Improving neural network

User 2334 | 10/26/2015, 4:59:30 PM


I am new to deep learning and wondering how to improve the accuracy of my neural network. I have a set of labeled images form cells which I want to classify in 3 groups (4500 images for each group).

I have tried to use a pre-trained network (imagenet) which gives me about 0,81 accuracy. Increasing the amount of images does not really increase my accuracy, I get approximately the same number if I use 300 or 3000 images.

However, if I use my own neural network the accuracy is slowly increasing (300 images about 0,63 -> 3000 images 0,73) but still not as good as for the pre-trained model.

Any hints would be very appreciated, Thanks


User 940 | 10/26/2015, 9:56:46 PM

Hi @ete,

In general, pre-trained features are really, really good.

However, if you want to train your network, in general you need quite a lot of training data to prevent overfitting.

What's your training error if you use your own neural network, and what is the network architecture you are using?

Cheers! -Piotr

User 2334 | 10/27/2015, 8:35:17 AM

Hi piotr!

pre-trained features gives me the best results so far but I can not improve them (perhaps because my images are to different from the imagenet images!?). 0.81 accuracy is not bad but still not good enough.

For my own neural network I used graphlab.neuralnet_classifier.create with the default settings (which means no architecture, network=None) and 20 iterations. With 4500 images per category (13500 total) I am getting an accuracy of 0.71 which slowly increases with the amount of input data.


User 2334 | 10/28/2015, 2:56:22 PM

Hi @piotr,

I was wondering whether image augmentation might help!? I realized that I have a small number of very light and very dark images. I was also thinking to try different images re-sizing (my original images all of different size) and changing the light intensity and gamma correction might bring something.


User 940 | 11/2/2015, 5:23:50 PM

Hi @ete,

13.5k images isn't very much to train from scratch! If you're seeing error go down with input data, I'd recommend doing quite a lot of this image augmentation!

Good luck! Cheers -Piotr

User 2334 | 11/3/2015, 10:27:59 AM

Thanks Piotr! I'm trying augmentation. My images are not very complex and I hope I will be lucky!