In my experience I used transfer learning paradigm on biomedical imaging domain. In most cases transfer learning can increase the error because either of a failed fine-tuning phase ( giving rise to overfitting). It strongly depends on parameters tuning (stride, output function of the overall architecture) or maybe because of a not balanced image samples with respect to the fine-tuning classes for the application domain.

Could you please tell me some further details?

I am sorry if my answer looks quite general, I give only my experience results with respect my application domain.

Let me hear from you,

Alessandro

Hey there,

I think transfer learning fails (increase the error) in two cases:

https://arxiv.org/pdf/1703.06857

http://www.mdpi.com/2072-4292/9/9/907/pdf

https://medium.com/nanonets/how-to-use-deep-learning-when-you-have-limited-data-part-2-data-augmentation-c26971dc8ced

http://ruder.io/transfer-learning/

https://towardsdatascience.com/understanding-learning-rates-and-how-it-improves-performance-in-deep-learning-d0d4059c1c10

Thank you guys.

For negative learning it decrease the performance. For example, between and source and target, you pick some thing as source which does not cover up all the criteria which belongs to target. In that case miss calculation or false prediction might come. Form confusion matrix you can observe those.

TIA