Glossary

Search results

3D Gaussian Splatting(also: 3DGS, Gaussian Splatting)

An explicit 3D scene representation, introduced by Kerbl et al. in 2023, in which a scene is modelled as a cloud of anisotropic 3D Gaussians whose positions, shapes, colours, and opacities are optimised to match input views. Unlike NeRFs, which require costly neural-network…

AI contestability(also: Algorithmic contestability, AI contestation)

The principle that users should be able to challenge, question, and seek recourse against decisions or outputs made by AI systems. In accessibility contexts, contestability is particularly important for blind users who rely on AI for visual access — they need mechanisms to flag…

Acoustic Activity Recognition(also: Sound Activity Recognition, Audio Activity Recognition, Environmental Sound Recognition)

The use of microphones and machine learning to automatically identify and classify sounds occurring in an environment, such as doorbells, alarms, appliances, speech, and other everyday acoustic events. Acoustic activity recognition is particularly relevant to accessibility for…

Acoustic Event Detection(also: Sound Event Detection, Audio Event Detection, Sound Event Classification)

The automated process of identifying and classifying specific sounds within an audio stream, such as recognizing a phone ringing, door knocking, fire alarm, or speech from continuous environmental audio. Acoustic event detection systems use machine learning trained on labeled…

Acoustic Model(also: AM)

An acoustic model is the component of an automatic speech recognition (ASR) system that maps short segments of audio (typically 10–25 ms frames of spectral features) to the linguistic units that produced them — most commonly phonemes or sub-phonetic states. Classical acoustic…

Adaptive Boosting(also: AdaBoost)

A machine learning ensemble method that combines multiple weak classifiers to create a strong classifier, with each successive classifier focusing on the examples that previous classifiers misclassified. In computer vision and accessibility applications, AdaBoost is widely used…

App Review Mining(also: App Store Review Analysis, User Review Mining)

The process of systematically extracting, classifying, and analyzing user reviews from app stores such as Google Play and the Apple App Store to identify patterns, issues, and feature requests. In accessibility research, app review mining is used to discover real-world…

Attention Mechanism(also: Attention)

A technique in neural networks that allows models to focus on relevant parts of the input when generating each part of the output, rather than relying solely on a fixed-length context vector. In sequence-to-sequence models, attention computes a weighted combination of all…

Automated speech scoring(also: Computer-aided speech assessment, Automatic speech evaluation)

The use of computational techniques — including voice activity detection, phoneme recognition, prosody analysis, and speaker diarization — to automatically evaluate the accuracy and quality of speech production without requiring real-time human assessment. In speech therapy…

Automatic Sign Language Processing(also: ASLP, Sign Language Processing)

An umbrella term encompassing three major areas of technology for sign language: automatic sign language generation (ASLG, creating sign language output from text or speech), automatic sign language recognition (ASLR, interpreting sign language input), and automatic sign…

Automatic text simplification(also: ATS, Text simplification, Automated readability improvement)

The use of natural language processing techniques to transform complex text into simpler, more readable versions while preserving meaning. ATS operates at two levels: lexical simplification (replacing difficult words with simpler synonyms) and syntactic simplification…

BLEU Score(also: BiLingual Evaluation Understudy, BLEU)

A metric for evaluating the quality of machine-generated text by comparing it to one or more reference (human-written) translations. BLEU calculates precision by counting how many n-grams (sequences of words) in the predicted text match n-grams in the reference text, with BLEU-1…

Background Subtraction(also: Foreground-Background Separation, Background Modelling)

Background subtraction is a computer vision technique used to identify moving objects (the foreground) in a video by comparing each frame against a model of the static background. Common approaches include adaptive Gaussian mixture models that continuously update the background…

Bayesian Network(also: Bayes Network, Belief Network, Probabilistic Graphical Model)

A statistical model that represents probabilistic relationships among variables using a directed graph structure. In accessibility and assistive technology applications, Bayesian networks are used for behavior recognition—inferring what action a user is performing based on…

Binary Classification(also: Two-Class Classification)

A type of supervised machine learning task where the goal is to categorize items into one of exactly two classes. In accessibility research, binary classification has been applied to automatically determine whether a bug report is accessibility-related or not, whether user…

Black Box Model(also: Opaque Model)

A machine-learning model whose internal workings are not directly inspectable or interpretable by a human, either because the model is architecturally complex (deep neural networks, large language models) or because it is proprietary and the developer does not disclose its…

Calibration-Free Interface(also: Zero-Shot Interface, Plug-and-Play Interface, Cross-User Model)

