Feature extraction using CNN and classification with SVM

Hi,

I have a question on feature extraction from 2D CNN and classifying features with SVM. First let me introduce what I am trying to do;

1) I use pretrained network AlexNet which is trained with ImageNet.

2) I have a small dataset and use transfer learning for the classification problem. First, I trained my database with AlexNet by retraining all the parameters inside the network (no freezing layers) and observed the accuracy.

3) Now I want to classify the extracted features from the network with SVM. Should I use the initial AlexNet network's layer for feature extraction (default AlexNet) or the retrained network's layer on step 2?

I actually tried both of them and acquired higher accuracy on retrained network with almost %20 difference compared with the initial AlexNet. Is it just because the retrained parameters performed well on SVM? That's why a higher accuracy observed? Which one should be used?

Answer 1

[Raynier Suresh]

Hi, It usually depends on the problem and the approach would change for different problems, there is no single answer for it. Based on the data and the problem you are trying to solve choose the most suitable method.