Glossary

Search results

LLM-as-Judge(also: LLM as a Judge, Model-as-Judge)

An evaluation methodology in which a large language model is prompted to assess the quality of some artifact — generated text, code, a UI, or a response from another model — according to a structured rubric. LLM-as-judge is attractive because it scales automated evaluation to…

LSTM(also: Long Short-Term Memory, LSTM Network)

A type of recurrent neural network architecture designed to learn long-term dependencies in sequential data by using special gating mechanisms that control the flow of information through the network. LSTMs are particularly effective for processing time-series data such as…

Language Model(also: Statistical Language Model, LM)

A computational model that assigns probabilities to sequences of words, enabling prediction of likely next words or sentences in text. In assistive technology, language models power word and sentence prediction systems by learning patterns from training corpora. Modern AAC…

Language Understanding Intelligent Service(also: LUIS, Azure LUIS)

A cloud-based Microsoft Azure service that applies machine learning to natural language text to predict meaning and extract relevant information. LUIS identifies user intents (what they want to do) and entities (key information in their utterance). In accessibility applications,…

Large Language Model(also: LLM)

A type of artificial intelligence model trained on vast amounts of text data to understand and generate human language. Large language models like GPT-4, Claude, and Gemini power many generative AI applications. In accessibility contexts, LLMs enable natural language interfaces…

Large Vision Model(also: LVM)

A large vision model is a foundation model trained on very large image (and often video) datasets to produce general-purpose visual representations - capable of object detection, segmentation, captioning, or feature extraction without task-specific retraining. Examples include…

LoRA(also: Low-Rank Adaptation)

A parameter-efficient fine-tuning technique, introduced by Hu et al. in 2022, in which a large pretrained neural network is specialised by training only a pair of small low-rank matrices that modify specific weight projections, while the original weights remain frozen. LoRA…

Machine Learning(also: ML)

A branch of artificial intelligence in which computer systems learn patterns from data to make predictions or decisions without being explicitly programmed for each scenario. In accessibility contexts, machine learning is used for a wide range of applications: predicting…

Machine Teaching(also: Interactive Machine Teaching)

A paradigm in human-computer interaction where non-expert users guide the training of machine learning models through interactive feedback, such as providing examples, labels, or corrections. Unlike traditional machine learning where data scientists prepare datasets and tune…

Markov Decision Process(also: MDP)

A mathematical framework for modelling decision-making in situations where outcomes are partly random and partly under the control of a decision-maker. In accessibility and assistive technology, Markov decision processes and their extension, partially observable Markov decision…

Markov Logic Networks(also: MLN, MLNs)

A machine learning framework that combines first-order logic with probabilistic graphical models to handle uncertainty in rule-based reasoning. In assistive technology contexts, MLNs enable context-aware systems to make intelligent decisions by weighing multiple factors—such as…

MediaPipe

An open-source framework by Google for building multimodal machine learning pipelines, commonly used for real-time face, hand, and body tracking. In accessibility applications, MediaPipe Holistic extracts 3D landmarks from the user's body and hands via webcam, while MediaPipe…

Meta-learning(also: Learning to Learn)

A branch of machine learning where models are trained to learn new tasks from very few examples by leveraging knowledge gained from previous tasks. In accessibility applications, meta-learning enables technologies like teachable object recognizers that can quickly adapt to…

Minimum Viable Description(also: MVD)

Minimum viable description (MVD) is an emerging framework for audio description that establishes the foundational level of visual information needed to provide equal access to video content without introducing bias or cognitive overload. Rather than attempting to describe…

Misgendering

The act of referring to someone using language that does not reflect their gender identity, such as incorrect pronouns, titles, or gendered terms. In digital accessibility and AI contexts, misgendering occurs when automated systems incorrectly classify a person's gender based on…

Misinterpretation(also: AI Misinterpretation)

An AI error where the model incorrectly identifies or describes something that is actually present in the input. In image descriptions, misinterpretation includes errors like mistaking one product for another (shampoo for cleaning product), reading numbers incorrectly ("6mg"…

Mixed-Initiative Interaction(also: Mixed-Initiative Systems, Human-Agent Collaboration)

An interaction paradigm in which both the human user and the computer system can take initiative in directing the task, rather than one party being entirely in control. In accessibility contexts, mixed-initiative interaction is particularly important for AI-powered assistive…

Mixture of Experts(also: MoE)

Mixture of experts is a neural network architecture that routes each input through a small subset of specialist subnetworks ('experts') rather than activating the whole model. A gating network decides which experts handle a given token or query, letting the overall model be much…

Model Reliability(also: AI Reliability, Model Trustworthiness)

The degree to which an AI model consistently produces accurate, truthful, and complete outputs across different inputs and contexts. In the context of visual access technology for BLV users, model reliability encompasses factual accuracy (not fabricating content), interpretive…

Multi-Model Comparison(also: Cross-Model Comparison, Ensemble Verification)

The practice of generating responses from multiple AI models for the same input and comparing their outputs to assess reliability, identify errors, and provide a more comprehensive understanding of the content. In accessibility contexts, multi-model comparison is used to help…

Multimodal AI(also: Multimodal Generative AI)

Artificial intelligence systems capable of processing and generating content across multiple modalities such as text, images, audio, and video. In accessibility contexts, multimodal AI is significant because a single tool can address diverse access needs — for example,…

Multimodal Large Language Model(also: MLLM, Vision-Language Model, VLM)

A deep learning model that can process and generate content across multiple types of input including text, images, audio, and video. In accessibility contexts, MLLMs like GPT-4o, Gemini, and Claude have become transformative tools for blind and low vision users, enabling…

Natural Language Command(also: Natural Language Input, Conversational Command)

A user input expressed in everyday spoken or written language rather than structured syntax or specific command formats. In accessibility contexts, natural language commands enable BLV users to interact with systems without memorizing precise command structures or navigating…

Natural Language Generation(also: NLG, Text Generation)

A subfield of artificial intelligence and computational linguistics focused on automatically producing human-readable text from structured data or other non-linguistic representations. In accessibility, natural language generation is used to create textual descriptions of visual…

Natural Language Processing(also: NLP, Computational Linguistics)

A branch of artificial intelligence that enables computers to understand, interpret, and generate human language. In accessibility, NLP powers voice-based assistive technologies, automatic captioning, text simplification for cognitive accessibility, and natural language query…

Natural Language Query(also: NLQ, Conversational Query, Natural Language Search)

An interaction paradigm where users pose questions or issue commands in everyday language rather than using structured search syntax, predefined filters, or navigation controls. In accessibility contexts, natural language querying is particularly valuable for screen reader users…

Natural Language Understanding(also: NLU, Intent Recognition, Language Understanding)

A branch of artificial intelligence that enables computers to interpret the meaning and intent behind human language input, rather than requiring exact predetermined phrases or commands. In accessibility contexts, NLU is valuable for voice-controlled interfaces because it allows…

Neural Network(also: Artificial Neural Network, ANN)

A machine learning model inspired by the structure of biological neural networks in the brain, consisting of interconnected layers of nodes (neurons) that process information by adjusting weighted connections during training. In accessibility and assistive technology, neural…

Neural Radiance Field(also: NeRF)

An implicit neural representation of a 3D scene, introduced by Mildenhall et al. in 2020, in which a small neural network is trained to map any 3D coordinate and viewing direction to a colour and density value. Rendering is performed by volumetric ray marching through this…

OCR(also: Optical Character Recognition)

Technology that converts images of text into machine-readable text. In accessibility contexts, OCR is used by visual assistance technologies and screen readers to read printed text from photos, signs, documents, and product packaging. While valuable for blind users, OCR has…

Obfuscation(also: Content Obfuscation, Visual Obfuscation)

AI-enabled techniques that automatically detect and remove or conceal private content in images and videos by applying filters such as blurring, masking, or blocking. In the context of visual assistance technologies, obfuscation aims to protect blind users from inadvertently…

Object Detection(also: Object Recognition)

A computer vision technique that identifies and locates specific objects within images or video frames, typically by drawing bounding boxes around detected items and classifying them. In video accessibility, object detection enables automatic identification of video elements…

Object Recognition(also: Object Detection)

A computer vision capability that identifies and classifies objects within images or video frames. In visual assistance technologies, object recognition enables automated description of what the camera captures, helping blind users identify items in their environment. However,…

Object-Based Cropping(also: Semantic Cropping, Object-Aware Cropping)

