The accepted ontology
Browse accepted Systema Constructum entries: human-made tools, institutions, methods, practices, works, and their definitions.
100 entries in this page
Working-memory-training is a human-constructed instructional method designed to systematically improve the capacity and endurance of working memory through repeated cognitive exercises. Its parameters are: (1) a defined protocol of repeated tasks that tax wor…
Deep learning is a subclass of machine learning defined by neural networks with three or more layers, where each layer learns a progressively higher-level abstraction of the data. Its parameters are the weights and biases of the neural network, persisted thro…
Memory consolidation is a human-made conceptual model that maps the natural process by which initially labile memories are stabilized into long-lasting storage. Parameters: (1) synaptic consolidation — cellular mechanisms (protein synthesis, AMPA receptor ins…
A predictive model is a human-constructed artifact that maps input features to predicted outcomes, calibrated through training on labeled or unlabeled historical data. Its parameters are optimized to minimize prediction error on held-out data, and it persists…
Unsupervised learning is a machine learning paradigm in which algorithms discover hidden structure or patterns in unlabeled data without external feedback. The learner identifies clusters, reduces dimensionality, or models distributions from raw input, rather…
Automaticity is the cognitive process by which repeated, deliberate practice transforms effortful actions into fast, unconscious operations that execute without conscious attention or working-memory load. Its parameters are: (1) acquisition through repetition…
Working-memory-capacity is the human-constructed measure of how much information the working-memory system can actively hold and manipulate at once. Its parameters are: (1) the storage limit (typically 4-7 chunks for humans), (2) the manipulation requirement…
Physics is the human-made discipline of systematic study of natural phenomena — matter, energy, space, time, and their interactions. It operates through a cycle of observation, mathematical modeling, experimental testing, and theoretical refinement. Its persi…
A policy gradient method is a reinforcement learning technique that parameterizes a policy πθ(a|s) with differentiable functions (typically neural networks) and updates parameters θ by ascending the gradient of expected return J(θ) = E[Σr_t] with respect to θ…
A reward model is a supervised learning model that maps text inputs to scalar scores representing human preference, trained on paired comparisons of model outputs where annotators select the preferred response. Its parameters are learned through logistic regr…
Information retrieval is the human-made discipline of organizing, storing, and retrieving information from collections to satisfy user information needs. It comprises: (1) a document corpus (text, structured data, or multimodal content), (2) an indexing mecha…
Stochastic gradient descent is an optimization algorithm that iteratively updates parameters by computing the gradient of an objective function on randomly sampled mini-batches of data rather than the full dataset, trading gradient accuracy for computational…
An objective function is a human-authored mathematical mapping that assigns a scalar value to each point in a parameter space, quantifying how closely a given configuration satisfies a defined goal. It takes as input a vector of decision variables (parameters…
A latent space is the human-made mathematical construct that organizes abstract representations of data into a structured coordinate system, where each dimension encodes a learned feature and distances between points reflect semantic or statistical similarity…
Curriculum learning is a training strategy in machine learning in which examples are presented to the model in a specific order—typically progressing from easy to difficult—so that the learner develops robust internal representations before tackling harder ca…
Optimization landscape is the conceptual model mapping all possible configurations of a parameterized system to the values of an objective function assigned to each configuration, forming a structured terrain of hills and valleys that guides algorithmic searc…
Significance level (alpha) is a human-made threshold in statistical hypothesis testing — the maximum acceptable probability of a Type I error (rejecting a true null hypothesis). Standard values are 0.05 and 0.01, chosen by convention to balance error rates ag…
Computational complexity is a theoretical framework that classifies computational problems by the resources — primarily time and space — required to solve them, regardless of any specific algorithm or machine implementation. It maps problems into complexity c…
Self-supervised learning is a machine learning paradigm where models extract supervision from unlabeled data by generating pseudo-labels from the data's own structure. The method creates a supervised learning task — such as predicting a missing token in a seq…
