I want to ask if my dataset is smaller like 1400 training examples and after using transfer learning my trainable parameters are 40k then also it is likely to overfit as number of parameters are greater than training examples .Please throw some light on this as i am confused in selecting how many fully connected layers on top of convolution base for model to work fine
Regarding Neural Networks
hey @amankharb ,
firstly , it is nothing as such .
You model can perform poor due to various reasons , might it be data leak problem , or you might using your model to dig very deeper into images which is leading to overfitting , learning rate , number of nodes , batch size , etc. anything can cause your model to overfit.
If your dataset is small , you can use augmentation to increase your dataset size and then train your model on it. And while using transfer learning , always first have a look at the first manually , to think whether it requires transfer learning or not. and with transfer learning , you need to pay more attention on learning rate also, it does has a big effect on it.
This is a task of experiments , you need to try various different layers combinations , to understand how the performance will be with this dataset . There is no such specific rule or explanation how to choose these feed forward layers for learning.
I hope this helps you.
Thank You and Happy Learning 
.
ok got it thanks for the help