I am working on a hybrid CNN-SVM where I aim to use CNN for feature extraction and SVM for classification. However, I am confused as after reading related works, I found many approaches:

-Some use SVM instead of the softmax layer.

-Some use SVM after flatten layer.

-Others replace the activation function from softmax to linear and the loss function from cross entropy to L1-SVM.

So I wanted to know which method is the correct one among these approaches.

  • $\begingroup$ Hello, First two methods seems to be correct. The first 2 seems to be very similar as softmax layer is applied after flattened layer only. So no matter you place SVM after flatten/linear layer or instead of softmax layer it will work similarly. Not sure about third if you could provide some reference it will be helpful $\endgroup$ Jan 14, 2022 at 16:48
  • $\begingroup$ @Ashwiniku918 Actually, I have tried the first two methods and they don't give me the same results (accuracy). When using an SVM classifier instead of a softmax layer, I obtained better results than after the flatten layer. Considering the third appoache here is a reference: scielo.org.mx/… $\endgroup$
    – root
    Jan 16, 2022 at 13:01


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Browse other questions tagged or ask your own question.