Few-shot learning is a machine learning paradigm where a model acquires the ability to perform a new task from K exemplars (typically K ∈ {1, 5, 10, 32}), by transferring structural knowledge acquired from a distribution of related source tasks. Parameters: K…
A sampling distribution is the probability distribution of a statistic (such as the mean, variance, or proportion) obtained by repeatedly drawing random samples of a fixed size from a population. It describes how the value of the statistic varies across diffe…
Bootstrap is a resampling technique that estimates the sampling distribution of a statistic by repeatedly drawing random samples with replacement from the observed dataset. Parameters: the original sample of n observations, the statistic to estimate (mean, va…
Measurement error is the human-made statistical construct that quantifies the difference between an observed measurement and the true value of a quantity. It decomposes into systematic error (bias) and random error (variance), and persists through calibration…
Autocorrelation is a statistical measure of the degree of similarity between a time series and a lagged version of itself over successive time intervals. It is computed as the correlation coefficient between observations separated by k time steps (the lag), r…
A statistical test is a formal procedure for evaluating evidence about a population parameter or model assumption using sample data. It operates by: (1) specifying a null hypothesis (H0) representing a baseline claim and an alternative hypothesis (H1); (2) se…
A hypothesis test is a statistical procedure for making decisions about population parameters based on sample data. It formalizes the process of evaluating evidence by specifying a null hypothesis (default assumption), an alternative hypothesis, a test statis…
Omitted variable bias is a statistical error that occurs when a regression model excludes one or more relevant variables that are correlated with both the dependent variable and one or more included independent variables. The effect of the omitted variable ge…
Degrees of freedom is a statistical concept specifying the number of independent values in a dataset that can vary freely when estimating parameters under given constraints. When fitting a model with k parameters to n observations, the residual degrees of fre…
Data cleaning is a quality assurance practice that identifies, corrects, or removes errors, inconsistencies, and irrelevant data from datasets to improve data quality for analysis. It encompasses operations including handling missing values (imputation, delet…
A prediction interval is the human-made statistical procedure that, given a sample and a statistical model, produces a range expected to contain a future observation with a specified probability. It differs from a confidence interval by accounting for both pa…
A standardized mean difference is a quantitative measure of the magnitude of difference between two groups or conditions, expressed in units of standard deviation. It is calculated by subtracting the mean of one group from the mean of another and dividing by…
Bayesian inference is the human-made statistical method that updates beliefs about unknown parameters using Bayes' theorem, combining prior distributions with observed likelihoods to produce posterior distributions. It persists through probabilistic programmi…
A graphical method for displaying multivariate data on a two-dimensional plane using a radial layout. Multiple quantitative variables are represented as axes originating from a common center, with data points plotted along each axis and connected to form a cl…
A quantile-quantile (Q-Q) plot is a human-made graphical method for comparing two empirical probability distributions. It carves by requiring (1) two datasets of observed values, (2) computation of quantiles for each dataset at matching probability levels, an…
A probability plot is a human-made graphical method for comparing the distribution of observed data against a theoretical probability distribution. It carves by requiring (1) observed data values sorted and assigned plotting positions, (2) a reference distrib…
A pipeline is a human-made sequence of processing stages arranged so that the output of each stage becomes the input of the next, creating a directed chain of transformations. It carves by requiring (1) ordered stages, (2) material or data flow between consec…
Graph theory is a mathematical framework for studying structures composed of nodes (vertices) connected by edges (links). It provides formal definitions for graph types (directed, undirected, weighted, bipartite), operations (traversal, embedding, decompositi…
A statistical model is a formal mathematical representation of relationships among variables, constructed to approximate, explain, or predict phenomena observed in data. It specifies a family of probability distributions parameterized by quantities of interes…
Intrinsic reward is a human-made psychological construct describing satisfaction that arises from the performance of an activity itself, rather than from separation of external incentives. It carves the boundary between internally generated positive affect (c…
Usability is a human-made quality attribute of interactive systems that measures how effectively, efficiently, and satisfactorily specified users can achieve specified goals in a given context of use. It carves by requiring (1) identifiable users with defined…
