Is Xavier init really better than Gaussian init?

[Andy Wong]

I found that sometimes using Xavier init may result in the objective not converging, while using Gaussian init could get good performance. Is Xavier init really better than Gaussian init? Thank you. 

[Sergio Guadarrama]

Good initialization depends on the problem as well on the network architecture. So there is not a single best initialization method.

You can also try MSRAFiller filler see [He, Zhang, Ren and Sun 2015]

[Ihsan Ullah]

As Sergio said it depends on the problem. One can not specifically says. How ever good initialization does effect convergence. 

Are their only these two ways of initialization of weights? Or there exist other techniques already implemented in caffe that one can use?
regards

ihsan

On Thursday, July 2, 2015 at 9:07:34 AM UTC-7, Andy Wong wrote: