MAI key terms

A training approach where a model learns from labeled input-output pairs, adjusting its parameters to map inputs to known target values so it can predict those targets on new data.

A training approach that finds structure in unlabeled data, such as clusters, dense regions, or lower-dimensional representations, without any target answer being supplied during learning.

The target value or correct answer attached to a training example, such as a class name, a numeric score, or a bounding box, which the model is trained to reproduce or predict.

A failure mode where a model fits idiosyncrasies and noise in training data so closely that it performs well there but poorly on unseen data, showing a large gap between training and validation error.

Any technique that deliberately constrains model complexity, such as adding a penalty on large weights, early stopping, or dropout, to reduce overfitting and improve generalization to new data.

A resampling procedure that splits data into complementary folds, training on some and validating on the held-out rest repeatedly, to estimate generalization performance more reliably than a single split.

A modeling error where information unavailable at prediction time, such as target values or statistics computed over the full dataset, contaminates training and produces optimistic evaluation results that collapse in production.

The practice of transforming raw variables into informative inputs, for example by encoding categories, scaling, binning, or deriving ratios, so a model can learn patterns more effectively.

A categorical encoding that replaces one column with a binary column per category, marking the present category with a one and all others with zeros, avoiding any implied ordering between categories.

Rescaling numeric inputs to comparable ranges, through normalization or standardization, so distance-based and gradient-based algorithms converge faster and are not dominated by large-magnitude variables.

A dataset condition where one class vastly outnumbers another, which biases models toward the majority and makes accuracy misleading, often addressed with resampling, class weights, or threshold tuning.

Precision is the share of predicted positives that are truly positive; recall is the share of actual positives that were found. Together they describe performance when class distributions are skewed.

The harmonic mean of precision and recall, giving a single balanced metric that penalizes models strong on one measure but weak on the other, useful for imbalanced classification tasks.

A table crossing predicted against actual classes, showing true positives, false positives, false negatives, and true negatives, from which many classification metrics can be derived.

An unsupervised technique that groups examples by similarity so members within a group are closer to each other than to members of other groups, used for segmentation and exploratory analysis.

A nonlinear function applied to a neuron's weighted sum, such as ReLU or sigmoid, that lets a neural network represent complex relationships rather than only linear ones.

The algorithm that computes gradients of a loss with respect to each network weight by applying the chain rule backward through the layers, enabling gradient-based parameter updates.

An iterative optimization method that updates parameters in the direction that reduces a loss function, scaled by a learning rate, until the model converges or training is stopped.

A hyperparameter controlling the size of each optimization step; too large causes instability or divergence, too small makes training slow or stuck in poor regions.

A training problem in deep networks where gradients shrink toward zero as they propagate backward through many layers, stalling learning in early layers; mitigated by ReLU, normalization, and residual connections.

A regularization technique that randomly deactivates a fraction of neurons during each training step, preventing units from co-adapting and improving generalization at inference time.

A network architecture that applies learned filters over local spatial regions with weight sharing, making it highly effective for images and other grid-structured data while using far fewer parameters than dense layers.

A neural architecture built on self-attention that weighs relationships between all positions in a sequence in parallel, forming the basis of most modern large language models.

A computation that assigns each position a weighted view of other positions based on learned query-key similarity, letting a model focus on relevant context regardless of distance in the sequence.

A unit of text, roughly a word fragment, produced by a tokenizer; models read and generate sequences of tokens, and token counts drive context limits and usage-based pricing.

The maximum number of tokens a language model can attend to in a single request, spanning prompt and generated output, which bounds how much material can be reasoned over at once.

A pattern that retrieves relevant documents from an external store and inserts them into the prompt so the model grounds its answer in current or private knowledge rather than relying only on trained parameters.

A dense numeric vector representing text, images, or other objects such that semantically similar items lie close together, enabling similarity search, clustering, and retrieval.

A model output that is fluent and confident but factually wrong or unsupported by any source, often arising when the model generates plausible-sounding content beyond what its evidence supports.

Further training of a pretrained model on a smaller, task-specific dataset so its weights adapt to a particular domain, style, or behavior, typically cheaper than training from scratch.

Reusing knowledge from a model trained on one large task as the starting point for a related task, cutting data and compute needs compared with training from random initialization.

A control that constrains model behavior, such as input filtering, output validation, topic restrictions, or human review, to keep responses safe, compliant, and within intended scope.

Systematic unfairness in data or model outputs that disadvantages particular groups, often mirroring historical inequities in training data or measurement choices, and requiring detection and mitigation.

The degree to which a person can understand why a model produced a given output, supported by techniques such as feature attribution or surrogate models, and increasingly required for accountability.

Degradation of model performance over time as real-world data distributions or input-output relationships change after deployment, detected through ongoing monitoring and remedied by retraining.

An experiment that randomly splits users between a control and a treatment variant and compares outcomes statistically, used to validate that a deployed AI change truly improves results.

The smallest release that delivers real value and generates learning, used in AI projects to test whether a use case is worth further investment before committing to full-scale development.

The financial benefit of an initiative relative to its cost, calculated from expected gains such as saved labor, higher revenue, or reduced risk against build and operating expenses.

The set of practices and tooling for reliably building, deploying, monitoring, and updating machine learning systems in production, combining software engineering discipline with model lifecycle management.

A central catalog that versions trained models with their metadata, metrics, and lineage, enabling reproducible deployment, rollback, and audit of which model is serving traffic.