ANC Workshop - Lazar Valkov

Title: Surrogate-based search for transfer learning


Speaker: Lazar Valkov



In transfer learning with deep neural networks, people reuse pre-trained layers on a new task in order to achieve better performance. However, as the number of available pre-trained layers grows, it becomes difficult to efficiently select which ones to reuse in order to obtain the best performance on the new task. We approach this by evaluating multiple combinations of pre-trained layers and using a surrogate function to learn to predict an unseen combination's performance.

I will go through our preliminary results and outline the difficulties we have encountered.





Oct 13 2020 -

ANC Workshop - Lazar Valkov

Tuesday, 13th October 2020