Distribution Truncation Effect
A concept in statistics and machine learning referring to the progressive suppression of the tail of the training data distribution—rare but important samples—through repeated iterative processes, causing the model to retain only the most common patterns and resulting in a loss of diversity.
Timeline (last 90 days)
Distributional truncation refers to the phenomenon whereby, when training data comes from model outputs, rare samples in the distribution tails are further suppressed in each iteration due to their low frequency.
Distributional truncation was systematically documented in corpus research in linguistics and coarse-grained modeling in statistical physics long before the current context.
All Facts (2)
Distributional truncation refers to the phenomenon whereby, when training data comes from model outputs, rare samples in the distribution tails are further suppressed in each iteration due to their low frequency.
50%UnverifiedDistributional truncation was systematically documented in corpus research in linguistics and coarse-grained modeling in statistical physics long before the current context.
50%