An input system that works for a new user without any per-user training or calibration data, typically by relying on models trained on large multi-user datasets that capture enough physiological and behavioural variation to generalise. Voice assistants and mixed-reality hand…

Camera-based assistive technology(also: Camera-based AT, Vision-based AT, VBAT)

Assistive technologies that use cameras (typically smartphone cameras or smart glasses) combined with computer vision and AI to provide visual information to blind and low-vision users. Applications include object recognition, text reading (OCR), scene description, face…

Cascading classifier(also: Cascaded detection, Multi-stage classifier)

A machine learning architecture that chains multiple detection stages in sequence, where each stage filters candidates before passing them to the next, progressively increasing detection precision while maintaining recall. In accessibility applications, cascading classifiers are…

Chain-of-Thought(also: CoT, Chain of Thought Reasoning, Step-by-Step Reasoning)

Chain-of-thought is a prompting and model-design technique in which a large language model produces its intermediate reasoning steps before giving a final answer. Modern reasoning models (e.g., OpenAI o-series, Claude thinking modes) expose chain-of-thought as visible internal…

Clustering Algorithm(also: Cluster Analysis, Unsupervised Clustering, K-means)

A clustering algorithm is an unsupervised machine-learning technique that groups similar data points together based on a distance or similarity measure, without needing pre-labelled training data. Common algorithms include K-means, PAM (Partitioning Around Medoids), CLARA…

Confidence score(also: Certainty score, Prediction confidence)

A numerical value (typically 0-100% or 0-1) indicating how certain an AI system is about its prediction or classification. In accessibility contexts, communicating confidence scores to users — particularly blind users who cannot visually verify AI output — helps them calibrate…

Constitutional AI(also: CAI)

A training method introduced by Anthropic in 2022 in which a large language model is aligned to a written set of principles (a 'constitution') through self-critique and reinforcement learning from AI feedback, rather than relying exclusively on human preference labels. The model…

Continual Learning(also: Continuous Learning, Lifelong Learning, Never-ending Learning)

A machine learning paradigm in which models learn incrementally from new data over time while retaining previously acquired knowledge, rather than being trained once on a fixed dataset. Continual learning is relevant to accessibility because it enables AI-powered accessibility…

Continuous Sign Language Recognition(also: CSLR)

A computer vision task that involves recognizing sign language from continuous, naturally produced signing — as opposed to isolated sign recognition, which identifies individual signs in segmented clips. Continuous sign language recognition deals with the complexities of natural…

Contrastive Decoding(also: Visual Contrastive Decoding, VCD)

Contrastive decoding is a technique for reducing hallucinations in large language model and multimodal AI outputs by comparing token probability distributions across different input conditions. The core principle is that tokens genuinely grounded in the input content will change…

Contrastive Learning(also: Contrastive Self-Supervised Learning)

Contrastive learning is a machine learning technique that trains models to produce vector embeddings by maximising similarity between representations of the same or augmented instance (positive pairs) while minimising similarity between representations of different instances…

Convolutional Neural Network(also: CNN, ConvNet)

A class of deep neural network that uses convolutional filters to automatically extract spatial features from data, originally designed for image processing but now widely applied to sensor data, audio, and video analysis. CNNs identify patterns like edges, textures, and shapes…

Counterfactual Explanation(also: Counterfactual XAI)

An explanation technique that communicates what minimal change to the input would have produced a different output from an AI model, for example 'if the applicant's income had been $5,000 higher, the loan would have been approved'. Counterfactual explanations are legally…

Cross-Validation(also: K-Fold Cross-Validation, Stratified Cross-Validation)

A statistical method for evaluating machine learning models by splitting data into multiple subsets (folds), training the model on some folds and testing on the remaining ones, then rotating through all combinations. Stratified cross-validation ensures each fold maintains the…

DBSCAN(also: Density-Based Spatial Clustering of Applications with Noise)

A density-based clustering algorithm introduced by Ester, Kriegel, Sander, and Xu (1996) that groups data points located in dense neighbourhoods and labels sparse points as noise. Unlike k-means, DBSCAN does not require the user to specify the number of clusters in advance and…

Data Descriptor(also: Training Data Descriptor)

An automated metric or feedback mechanism that characterizes the quality or properties of a dataset, particularly training images used in machine learning. In accessibility research, data descriptors provide non-visual feedback to blind users about the quality of photos they…

Data Mining(also: Knowledge Discovery, KDD, Knowledge Discovery in Databases)

Data mining is the computational process of discovering patterns, rules, and relationships in large datasets, drawing on techniques from statistics, machine learning, and database systems. Common tasks include classification, clustering, association-rule mining, anomaly…