A skeuomorph is a design element whose visual appearance references an earlier physical object or medium, while the referenced function or form has been lost or altered in the new context. Skeuomorphs preserve the visual language of analog predecessors — the…
A user feedback construct is a deliberately designed signal produced by a computing system to communicate operational state, action results, or progress to the human operator. It manifests through visual indicators (color changes, animations, progress bars),…
Underfitting is a condition in supervised machine learning where a model lacks sufficient capacity to capture the underlying pattern in the training data, producing high error on both training and held-out sets. The parameters are: (1) a trained model with fi…
Gesture recognition is a computational technique that maps patterns in sensor data to intentional human body movements — hand signs, head nods, full-body motions — using pattern-matching algorithms trained on annotated gesture corpora. Its parameters are: (1)…
A capacitive stylus is a pointing instrument that emulates the electrical charge of a human fingertip so it can be detected by capacitive touchscreens. It consists of a conductive tip (typically rubber or transparent conductive material) attached to an insula…
A gradient is the vector of partial derivatives of a multivariable function, computed by differentiating the function with respect to each independent variable. It points in the direction of steepest ascent, and its magnitude gives the rate of change in that…
Haptic feedback is a human-made sensory interface technique that renders tactile sensations — vibration, force, texture simulation, or pressure — to transmit data through the user's sense of touch. It operates via actuators such as eccentric rotating mass mot…
A touchscreen is a human-made display surface that combines a visual monitor with an input device, allowing direct tactile interaction with on-screen elements. Built to persist through standardized interfaces (capacitive, resistive, infrared), teaching in HCI…
A sensor is a transducer device that converts a physical, chemical, or biological stimulus into a quantifiable electrical signal. Its parameters are: (1) a sensing element that interacts with the measurand, (2) a transduction mechanism (piezoelectric, thermoe…
Sensor fusion is a signal processing methodology that integrates data streams from multiple heterogeneous sensors through computational algorithms (Kalman filtering, Bayesian inference, Dempster-Shafer theory) to produce a unified state estimate with greater…
A washing machine is a human-made household appliance that cleans textiles (clothing, linens, towels) using water, detergent, and mechanical agitation within a programmable cycle. It operates through a sequence of filled, agitated, drained, and spun phases co…
An epoch is a training hyperparameter specifying one complete cycle through the full training dataset. Its parameter is the integer count of cycles; its persistence mechanism is procedural discipline — it structures the temporal organization of training by go…
Mini-batch gradient descent is an optimization algorithm that partitions training data into subsets of fixed size (the mini-batch size), computing a gradient estimate from each subset before updating model parameters. Its defining parameters are the batch siz…
Iteration is the human-made practice of repeating a structured process or sequence of operations to progressively approach a desired result. Each cycle takes the output of the previous cycle as input, refining the state through systematic repetition. The para…
Batch size is a hyperparameter in machine learning training that determines the number of training samples processed in a single forward and backward pass before the model's weights are updated via an optimizer. It partitions the training dataset into mini-ba…
Optics is the branch of physics that studies the behavior and properties of light — electromagnetic radiation across a range of wavelengths — and its interactions with matter. It carves out a domain delimited by its subject (light and its phenomena: reflectio…
A learning rate schedule is a protocol for modulating the learning rate — the step size parameter in iterative optimization — as training progresses. It specifies a function or policy that maps training progress (measured in epochs, steps, or wall-clock time)…
A learning rate is a scalar hyperparameter in machine learning optimization that determines the step size at each iteration while moving toward a minimum of a loss function. Parameters: the initial learning rate (a positive real number), optional scheduling s…
A training loop is a human-made procedural construct in machine learning: an iterative cycle that repeatedly feeds data through a model, computes error via a loss function, computes gradients via backpropagation, and updates parameters via an optimizer. Its p…
An optimizer is a human-made algorithm that computes parameter updates from gradients during model training. It specifies the rule for combining gradient information with previous updates — through mechanisms like momentum, adaptive scaling, or decoupled weig…
