An empirical study of task and feature correlations in the reuse of pre-trained models

The empirical study published on November 13, 2025, investigates how pre-trained neural networks are reused in machine learning, particularly examining the correlation between tasks. It reveals that Bob's success in reusing Alice's model is largely attributed to the correlation between their tasks, as task accuracy increases with this correlation. Interestingly, even when tasks are uncorrelated, Bob can still achieve better-than-random performance due to the characteristics of Alice's network and optimizer. The study posits that when task correlation is low, it is advisable to reuse only the lower layers of the pre-trained model. Furthermore, it hypothesizes that the optimal number of retrained layers can indicate the level of task and feature correlation. This research underscores the significance of understanding task relationships in enhancing the effectiveness of model reuse in machine learning.