Data Representativeness(also: Dataset Representativeness, Demographic Representativeness)

The degree to which a dataset reflects the diversity of the population it is intended to serve, particularly across demographic dimensions such as age, gender, race, ethnicity, disability, and socioeconomic status. In AI and machine learning, unrepresentative training data leads…

Dataset Bias(also: Training Data Bias, Data Representation Bias, Sampling Bias)

A systematic skew in the composition of training data used to build machine learning models, resulting in models that perform well for overrepresented groups but poorly for underrepresented ones. In accessibility contexts, dataset bias is a pervasive problem: activity…

Debiasing(also: Bias mitigation, Bias correction)

Debiasing refers to techniques and processes applied to AI systems—particularly machine learning models and large language models—to detect, reduce, or eliminate unfair biases that cause the system to produce outputs that discriminate against or misrepresent specific demographic…

Decision Tree(also: Classification Tree, Regression Tree, C4.5)

A decision tree is a supervised machine-learning model that represents a classification or regression decision as a tree of yes/no tests on input features, with predictions at the leaves. Well-known algorithms include ID3, C4.5, CART, and Random Forests. Decision trees are…

Diffusion Model(also: Diffusion-based Generator, Denoising Diffusion Model)

A diffusion model is a class of generative AI that learns to produce images or videos by iteratively denoising a random noise input, reversing a forward process that gradually adds noise to training data. In accessibility work, diffusion models are used to synthesize sign…

Document Expansion(also: Query Prediction, Document Enrichment)

An information retrieval technique that enhances a document by augmenting it with additional terms or predicted queries that users might use to search for that content. Methods like DocTTTTTQuery use sequence-to-sequence machine learning models to generate likely search queries…

Document Layout Analysis(also: DLA, page layout analysis)

A computer-vision task that identifies and classifies the visual regions of a document page—headings, paragraphs, tables, figures, captions, lists, headers, and footers—typically using object-detection models trained on datasets such as DocLayNet, PubLayNet, or DocBank. Document…

Domain Adaptation(also: Cross-Domain Transfer, UDA, Unsupervised Domain Adaptation)

A machine learning technique that enables models trained on data from one domain (such as web interfaces) to perform well on a different but related domain (such as mobile app interfaces). Domain adaptation is valuable for accessibility because it allows models trained on…

Dynamic Bayesian Network(also: DBN, Temporal Bayesian Network)

A probabilistic graphical model that represents sequences of variables over time, extending standard Bayesian networks to handle temporal relationships. In accessibility and affective computing contexts, Dynamic Bayesian Networks are used to model how facial expressions, head…

Dynamic Programming(also: DP, DP Matching)

A mathematical optimization technique used in pattern recognition that breaks complex problems into simpler overlapping subproblems. In accessibility technology, dynamic programming matching (DP matching) is commonly used in sign language recognition and speech recognition…

Dynamic Time Warping(also: DTW)

An algorithm for measuring similarity between two temporal sequences that may vary in speed or timing. Dynamic time warping aligns sequences by warping the time axis to find the optimal match, making it robust to variations in how quickly gestures are performed. DTW is commonly…

Eigenfaces

A computer vision technique for face recognition that uses Principal Component Analysis to represent faces as a linear combination of standardized face components (eigenvectors derived from a training set of face images). Developed by Turk and Pentland in 1991, Eigenfaces was…

Element Detection(also: UI Element Detection, Widget Detection, Object Detection)

The task of automatically identifying the locations and types of user interface components (such as buttons, text fields, images, and checkboxes) from a screenshot using computer vision models. Element detection is important for accessibility because it can identify interactive…

Explainable AI(also: XAI, Interpretable AI)

A set of methods and design approaches that make the outputs and decision-making processes of artificial intelligence systems understandable to human users. Explainable AI aims to provide transparency about why an AI produced a particular result, typically through confidence…

Feature Extraction(also: Feature Engineering, Representation Learning)

Feature extraction is the process of identifying and isolating measurable properties or characteristics (features) from raw data such as images, audio, or text, for use in machine learning tasks. In image processing, features may include edges, textures, colours, shapes, or…

Feature Hashing(also: Hashing Trick)

A technique used in machine learning to convert text or categorical data into fixed-length numerical feature vectors by applying a hash function. Feature hashing is particularly useful for handling high-dimensional sparse data, such as the text of bug reports or user reviews. It…

Federated Learning(also: FL)

A machine-learning approach in which a shared model is trained across many user devices without the raw training data ever leaving those devices: each device computes updates locally and sends only model parameters or gradients to a central server for aggregation. Federated…