Recency Bias
A phenomenon in continual learning in which a model becomes overconfident about the most recently trained task, causing decision boundaries and confidence to skew toward recent tasks and thereby affecting the reliability of out-of-distribution detection.
Timeline (last 90 days)
In continual learning, standard confidence scores shift systematically as tasks progress: the model is overconfident on inputs from recent tasks and gives muddled confidence for old tasks and truly unknown inputs.
Recency bias refers to the model repeatedly encountering the latest task data at the end of training, causing the decision boundary and output confidence to skew toward recent tasks, rendering threshold-based OOD discrimination ineffective, and this cannot be corrected by simple temperature scaling or post-processing.
All Facts (2)
In continual learning, standard confidence scores shift systematically as tasks progress: the model is overconfident on inputs from recent tasks and gives muddled confidence for old tasks and truly unknown inputs.
50%UnverifiedRecency bias refers to the model repeatedly encountering the latest task data at the end of training, causing the decision boundary and output confidence to skew toward recent tasks, rendering threshold-based OOD discrimination ineffective, and this cannot be corrected by simple temperature scaling or post-processing.
50%