Why is there a decrease in the performance of pre-trained Deep Learning models?

Question

Doing binary classification with infected/uninfected RBCs (something the pretrained DL models have never seen before) using models and weights from Keras. I find the performance of the models (vgg16,19,xception) decrease with increase in the number of training and validation instances. Why?

Are you adding a new class to the final prediction? Output dimension increased? — Daniel Möller, May 04 '17 at 18:52

score 1 · Accepted Answer · answered May 05 '17 at 08:16

1

Maybe it is related to resource management where you are doing inference and the model expands in the memory and it can decrease the performance. This situation will create a lot of Main memory access to perform the forward pass computations and page faults are occurring and it can decrease the performance.

Hope this helps.

answered May 05 '17 at 08:16

Chitrang Talaviya

26
3

Thanks for that. The issue is that the I'm not using adequate number of epochs while fine-tuning the network to learn the new features. As I increase the number of epochs and do permutations with the SGD learning rate, I could find that the performance is increasing. – shiva May 05 '17 at 20:26

Why is there a decrease in the performance of pre-trained Deep Learning models?

1 Answers1