An image cropping approach that allows users to select which objects to keep in an image rather than specifying pixel coordinates or spatial boundaries. Object-based cropping uses computer vision to identify and segment objects (e.g., "the dog", "the chair"), then crops the…

Omission(also: AI Omission, Information Omission)

An AI error where the model fails to mention important information that is present in the input. In image descriptions for BLV users, omission can include failing to mention warning labels on medication, not describing important text in a document, skipping relevant objects in a…

Optical Character Recognition(also: OCR, Text Recognition)

Technology that converts images of text—whether typed, handwritten, or printed—into machine-readable text data. OCR is used in accessibility to extract text from images, documents, video frames, and real-world scenes, enabling screen readers to read text that would otherwise be…

Outlier detection(also: Anomaly detection, Novelty detection)

An algorithmic technique that identifies data points or behaviors that deviate significantly from expected patterns, used in fraud detection, quality assurance, CAPTCHAs, and crowd labor platforms. People with disabilities are disproportionately flagged as outliers because…

Parameter-Efficient Fine-Tuning(also: PEFT, Lightweight Fine-Tuning)

Parameter-efficient fine-tuning is a family of techniques (LoRA, adapters, prefix tuning, prompt tuning) that adapt a large pretrained model to a new task or domain by updating only a small fraction of its parameters - typically under 1% - while freezing the rest. This…

Participatory AI(also: Community-Centered AI, Participatory Machine Learning)

An approach to artificial intelligence development that actively involves the communities affected by AI systems in defining problems, setting priorities, designing solutions, collecting data, evaluating outcomes, and governing deployment. Participatory AI goes beyond…

Pattern Recognition

A branch of machine learning and artificial intelligence focused on identifying regularities, patterns, and structures in data such as images, sounds, or sensor readings. In accessibility, pattern recognition is fundamental to technologies like sign language recognition systems…

Pedestrian Detection(also: Person Detection, Human Detection)

A computer vision task that identifies and locates people in images or video frames, typically using deep learning models such as convolutional neural networks. In accessibility applications, pedestrian detection is used in wearable assistive technologies for blind and low…

Perceptual Gap

A design failure identified by Choudhury (2026) in which an AI system's explanation is delivered through exactly the sensory channel that its user cannot access. For example, a Grad-CAM heat map overlaid on an image tells a blind user where the model looked but cannot be seen by…

Perturbation testing(also: Counterfactual testing, Template-based testing)

A bias evaluation methodology for NLP models that systematically substitutes identity-related terms (e.g., disability phrases) in otherwise identical sentences to measure whether the model produces different predictions based on the identity mention alone. By holding all other…

Predictive AI(also: Predictive Analytics, Predictive Algorithm)

AI systems that use machine learning to identify patterns in data and anticipate future outcomes, behaviors, or events. In the context of disability, predictive AI systems have been documented causing significant harm by making decisions about resource allocation (welfare…

Proactive description(also: Proactive notification, Unsolicited description)

The ability of an assistive system to automatically provide relevant visual or environmental information without requiring the user to explicitly request it. For blind and visually impaired users navigating real-world environments, proactive description is critical — a human…

Prompt Chaining(also: Chained Prompting, Sequential Prompting)

A technique for interacting with large language models where multiple prompts are issued in sequence, with each prompt building on the output of the previous one to achieve a more refined or accurate result. In accessibility and bias mitigation contexts, prompt chaining enables…

Prompt Engineering(also: Prompt Design, Prompt Crafting)

The practice of designing and structuring input prompts to guide large language models (LLMs) toward producing more accurate, relevant, and useful outputs. In accessibility contexts, prompt engineering techniques such as role-play prompting (assigning expert personas),…

Provenance Indicator(also: Source Attribution)

Information that identifies which AI model, trial, or prompt produced a particular piece of content in a multi-model comparison. Provenance indicators help users understand which models generate which claims, enabling them to build mental models of individual model strengths and…

Recurrent Neural Network(also: RNN)

A recurrent neural network (RNN) is a type of artificial neural network designed to process sequential data by maintaining an internal state (memory) that captures information from previous inputs in the sequence. Unlike feedforward networks, RNNs have connections that loop…

Reinforcement Learning(also: RL)

A type of machine learning where a system learns to make decisions by performing actions in an environment and receiving rewards or penalties based on the outcomes. Unlike supervised learning, which learns from labelled examples, reinforcement learning discovers optimal…