A neural network architecture is a structured design specifying how computational layers are organized, interconnected, and oriented to process information. It defines the layer types (convolutional, recurrent, attention-based, fully connected), their arrange…
An optimization algorithm is a procedure that iteratively modifies a candidate solution to a problem so as to minimize or maximize a objective function. Parameters include the search space, the objective function, the update rule (e.g. gradient descent, simul…
The chain rule is a rule in calculus for computing the derivative of a composite function. Given two differentiable functions f and g, the derivative of f∘g is (f∘g)\'(x) = f\'(g(x)) · g\'(x). The rule is parameterized by the outer function, the inner functio…
backpropagation is a human-made algorithm for computing the gradient of a loss function with respect to all weights in a multi-layer computational graph, most commonly a neural network. It works by applying the chain rule of calculus recursively from the outp…
Gradient clipping is a technique that bounds the magnitude of gradients computed during backpropagation to prevent them from exceeding a specified threshold. It is human-made: a deliberate algorithmic intervention designed to stabilize training of neural netw…
Weight decay is a human-made regularization technique used in training neural networks and other statistical models. It adds a penalty term proportional to the squared L2 norm (sum of squared weights) to the loss function, constraining model capacity by penal…
Batch normalization is a human-made neural network training technique that normalizes the activations of each layer across the mini-batch to produce zero-mean, unit-variance outputs, reducing internal covariate shift and enabling higher learning rates. It com…
Hyperparameter optimization is a human-made search methodology for finding the optimal configuration of hyperparameters — learning rates, network depths, regularization coefficients, batch sizes — in machine learning model training. It operates over a defined…
Dropout is a human-made regularization technique in neural network training that randomly sets a fraction of neuron activations to zero during each forward pass, with the dropout rate controlled by a hyperparameter. By preventing neurons from co-adapting — fo…
Learning rate scheduling is a human-made optimization technique in iterative machine learning training that adjusts the learning rate — the step-size parameter controlling optimizer update magnitude — according to a predefined schedule across training steps o…
A transfer function (also called activation function in neural networks) is a mathematical function that transforms an input signal or vector into an output signal, typically introducing non-linearity into a computational system. In neural networks, each neur…
Early stopping is a human-made regularization technique in iterative machine learning training that halts the optimization process when a held-out validation metric ceases to improve, thereby preventing overfitting to training data. It persists through standa…
Model evaluation is the structured practice of testing a trained statistical or machine learning model against held-out data and acceptance criteria to confirm it meets its intended performance standards. It persists through standardized benchmark suites, def…
Training data is the curated set of examples (inputs paired with target outputs, or raw observations) that an agent or model consumes during a learning procedure to adjust its internal parameters. It is human-made: datasets are collected, filtered, annotated,…
Representation learning is a human-made machine learning approach that discovers or constructs internal encodings of data to make downstream learning easier, rather than relying on hand-engineered features. It transforms raw inputs into structured internal re…
A student model is a machine learning construct: a neural network trained to mimic the outputs or latent representations of a larger teacher model, typically for compression or efficiency. Training minimizes divergence between student and teacher predictions…
Model compression is the family of techniques that reduce the size or computational cost of a trained machine learning model while preserving as much of its predictive performance as possible. It persists through standardized methods (quantization, pruning, k…
Teacher model is a machine learning construct: a pre-trained neural network or ensemble whose outputs (predictions, embeddings, or hidden representations) serve as supervisory targets for training a smaller student model in knowledge distillation. Parameters:…
Knowledge distillation is a model compression technique in which a smaller 'student' neural network is trained to replicate the output probability distribution of a larger 'teacher' model. Parameters: (1) a pre-trained teacher model with established weights,…
Retrieval-augmented-generation (RAG) is a machine learning architecture in which a pretrained language model is combined with an external document retrieval component: given a query, the system first retrieves relevant passages from a knowledge store, then ge…
