einstein was born in 1950,’ which is a hallucination.

Hyperparameters

Configurable variables used to control the learning process in machine learning models.

Example: Changing the learning rate affects model training speed and accuracy.

Inference

The process of using a trained model to make predictions or generate outputs.

Example: An AI model classifies a photo of a dog during inference.

Input Prompt

The initial text or data given to a generative model to guide its output.

Example: Typing ‘write a poem about spring’ is an input prompt for ChatGPT.

Frugal AI framework

Knowledge Distillation

Training a smaller ‘student’ model to reproduce the behaviour of a larger ‘teacher’ model.

Example: DistilBERT keeps most of BERT’s accuracy at a fraction of its size.

Large Language Model (LLM)

A type of AI model trained on vast amounts of text to understand and generate human-like language.

Example: Claude and LLaMA are examples of large language models.

Latency

The delay between input and response time in computing systems.

Example: Low latency is important in AI-powered real-time video conferencing tools.

Frugal AI framework

Low-Resource Language

A language with little digital text, speech data or tooling available for building AI.

Example: Soliga, spoken in the Biligiri Rangaswamy Hills of Karnataka, has no widely used script.

Machine Learning (ML)

A branch of AI focused on systems that learn from data to improve their performance over time.

Example: Spotify uses ML to generate Discover Weekly playlists.

Model Training

The process of teaching an AI model to make predictions or decisions based on data.

Example: A model is trained on labeled photos to recognise traffic signs.

Natural Language Processing (NLP)

A field of AI that enables computers to understand and interpret human language.

Example: NLP helps AI understand customer service emails.

Neural Network

A computing system inspired by the human brain’s structure, used in deep learning.

Example: Neural networks power handwriting recognition apps.

Open Source

Software whose source code is freely available and can be modified or shared.

Example: Hugging Face hosts open-source models for developers.

Optimisation

The process of adjusting a model or algorithm to improve its performance.

Example: Optimising a model can reduce energy usage and improve speed.

Parameter

A value within a model that the learning algorithm adjusts to fit the training data.

Example: Transformer models have billions of trainable parameters.

Pretraining

The initial phase of training a model on a large dataset before fine-tuning it for specific tasks.

Example: BERT was pretrained on Wikipedia before fine-tuning.

Frugal AI framework

Quantisation

Storing a model’s numbers at lower precision, for example 8-bit instead of 32-bit, to cut memory, cost and energy use.

Example: An 8-bit model that runs on a phone instead of a server.

Reinforcement Learning

A machine learning approach where an agent learns by receiving rewards or penalties.

Example: Robots learn to walk using reinforcement learning.

Responsible AI

The practice of developing and deploying AI in a manner that is ethical and beneficial to society.

Example: A company publishes its AI ethics policy to ensure responsible AI.

Frugal AI framework

Sovereign Spine

At each layer of the Frugal Stack, the decision about what must be controlled, what can be influenced and what may safely be rented.

Example: A ministry may rent compute while keeping control of its data and governance.

Frugal AI framework

Sufficiency Test

The prior question for any AI project: does this workflow need AI at all, and if so, what level of capability is enough for the outcome?

Example: A rules-based form may serve a fixed eligibility check better than a language model.

Sustainable AI

AI technologies designed to have a minimal environmental footprint across their lifecycle.

Example: Using solar-powered edge devices to run crop monitoring AI models in the field.

Text-to-Speech (TTS)

Technology that converts written text into spoken voice.

Example: TTS allows visually impaired users to listen to webpages.

Token

A unit of text (like a word or part of a word) used by language models during processing.

Example: The sentence ‘I love AI’ has 3 tokens.

Frugal AI framework

Total Cost Architecture

How design choices create, move and amplify cost across the Frugal Stack, including the layers procurement leaves uncosted.

Example: A low licence fee can hide large costs in data preparation and human oversight.

Frugal AI framework

Total Cost of Ownership (TCO)

The full cost of an AI system over its life: compute, data, software, people, integration, maintenance, governance and energy.

Example: The Hub and UNICC framework counts seven cost pillars for each use case.

Transformer

A deep learning model architecture that underlies most modern language models.

Example: Transformers are used in ChatGPT for language understanding.

Vector Database

A type of database optimised for storing and retrieving vector representations (e.g., embeddings).

Example: Pinecone stores embeddings for fast semantic search.

Voice Synthesis

The creation of artificial speech using AI techniques.

Example: ElevenLabs creates realistic synthetic voices.

Workflow Automation

The use of technology to perform recurring tasks or processes with minimal human intervention.

Example: Zapier automates tasks like posting Slack messages.

Zero-shot Learning

The ability of an AI model to correctly make predictions on tasks it was not explicitly trained on.

Example: GPT-4 answers questions on topics it hasn’t seen before.

A living resource

Suggest a term

To propose a new term, a correction or an improvement, especially on sustainable, frugal or responsible AI, use the contact form or post in the Frugal AI Initiative group on LinkedIn.