AAI-M key terms

The field of building computer systems that perform tasks normally requiring human intelligence, such as reasoning, learning, perception, and language use. It is an umbrella covering many techniques, not a single method.

A branch of AI where systems learn patterns from data rather than following hand-written rules. Models improve at a task as they are exposed to more examples during training.

A machine learning approach using artificial neural networks with many stacked layers. It excels at tasks like image, speech, and text processing because deeper layers capture increasingly abstract features.

A computing structure of interconnected nodes arranged in layers that transform input data into outputs. Each connection carries a weight adjusted during training to reduce prediction error.

The labeled or unlabeled examples used to teach a model its patterns and parameters. Its quality, representativeness, and licensing heavily influence how well the resulting model behaves.

A held-out portion of data used during model development to tune settings and compare candidate models. It is kept separate from training data to reveal whether the model generalizes.

Data reserved and never used during training or tuning, used only for a final, unbiased estimate of model performance. Leakage from test data into training invalidates this estimate.

When a model memorizes training examples, including their noise, instead of learning general patterns. It performs well on training data but poorly on new, unseen data.

When a model is too simple or insufficiently trained to capture the underlying patterns in the data. It performs poorly on both training and new data.

A training approach where each example includes the correct answer, called a label. The model learns to map inputs to those labeled outputs, enabling prediction or classification on new data.

A training approach that finds structure in data without labeled answers. Common uses include clustering similar records and reducing dimensionality to reveal patterns.

A training approach where an agent learns by taking actions and receiving rewards or penalties. Over many trials it discovers strategies that maximize cumulative reward.

The known correct output attached to a training example, such as a category or numeric value. High-quality labels are essential for supervised learning to succeed.

An individual measurable input variable used by a model to make predictions, such as age, price, or word count. Feature selection and engineering strongly affect model accuracy.

The mathematical structure produced by training that maps inputs to outputs. It encodes learned patterns and is what gets deployed to make predictions or generate content.

The process of feeding data to an algorithm so it adjusts its internal parameters and improves at a task. It produces a model that can then be evaluated and deployed.

Using an already-trained model to produce predictions or outputs on new inputs. It happens in production and is typically far cheaper than training.

AI systems that create new content, such as text, images, audio, or code, rather than only classifying or predicting. They learn patterns from large datasets and generate novel outputs from prompts.

A neural network trained on vast text corpora to predict and generate language. It can answer questions, summarize, translate, and write, based on patterns learned during training.

A chunk of text, often a word or word fragment, that a language model processes as a unit. Token counts drive context limits, latency, and API cost.

The instructions or input text given to a generative AI model to elicit a response. Clear, specific prompts with context and examples generally yield better outputs.

The practice of designing and refining inputs to a generative model to get accurate, useful, and safe outputs. Techniques include role instructions, examples, and structured formatting.

When a generative model produces confident but false or fabricated information. It occurs because the model predicts plausible text rather than verifying facts, so outputs need human review.

A technique that retrieves relevant documents from a knowledge base and supplies them to a language model before it answers. This grounds responses in current, verifiable sources and reduces fabrication.

Anchoring a model's output in trusted, verifiable sources such as documents or databases. Grounded responses are more accurate and easier to audit than unconstrained generation.

The maximum amount of text, measured in tokens, that a model can consider at once when generating a response. Content beyond this limit is not seen by the model.

Further training of a pre-trained model on a smaller, domain-specific dataset to adapt its behavior. It can improve specialized performance but requires quality data and careful evaluation.

The degree to which data is accurate, complete, consistent, timely, and fit for its intended use. Poor data quality directly degrades model performance and business trust.

The policies, roles, and processes that manage data availability, usability, integrity, and security across an organization. It establishes who can access and use data, and how.

A record of where data comes from, how it moves, and how it is transformed over time. It supports traceability, debugging, and compliance audits of AI systems.

Systematic unfairness in model outputs that disadvantages certain groups, often arising from biased training data or flawed design. It can cause discriminatory outcomes in hiring, lending, and other decisions.

The principle that AI systems should treat people and groups equitably and avoid unjustified disparate impact. It is often measured and monitored across demographic groups.

The ability to describe how and why a model produced a given output in terms understandable to stakeholders. It supports accountability, debugging, and regulatory compliance, especially for high-stakes decisions.

Openness about how an AI system works, what data it uses, and its limitations. It helps users and auditors assess whether to trust and rely on the system.

The practice of designing, building, and deploying AI in line with ethical principles such as fairness, safety, accountability, and human oversight. It aims to prevent harm and build trust.

The structures, policies, and controls an organization uses to ensure AI systems are developed and used lawfully, ethically, and safely. It assigns accountability and sets risk-management processes.

The process of identifying, assessing, and mitigating risks that AI models produce wrong or harmful outputs. It includes validation, monitoring, and documented controls throughout the model lifecycle.

Ongoing tracking of a deployed model's performance, inputs, and outputs to detect degradation or drift. It triggers retraining or intervention when behavior falls outside acceptable limits.

A change over time in the statistical properties of production data compared with the data a model was trained on. It can silently degrade accuracy and is a key monitoring signal.

A design where people review, approve, or correct AI outputs before they are acted upon. It adds oversight for high-stakes or uncertain decisions and helps catch errors.

Data that can identify a specific individual, directly or in combination, such as names, addresses, or identifiers. Handling it requires privacy safeguards and legal compliance.

The protection of personal information from unauthorized access, use, or disclosure. AI systems must collect and process personal data lawfully and with appropriate consent and security.

A short document describing a model's purpose, training data, performance, limitations, and intended uses. It promotes transparency and helps users judge whether the model suits their needs.

A small, time-boxed pilot that tests whether an AI idea is technically feasible and worth pursuing. It is not production-ready but informs investment and design decisions.

The end-to-end stages of an AI initiative: problem framing, data preparation, model development, evaluation, deployment, monitoring, and retirement. Each stage has distinct deliverables and risks.

The practices and tooling for reliably building, deploying, and maintaining machine learning systems in production. It applies software engineering and automation to the model lifecycle.

The financial benefit gained relative to the cost of an AI initiative. It helps justify investment and compare competing use cases on measurable business value.

A specific business problem or opportunity where AI can deliver value, such as automating document review or forecasting demand. Use cases should be prioritized by feasibility and impact.