Transfer learning on a very small dataset

Hi, I have total 20 images from two datasets i.e. Train and Test.

I want to perform a pilot study on transfer learning using Alexnet architecture. Is this going to hurt me? Can I not perform this transfer learning on such a very small dataset?

[Kevin Chng]

It's subjective. Generally, we believe larger dataset would provide more information to your network to get higher accuracy. However, why don't you try out with your current dataset first? Subsequently, you may justify the performance, and think how to improve the accuracy.

https://www.mathworks.com/help/deeplearning/examples/transfer-learning-using-alexnet.html

[Nbillah]

Thank you so much, and I checked with my dataset and its very low accuracy i.e 66.67%.

[Image Analyst]

Then I guess you have your answer. Can you post your script? Maybe there's something that can be done. What is in the images that you want to find? Can you post 1 or 2 training images with the things identified?

[Nbillah]

Thank you. I have total 20 images, 10 each group for Normal and Glaucoma. I did ROI extraction from real images and then tried on transfer learing example.

Would you please tell me what should be the maxepoch and minibatchsize for my data?

[Image Analyst]

It seems that deep learning is not needed. Simply find the area. Glaucoma has small area while normal has large area. Would you consider abandoning Deep Learning to do a much simpler traditional method using regionprops() to find the area of the blobs in your binary image?

Or do you just want to use deep learning as a learning exercise even though it's probably not the best approach?

[Nbillah]

Thank you for your comment. Yes I did the calculation for area. Actually I want to train a machine with the shape and size of this extract ROI.

[Nbillah]

Hi, I used Alexnet pretrained CNN to train my images . I got the following graph, I used 120 binary images per two group. And I got the following result. Please have a look the following graph and comment about the result. The graph is good or is it overfit/underfit? Can I go further with this CNN to train same segmented images with increasing my data size?

[Shounak Mitra]

Nahida,

The training progress chart looks 'just right'. Is the 80% accuracy acceptable in your case? Sometimes even 99% is bad and other times even 70% is good! You can go ahead with more data. I don't see any problem. Good luck.

Shounak

[Nbillah]

Thank you so much Shounak. In my case 80% is not acceptable, but not bad. Yes I will try with more data.