Negative Transfer
A phenomenon in transfer learning where pretrained weights not only fail to benefit target-domain learning but actively interfere with it, degrading training performance on the target task. It is one of the most common underlying causes of failure when fine-tuning models across different domains.
Timeline (last 90 days)
Covariate shift and model drift can trigger negative transfer, causing additional data to drag down model performance.
When two tasks exhibit a high degree of overlap in critical components, fine-tuning on one task may actually lead to performance degradation on the other.
Multi-task learning faces the risk of negative transfer, where joint training can harm the performance of individual tasks when the gradient directions between tasks conflict.
All Facts (3)
Covariate shift and model drift can trigger negative transfer, causing additional data to drag down model performance.
50%UnverifiedWhen two tasks exhibit a high degree of overlap in critical components, fine-tuning on one task may actually lead to performance degradation on the other.
50%UnverifiedMulti-task learning faces the risk of negative transfer, where joint training can harm the performance of individual tasks when the gradient directions between tasks conflict.
50%