Publication
An intellectual research work, independent of venue or version, that addresses research areas and may use, study, or produce research objects.
Ontology Dictionary
A typed, navigable dictionary of the classes, relationships, inference rules, and reusable entries in this research ontology.
Dictionary
Browse the vocabulary used to describe and connect the research on this site.
Ontology classes
Research area
A sustained field or line of inquiry addressed by the research.
Research object
An entity, structure, process, property, phenomenon, statement, or system that research can use, study, or produce.
AlgorithmSubclass of Research object
A specified computational or mathematical procedure for training, inference, intervention, or analysis.
ModelSubclass of Research object
A mathematical or computational system used to represent data, make predictions, generate observations, or describe dynamics.
ObjectiveSubclass of Research object
A scalar criterion, loss, or utility whose value defines what an algorithm seeks to optimize.
Relationship vocabulary
Predicates are authored in the displayed direction. The inverse reading names the same assertion when followed from range to domain.
| Predicate | Inverse reading | Meaning | Domain | Range |
|---|---|---|---|---|
| Research assertions | ||||
addressesaddresses | addressed by | Connects a publication to a principal research area to which it contributes. | Publication | Research area |
applies-toapplies to | has applicable result | Connects a research result or other research object to the object or setting for which it holds. | Research object | Research object |
characterizescharacterizes | characterized by | Connects a research result or other research object to a phenomenon or quantity whose behavior it describes. | Research object | Research object |
compares-withcompares with | compared with | Connects publications that directly compare their findings, methods, or conclusions. | Publication | Publication |
enablesenables | enabled by | Connects a research object to an outcome or capability it makes possible. | Research object | Research object |
extendsextends | extended by | Connects a publication to an earlier publication whose result or analysis it directly generalizes or develops. | Publication | Publication |
learnslearns | learned by | Connects an algorithm to the research object it learns or recovers. | Algorithm | Research object |
optimizesoptimizes | optimized by | Connects an algorithm to the objective it is designed to optimize. | Algorithm | Objective |
part-ofpart of | has part | Relates a research object to a larger research object of which it is a component. | Research object | Research object |
producesproduces | produced by | Connects a publication to a research object introduced or established as an output of the work. | Publication | Research object |
relatedrelated to | related to | Records a direct associative connection when no more specific predicate applies. | Research area, Research object | Research area, Research object |
reproducesreproduces | reproduced by | Connects a publication to a publication whose reported result it attempts to reproduce. | Publication | Publication |
specialization-ofspecialization of | has specialization | Relates a more specific research object to the broader research object of which it is a kind. | Research object | Research object |
studiesstudies | studied by | Connects a publication or research area to a research object it investigates as a subject. | Publication, Research area | Research object |
subarea-ofsubarea of | has subarea | Relates a narrower research area to the broader research area that contains it. | Research area | Research area |
used-inused in | uses | Connects a research object to a research area in which it is used. | Research object | Research area |
usesuses | used by | Connects a publication to a research object it employs without presenting that object as its primary subject or output. | Publication | Research object |
| Graph structure | ||||
instance-ofinstance of | has instance | Assigns an ontology entity to the class whose membership it instantiates. | Ontology class | |
subclass-ofsubclass of | has subclass | Relates a narrower ontology class to a broader class whose members include every member of the narrower class. | Ontology class | Ontology class |
Inference rules
These entailments are computed from authored assertions and remain marked as inferred in the graph.
Defined entries
Adversarial robustness
The study of how learning systems behave under deliberately chosen perturbations intended to cause errors or expose vulnerabilities.
Computer vision
The study of computational systems that infer structure, meaning, or action from visual data.
Developmental interpretability
The study of how interpretable representations, mechanisms, and circuits form and change during training.
Implicit bias
The study of how a learning algorithm selects among multiple solutions even without an explicit preference in the objective.
Mechanistic interpretability
The study of how internal components and computations of learned models produce their behavior.
Meta-learning
The study of systems that use experience across tasks to learn how to adapt to a new task.
Statistical learning theory
The mathematical study of what can be learned from finite data and how performance depends on data, model classes, objectives, and algorithms.
Statistical modeling
The construction and analysis of explicit probabilistic models for how data are generated.
Training dynamics
The study of how a model's parameters, representations, and computations evolve during learning.
Convergence guarantee
A formally justified statement that an iterative procedure approaches a specified solution set or performance criterion under explicit assumptions.
Forward invariance
A property of a subset of a dynamical system's state space such that every trajectory beginning in the subset remains there at all later times for which it is defined.
Generalization
The ability of a learned model to perform well on data that was not used to fit its parameters.
