Dear Ajit Sinam ,

To training from the stracth is computationally expensive and required many data (w.r.t network architecture). On the other hand, transfer learning is computationally efficient and provides good results.

In general, I think you need to check this empirically. Since transfer learning is low-cost, I recommend you starts by testing this approach. In this work Conference Paper Face Verification: Strategies for Employing Deep Models

I think learning from scratch is better because it is more efficient in education and with accuracy

Dear @Luaay Abdulwahid Shihab can u provide me some pros and cons for transfer learning.

Dear Ajit Sinam ,

Here you can find some guidelines that need to be contextualized in the specific use case. Generally speaking, given a task A and a task B, Transfer Learning (from A -> B) makes sense if three conditions are satisfied:

In case these three conditions are satisfied, you can decide to

• just fine tuning the network meaning that you take the pre-trained network of task A, just replace the last layers and retrain the network using the samples available for task B assuming immutable the weights of the pre-trained network of task A; basically, you are just training the last layers on the new samples of task B; this practice is typically recommended if you have not many (
  1 votes 1 thanks

I suggest to use recurrent networks using long short-term memory layers, or gated recurrent layers.

I can help you if you want.

Wolfgang

Dear Wolfgang Hitzl , please provide me the necessary details

Thank you btw.

You can only use transfer learning when there is a relation between your dataset and the pretrained model. For example you can use different pretrained model over ImageNet dataset with Cifar10,100 datasets, but you can not use the pretrained model of ImageNet with biomedical images because of ImageNet not contains images belongs to biomedical field, so in such case you should trains the model from scratch over biomedical images and then you can uses this model for transfer learning.