AIT key terms

The field of building computer systems that perform tasks normally requiring human intelligence, such as reasoning, learning, perceiving, and decision-making. AI is the broad umbrella covering machine learning, NLP, robotics, and related subfields.

AI systems designed to perform a specific, bounded task, such as image classification or language translation. Nearly all deployed AI today is narrow, lacking general reasoning ability outside its trained purpose.

A hypothetical system that can learn and reason across any domain at human level or beyond, transferring knowledge between tasks. It does not exist today; claims about it are speculative and often overstated.

A finite, step-by-step procedure for solving a problem or producing an output from inputs. In AI, algorithms define how a model learns from data and how it computes predictions.

The dataset used to fit a model's parameters. Its quality, representativeness, and licensing strongly shape model behavior, making data provenance a core governance concern.

The individual measurable input variables a model uses to make predictions, such as pixel values, word counts, or customer attributes. Good feature selection often matters more than model choice.

The target or correct answer attached to a training example in supervised learning, such as a class name or numeric value. Labels are produced by annotation or measurement and can carry error and bias.

A training approach where the model learns a mapping from inputs to known labels, such as classifying emails as spam or not. It requires labeled data and is used for classification and regression.

A training approach that finds structure in unlabeled data, such as clusters or lower-dimensional representations. It is used for segmentation, anomaly detection, and exploratory analysis.

A training approach where an agent learns by taking actions in an environment and receiving rewards or penalties. It suits sequential decision problems like game play and robotic control.

A supervised task predicting a discrete category, such as fraud or not fraud. Performance is judged with metrics like accuracy, precision, recall, and F1 score, not just overall correctness.

A supervised task predicting a continuous numeric value, such as price or demand. Performance is typically measured with error metrics like mean absolute error or root mean squared error.

An unsupervised technique that groups similar data points together based on distance or similarity measures. It is used for customer segmentation, topic discovery, and anomaly detection.

When a model fits training data too closely, including its noise, and performs poorly on new data. It is countered with more data, regularization, simpler models, and validation.

When a model is too simple or poorly trained to capture the underlying pattern, performing badly on both training and new data. It signals the model or features need more capacity or better inputs.

A held-out portion of data used to tune model settings and compare candidates during development. It must remain separate from the test set to give an honest performance estimate.

A held-out dataset used only once, at the end, to estimate how a finished model will perform on unseen data. Leakage from the test set into training invalidates evaluation.

A model's ability to perform well on data it was not trained on. It is the central goal of machine learning and the reason for held-out evaluation and regularization.

The practice of creating, transforming, or selecting input variables to improve model performance, such as normalizing values or deriving ratios. It encodes domain knowledge into the data.

The process of assigning target answers to raw data so supervised models can learn from it. It is labor-intensive, sometimes outsourced, and a common source of error and bias.

A model composed of layers of connected units that transform inputs through weighted sums and nonlinear activations. Stacking layers lets it learn increasingly abstract patterns.

Machine learning using neural networks with many layers, which learn hierarchical representations directly from raw data. It underpins modern vision, speech, and language systems.

A nonlinear function applied to a neuron's weighted input, such as ReLU or sigmoid, enabling networks to model complex relationships. Without nonlinearity, stacked layers collapse to a single linear transform.

The algorithm that computes how much each network weight contributed to the output error, using the chain rule, so weights can be adjusted. It is the core of neural network training.

An optimization method that repeatedly adjusts parameters in the direction that reduces error, following the negative gradient. Learning rate controls the size of each step.

One complete pass through the entire training dataset during model training. Multiple epochs are usually needed, but too many can cause overfitting.

A setting chosen before training that controls how learning proceeds, such as learning rate, batch size, or number of layers. It is tuned via validation, not learned from data directly.

A neural architecture built on self-attention that weighs relationships between all tokens in a sequence in parallel. It is the foundation of modern large language models and many multimodal systems.

A technique that lets a model weight which parts of its input are most relevant to each output position. Self-attention within transformers captures long-range dependencies in text and other sequences.

The field of enabling computers to process and generate human language, covering tasks like translation, summarization, sentiment analysis, and question answering. Modern NLP is dominated by deep learning and transformers.

A unit of text, often a word piece or subword, that a language model processes as input or output. Token counts drive context limits and API pricing.

A neural network trained on large text corpora to predict and generate language, enabling summarization, coding, and dialogue. LLMs can produce fluent but false content and require safeguards.

When a generative model produces confident output that is fabricated or unsupported by fact. It arises from statistical generation rather than knowledge lookup and requires verification or grounding.

The instruction or input text given to a generative model to steer its output. Prompt design affects quality, safety, and consistency of responses.

An approach where a model retrieves relevant documents from an external store and includes them in the prompt before generating an answer. It improves factual grounding and lets knowledge be updated without retraining.

The maximum amount of text, measured in tokens, a model can consider at once when generating a response. Exceeding it forces truncation or summarization of earlier content.

Systematic error or unfair skew in data, models, or outputs that disadvantages groups or distorts predictions. It can enter through sampling, labeling, historical inequity, or design choices.

The principle that AI systems should not produce unjustified disparate outcomes across groups. It is hard to formalize, and competing mathematical definitions often conflict in practice.

The degree to which an AI system's purpose, data, logic, and limitations are disclosed and understandable to stakeholders. It supports scrutiny, trust, and informed use.

The ability to describe how and why a model produced a given output in terms humans can follow. It matters most in high-stakes decisions like credit, hiring, and healthcare.

The assignment of clear responsibility for an AI system's design, deployment, and harms to identifiable people or organizations. It requires oversight, documentation, and remedy mechanisms.

The policies, roles, and controls an organization uses to direct and monitor AI development and use. It covers risk assessment, approval gates, monitoring, and incident response.

The ongoing process of identifying, assessing, mitigating, and monitoring AI risks such as harm, bias, security, and legal exposure. It is proportionate to use case and impact.

A design where a person reviews, approves, or corrects AI outputs before consequential action is taken. It adds oversight and accountability, especially in high-stakes settings.

A short document describing a model's intended use, training data, evaluation results, and limitations. It helps users judge fitness and supports transparency and governance.

The process of making a trained model available in production so applications can send it inputs and receive predictions. It includes serving, scaling, monitoring, and versioning.

Running a trained model on new inputs to produce predictions or generations. It is the production-time counterpart to training and is often optimized for latency and cost.

The degradation of a deployed model's performance as real-world data diverges from training data over time. It is detected through monitoring and addressed by retraining or redesign.

A defined interface that lets software systems request services from one another, such as sending a prompt to a hosted model. APIs enable integration of AI capabilities into applications.

Further training of a pretrained model on a smaller, task-specific dataset to adapt its behavior. It is more expensive than prompting but can improve specialized performance.