Artificial Intelligence Technical Terminology & Mathematical Glossary

An autonomous computational entity that perceives its environment through sensors and executes actions via actuators to maximize its expected performance metric based on its perceptual history.

The formal engineering specification of an intelligent agent: Performance measure (utility to maximize), Environment (external context), Actuators (action tools), and Sensors (perceptual inputs).

A mathematical framework for modeling decision-making in environments where outcomes are partly random and partly under the control of a decision-maker, satisfying the Markov memoryless property.

A fundamental recursive equation expressing that the value of an optimal state equals the immediate reward plus the discounted expected value of the best successor state.

The discrete or continuous graph of all reachable problem states. Heuristic search (e.g. A* search) uses an evaluation function combining path cost g(n) with estimated distance to goal h(n) to avoid combinatorial explosion.

An operational test of machine intelligence proposed by Alan Turing (1950). A computer passes if an interrogator cannot reliably distinguish its natural-language responses from those of a human.

John Searle's 1980 thought experiment demonstrating that formal syntactic rule execution (e.g. lookup tables or program code) does not inherently generate genuine semantic understanding or conscious intentionality.

A machine learning paradigm where an algorithm learns a mapping function from input features x to target labels y using a ground-truth labeled training dataset.

Discovering latent patterns, geometric clusters, or probability distributions from unlabeled empirical data without explicit human supervision.

A goal-oriented framework where an agent learns an optimal behavioral policy through trial-and-error interactions with a dynamic environment, receiving scalar rewards or penalties.

The central dilemma in statistical modeling: High bias causes underfitting (oversimplifying patterns), while high variance causes overfitting (capturing random training noise).

Overfitting occurs when a model memorizes idiosyncrasies in the training set and fails to generalize to validation data. Underfitting occurs when the model is too simple to capture the underlying data manifold.

Mathematical penalties added to the loss function to constrain weight magnitude, discouraging complexity and preventing overfitting. L1 promotes feature sparsity; L2 penalizes large outlier weights.

An iterative first-order optimization algorithm that calculates the gradient vector of the loss function with respect to model parameters and steps in the direction of steepest descent. Stochastic Gradient Descent (SGD) uses mini-batches.

A contingency table recording True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN). Precision = TP/(TP+FP); Recall = TP/(TP+FN); F1 is their harmonic mean.

Receiver Operating Characteristic plotting True Positive Rate vs False Positive Rate across all classification decision thresholds. Area Under the Curve (AUC) measures aggregate threshold-independent separability.

A biologically inspired computational model composed of interconnected artificial nodes (neurons) organized in layers, performing non-linear affine transformations.

Mathematical non-linearities applied to weighted sums, enabling neural networks to approximate arbitrary non-linear functions (Universal Approximation Theorem).

An algorithmic implementation of the multivariable calculus chain rule that computes the partial derivative of the loss function with respect to every weight in reverse order from output to input.

In deep networks, backpropagating gradients through many layers can exponentially decay to zero (halting learning) or compound to infinity (causing numeric overflow).

A neural architecture tailored for grid-like topology (images) utilizing parameterized convolutional kernels that slide over inputs, providing spatial translation invariance and parameter sharing.

Dropout randomly deactivates a fraction of neurons during training to prevent co-adaptation. Normalization stabilizes internal covariate shifts across layer distributions, accelerating training convergence.

A compressed, high-dimensional geometric vector space where semantically similar entities (words, images, concepts) are positioned close together according to distance metrics.

A computational mechanism mapping Query (Q), Key (K), and Value (V) projections to dynamic pairwise attention scores between all tokens in a sequence, eliminating sequential recurrent bottlenecks.

Projecting Q, K, and V into multiple lower-dimensional representation subspaces in parallel, allowing the model to simultaneously attend to syntax, semantics, and long-range dependencies.

A sequential generation strategy where each output token is predicted based solely on preceding tokens, enforced during training using lower-triangular causal attention masking.

The algorithmic conversion of raw text strings into discrete integer indices using subword algorithms such as Byte-Pair Encoding (BPE), resolving vocabulary bloat and out-of-vocabulary terms.

The generation of factually incorrect, unverifiable, or nonsensical assertions presented with authoritative linguistic confidence by an autoregressive model due to statistical probability sampling.

An enterprise AI architecture that queries an external vector database for verified reference documents matching a user prompt and injects the retrieved context into the model's prompt.

Low-Rank Adaptation freezes pre-trained model weights W_0 and injects trainable rank decomposition matrices (A and B). QLoRA quantizes the base model to 4-bit NormalFloat (NF4) for low-VRAM GPUs.

A fine-tuning pipeline where a reward model is trained on human preference ratings, and the generative policy is optimized via Proximal Policy Optimization (PPO) or Direct Preference Optimization (DPO).

An inference optimization technique that stores computed Key and Value tensor projections of historical tokens in GPU memory, avoiding redundant matrix multiplications during next-token decoding.

A sparse neural architecture where a gating router network dynamically activates only a small subset of total parameters (experts) for each individual token.

Simultaneously predicting the categorical label and spatial localization coordinates (bounding box) of multiple distinct foreground objects within a 2D image or 3D point cloud.

Semantic segmentation assigns a categorical label to every pixel without distinguishing separate objects. Instance segmentation delineates individual object boundaries as distinct entities.

An evaluation metric quantifying the spatial overlap accuracy between a predicted bounding box and the ground-truth annotation. An IoU ≥ 0.5 is typically considered a True Positive.

Training robotic control policies inside ultra-fast physics simulation engines while aggressively randomizing physical parameters (mass, friction, lighting) so the policy generalizes to reality.

The fundamental engineering challenge of ensuring autonomous AI systems pursue objectives and terminal goals that are strictly aligned with human values, ethics, and intended outcomes.

Methodologies providing human-interpretable explanations of black-box model decisions. SHAP calculates Shapley values from cooperative game theory to quantify the exact contribution of each feature.

An automated safety alignment framework pioneered by Anthropic where AI systems use a list of explicit constitutional principles to critique, revise, and align their own responses without human labeling.

Global benchmark legislation classifying AI systems into four risk tiers: Unacceptable risk (banned social scoring), High risk (strict audits for medical/employment AI), and Minimal risk.