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.
Core Facts
Timeline (last 90 days)
When facing rule instances that never co-occurred during training, the performance of symbolic descriptions falls below the no-information baseline, while the example approach remains above the baseline.
Negative transfer refers to when knowledge or priors a model learns on a source task actually harm performance when applied to a target task.
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 (5)
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.
75%UnverifiedWhen facing rule instances that never co-occurred during training, the performance of symbolic descriptions falls below the no-information baseline, while the example approach remains above the baseline.
50%UnverifiedNegative transfer refers to when knowledge or priors a model learns on a source task actually harm performance when applied to a target task.
50%UnverifiedCovariate 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%