Out Of Distribution (OOD) behavior in the context of AI pertains to situations where a model or system is faced with input data that significantly diverges from the data the model was trained on. Deep learning models, in particular, are prone to OOD behavior, as they are often incapable of extrapolating their learning to new, unseen data types. This is a significant issue, mainly when the AI model is used in real-world scenarios that often present unpredictable and diverse data not accounted for in the model's training phase. As a result, the model might produce unreliable outputs, severely impacting its performance and applicability.
The term came into use with the advent of machine learning and deep learning models in the 1990s and 2000s. OOD behavior became a focus area of study as these models were increasingly deployed in diverse and unpredictable environments, leading to a recognition of the challenges OOD data pose to model reliability.
While no specific individuals or groups are singularly credited with defining this issue, the broader AI and ML research community have collectively contributed to the ongoing study of OOD behavior in AI models. Researchers focusing on improving AI robustness and generalizability often address the challenge of OOD behavior.