Deep Learning Algorithm Global Minimum

40 Views Asked by At

I am struggling with a problem of calculating the global minima of Neural Networks in Natural Language Processing. The first method I used is to finde the global minimum based on convexity prperties. However, as you know that the categorical cross entropy loss function is nonconvex and I could not get further. So I am looking now for another method, any ideas? This would be very appreciated.

1

There are 1 best solutions below

0
On

Loss function for Neural Network is non-convex and hence, finding global minima is very tough. Infact, verifying if the minima reached is global minima is also tough. This is one of the major problem in deep learning. You can try the following methods and try your luck.

Try different initialisations.
Use accelerated optimization techniques such as Adam, Adagrad.
Using Autoencoder, RBMs for weight initialisation.