Neural Network

[roberto]

[nickoppen]

Hi Roberto,

I'm very interested in your XOR solution. Could you post a reference please.

I've been out of uni for quite a while now and I've not kept up with developments in machine learning. Someone on reddit told me about asynchronous SGD, a technique introduced by Google in a 2012 NIPS. That sounds like a worthwhile development especially with lots of cores and small networks.

nick

[over9000]

[roberto]

[nickoppen]

Hi Roberto,

The Google paper is here: . As I understand it, they have shown that training the same network asynchronously one many cores each using part of the training set is as good or better than running the whole training set on the network in a linear of synchronous manner. There is a commentary here: .

Can you dig up the 1989 paper by the three smart guys? You've got me intrigued. Your XOR redefinition seems like it is adding non-orthogonal inputs to the network - something that I thought was a problem.

nick

[roberto]

[over9000]

[over9000]

[nickoppen]

Thanks Roberto.

That's going to take me a while to get through.

nick

PS the link is if anyone wants to have a read.

[roberto]