8 terms
Showing all terms starting with H
When an AI model generates plausible-sounding but factually incorrect or fabricated information, often presenting it with false confidence.
Reinforcement Learning from Human Feedback - a training technique where human raters score model outputs to guide the model toward preferred behaviour.
Configuration settings set before training (like learning rate or batch size) that control how a model learns, distinct from learned parameters.
Techniques like RAG, grounding, temperature reduction, and self-consistency that reduce the rate of fabricated outputs in LLMs.
In multi-head attention, each head independently learns different aspects of token relationships, with results concatenated for richer representations.
Layers in a neural network between the input and output layers where learned representations and feature transformations occur.
An AI workflow design where humans review, validate, or correct model outputs at critical steps before actions are taken.
Systems that combine symbolic AI (rules, logic) with machine learning approaches to leverage the strengths of both paradigms.
