I'm using a 2 layered neural network with 7 neurons for hidden layer to predict diabetes. But according to plot that Matlab shows me, my network's error first increases and then doesn't change, while as I know it should first decreases and then stops changing. Is my case possible or my network is not working correctly?
Error plot is attached bellow:
The back propagation uses "gradient descent", it is basically the RMSE of the desired vs actual (target you give ) .
Dear Mohammad Mahdi,
Because your network does not converge, meaning that its architecture is not proper for your data.
First check to see if your input and output layers are correct.
From architecture standpoint, things to look into is to increase the number of hidden neurons, change transfer functions, add additional hidden layers, among others.
Redesigning input layer is also another way to look at.
If your data is too sophisticated, which I highly doubt that, then try to use deep learnings and your classification would be accomplished.
Best regards,
J.
The back propagation uses "gradient descent", it is basically the RMSE of the desired vs actual (target you give ) .
Dear Mohammad Mahdi,
Well, initial random weights become important when dataset (inputs) is small or when the network has few layers with few neurons (e.g., in your case).
8% error is rather high. Another way is to feed and train the network with a larger input.
Another subject is that sometimes data are too scattered and dataset is not sufficient for small-sized networks to be trained, yours may fall in this category.
Best regards,
J.
J. Rafiee I Really appreciate your dedication. Your answers are really helpful for me.
Please see this link. Its in more detail
https://blog.slavv.com/37-reasons-why-your-neural-network-is-not-working-4020854bd607
Seyyed Mohammad Mahdi Bahadornia
- Badges
- Science topic