Meta-learning is a machine learning paradigm in which the learning algorithm is itself optimized using experience — typically measured as performance across a distribution of related tasks or training episodes. The system learns an induction bias, initializat…
Fine-tuning is a machine learning technique in which a model trained on a broad pre-training dataset undergoes additional supervised training iterations on a narrower, task-specific dataset. The process preserves the general representations acquired during pr…
Generalization error is the gap between a machine learning model's performance on its training data (empirical risk) and its performance on new, unseen data drawn from the same distribution (expected risk). It arises because a model fits not only the underlyi…
Multicollinearity is the condition in multiple regression analysis where two or more predictor variables exhibit substantial linear correlation, such that the design matrix approaches singularity and coefficient estimates become imprecise and unstable. It is…
Goodness-of-fit is a class of statistical measures that quantify the agreement between observed data and a hypothesized model or distribution. It operates by computing a test statistic from sample data and comparing it against a theoretical null distribution…
Model validation is the structured practice of testing a trained statistical or machine learning model against held-out data and acceptance criteria to confirm it generalizes beyond its training set. It carves a distinct phase from model evaluation (which mea…
Outlier detection is the systematic practice of identifying observations that deviate markedly from the dominant pattern in a dataset. It operates through three complementary mechanisms: (1) threshold-based statistical tests (z-score, IQR rule, Grubbs', Dixon…
Language assessment is a systematic methodology for evaluating a person's proficiency in one or more languages. It operates through standardized instruments (tests, rubrics, performance tasks) that measure specific language domains—listening, speaking, readin…
Model diagnostics is the systematic practice of checking whether a fitted statistical model meets its underlying assumptions and provides an adequate fit to the data. It comprises four core components: (1) residual analysis — examining patterns in residuals t…
Robust statistics is a branch of statistical methodology designed to produce estimates and inferences that remain reliable even when model assumptions are violated — particularly in the presence of outliers, heavy-tailed error distributions, or contaminated d…
Conformity is the act of adjusting one's behavior, beliefs, or expressed opinions to align with perceived group norms or standards. It operates through two mechanisms: behavioral compliance (public alignment without private acceptance) and internalization (ge…
DFFIT is a regression diagnostic measure that computes the standardized change in the fitted value for each observation when that observation is deleted from the model. Parameters: dffiti = (yhat_i - yhat_i(i)) / (s_i * sqrt(h_ii)), where yhat_i is the fitted…
The Breusch-Pagan test is a statistical procedure for testing whether the variance of errors in a regression model depends on values of the explanatory variables. It fits an auxiliary regression of squared OLS residuals on the predictors, computes a test stat…
The Durbin-Watson test is a statistical procedure for detecting first-order autocorrelation in the residuals of an ordinary least squares regression. It computes the statistic d = Σ(eₜ − eₜ₋₁)² / Σeₜ², where eₜ denotes the residual at observation t. The stati…
Press residual (also called the studentized deleted residual or leave-one-out residual) is a diagnostic measure in regression analysis that quantifies the discrepancy between an observation's actual value and the value predicted by a model fit to all other ob…
A human-made criterion framework for assessing and categorizing language proficiency levels. The constructed scales (CEFR A1–C2, ACTFL Novice–Distinguished, IELTS 1–9) partition the continuous dimension of communicative ability into discrete, teachable levels…
Regression diagnostics are statistical methods for evaluating whether a fitted regression model meets its assumptions and identifying observations that disproportionately affect the fit. Parameters: diagnostics operate across three axes — (1) goodness-of-fit…
The covariance ratio (covratio) quantifies the effect of deleting observation i on the estimated covariance matrix of regression coefficients. Formally, covratio_i = |XᵀX| / |(X_{-i})ᵀ(X_{-i})| × |s²_{-i}(X_{-i})ᵀ(X_{-i})| / |s²(XᵀX)| where X is the design ma…
Influence diagnostics are regression techniques that quantify how much individual observations affect model outputs — coefficients, predictions, and fit statistics. Parameters: each observation is individually removed or down-weighted, and the resulting chang…
Taxonomies are human-made hierarchical classification systems that organize entities into structured categories based on shared properties and distinguishing criteria. Parameters: root category, depth of hierarchy, branching factor at each level, classificati…