AI Glossary

Techniques that recover a 3D shape or scene from 2D images or video, often using multi‑view geometry, NeRFs, or depth estimation.

Controlled experiments comparing variants to measure impact on quality, conversion, or engagement.

Mathematical functions that determine whether a neuron should be activated (fire) based on its input. Common examples: ReLU, Sigmoid, Tanh, GELU.

A training strategy where the model identifies which unlabeled examples would be most valuable to label, reducing annotation costs by focusing human effort on informative samples.

Adaptive learning rate optimizer that refines AdaGrad by limiting aggressive decay via running averages.

Optimizer that adapts learning rates per parameter based on historical squared gradients, aiding sparse features.

Adaptive moment estimation optimizer combining momentum and RMSProp‑like variance scaling; widely used in deep learning.

A widely used optimizer for training transformers. AdamW decouples weight decay from gradient-based updates, often improving generalization.

Lightweight trainable layers inserted into a frozen model to adapt it to new tasks without updating all parameters.

Regression metric adjusting R² for the number of predictors; penalizes overfitting and aids model comparison.

Intentionally crafted inputs designed to fool AI models into making mistakes or producing unintended outputs. Examples include adding imperceptible noise to images to cause misclassification, or prompt injections in LLMs. Important for testing model robustness.

Intentional manipulations of AI model inputs to cause incorrect outputs. These attacks exploit model vulnerabilities and are critical for AI security research.

Evaluating models with stress tests and adversarial prompts/inputs to find failure modes (safety, jailbreaks, hallucinations, tool misuse) before production.

Carefully crafted inputs designed to fool AI models into making wrong predictions. A single pixel change can cause a model to misclassify images.

Learning actionable possibilities of objects or environments (what actions are feasible), often for robotics.

An autonomous system that can perceive its environment, process information, and take actions to achieve specific goals. In the context of LLMs, an agent can use tools (like web search or APIs) to gather information and perform tasks.

Simulation approach where individual agents with simple rules interact to produce emergent system behavior.

AI systems that can autonomously plan, make decisions, and take actions to achieve goals, often using multiple tools and iterating based on feedback. Unlike simple chatbots, agentic AI can break down complex tasks, use external tools (APIs, databases), and adapt its strategy. Examples include autonomous research assistants and coding agents.

AI system design where models autonomously plan, execute multi-step tasks, use tools, and iterate based on feedback. Contrasts with simple request-response patterns. Requires careful safety controls.

A White House blueprint outlining principles to protect the public in automated systems: safe/effective systems, protection against discrimination, data privacy, notice/explanation, and human alternatives.