Generalization bound
A formally justified bound on the gap between performance on observed data and performance on an underlying population.
Hypothesis-class stability
A uniform regularity property that bounds how much the predictions of models in a hypothesis class can change under specified transformations of their input context.
In-context learning
The adaptation of a model's predictions to examples or instructions supplied in its input context without updating model parameters.
Induction head
An attention-head mechanism that uses earlier context matches to promote a corresponding continuation token.
Interpretable direction
A direction in a learned feature or latent space associated with a recognizable factor of variation.
Invariance
A property under which a specified transformation or evolution preserves a mathematical object, relation, set, or system behavior.
Label noise
Corruption in which observed training labels differ from the labels specified by an underlying target process.
Learned representation
A feature space or encoding produced by learning that exposes structure useful for prediction, transfer, or analysis.
Learning without mixing
A statistical phenomenon in which learning from dependent observations avoids a multiplicative sample-complexity penalty from the mixing time.
Mechanistic characterization
A scientific account that identifies an internal component, computation, or circuit responsible for observed behavior.
Next-token prediction
The task of predicting a conditional distribution for the next token given a preceding token context.
Overparameterization
A regime in which a model has more adjustable degrees of freedom than are needed to represent or interpolate the target structure.
Permutation invariance
An invariance under a specified relabeling or reordering of elements.
Pretraining
Training performed before a model is adapted to a target task or used in its target setting, typically on broad data with a general or self-supervised objective.
Rephrasability
A property of a sequential data-generating process that relates prediction on a new context to denoising problems over contexts generated by the process.
Reproduction result
A documented assessment of whether and under what conditions previously published computational findings can be reproduced.
Sample complexity bound
A formally justified bound on how much data is sufficient or necessary to achieve a specified learning objective.
Sparse attention pattern
An attention structure in which substantial interaction weight is concentrated on a restricted subset of token positions.
Theoretical guarantee
A formally stated and justified claim about the behavior of a system under explicit assumptions.
Adversarial training
A training algorithm that fits a model on deliberately challenging perturbations chosen to expose its vulnerabilities.
Contrastive learning
A family of learning algorithms that forms representations by comparing samples designated as similar or dissimilar.
Data augmentation
An algorithmic procedure that expands or transforms training examples to encode useful variation and invariance.
Empirical risk minimization
A learning rule that selects a predictor by minimizing average loss over observed data within a specified hypothesis class.
First-order ANIL
A first-order meta-learning algorithm that adapts a task-specific output layer while learning a shared representation across tasks.
Gradient clipping
An algorithmic operation that limits the magnitude of gradients or parameter updates during optimization.
Regularization
A family of algorithmic techniques that constrain or bias learning to promote selected properties of the learned solution.
Self-attention
An attention operation in which queries, keys, and values are derived from representations in the same sequence.
Autoregressive process
A sequential stochastic model whose conditional distribution at each step is parameterized as a function of preceding observations.
Disentangled Transformer
An attention-only Transformer variant that removes multilayer perceptrons and replaces additive residual connections with concatenation so that successive computations remain explicit.
Generative adversarial network
A generative model comprising a generator trained against a discriminator that distinguishes generated samples from observed data.
Generative model
A model of a data-generating distribution from which new observations can be sampled.
Language model
A probability model over sequences of tokens, used to score or generate text and other sequential data.
Linear dynamical system
A dynamical model whose state evolution and observation equations are linear in their state and input variables.
Markov chain
A stochastic process whose next-state distribution depends on the past only through a specified current state.
Neural network
A parameterized model composed of interconnected computational units arranged in layers or another directed computation graph.
Normalizing flow
A probabilistic model family that transforms a simple base distribution through a sequence of invertible mappings.
Training-trajectory model
A formal or empirical model that describes how a learning system changes over the course of training.
Transformer
A family of neural sequence architectures built around attention-based interactions between positions.
Vector autoregressive process
A multivariate stochastic process in which each vector-valued state is a linear combination of a fixed number of preceding states plus an innovation term.
Adversarial training objective
A robust objective that evaluates loss on perturbations selected to increase error within a specified threat set.
Contrastive objective
A loss that rewards selected examples or views for receiving more similar representations than designated alternatives.
Cross-entropy loss
A loss that scores a predicted probability distribution by the negative log-probability assigned to the observed outcome.
Empirical risk
An objective formed by aggregating a loss over an observed sample.
Loss function
An objective that assigns a scalar cost to a prediction, action, parameter setting, or comparison with a target.
Orthogonality penalty
A regularization objective that penalizes deviation from an orthogonality constraint among selected vectors or transformations.
No ontology entries match these filters.