Mapping a vector to another using Caffe and Deep Learning

[Seyed Hassan Tabatabaei]

Hi Friends,

I want to map a 512*1 float vector to another 512*1 vector, I have create a fairly simple net in caffe with one or two fully connected layers. It result in a average 0.005 error per element but i wanna a more precise result (about 0.0001 error per element). also using a convolution layer before fully connected layers doesn't change the result.

what is your proposed net structure? does using more convolution layers cause in better result? if so how?

thanks.

[Przemek D]

Convolution is a contextual operation and as such is only useful if order is important in your data. If it's 512 sensors of various kinds, you probably don't need convolution. But if it's a audio signal represented by 512 samples - you'll find it useful.
As to fully connected layers, choosing their sizes is a tradeoff between underfitting and overfitting: you don't want your network to "memorize" data, but generalize it. More neurons = better capacity, greater chance of overfitting. Less neurons = less overfitting, possibly worse classification (insufficient capacity to accurately represent data). There's no golden rule though.