Hey everyone,
does anyone know if it is useful or even possible to train Full Dense Networks, i.e. networks without any convolutional layers using transferlearning?
Unfortunately I haven’t found any sources for this anywhere and I would like to know if it is worthwhile to put some work into it.
Yes,
i want to train my Network with a bunch of (at least similar) simulated Data and then retrain it with a few real measured data.
right know i work with self normalizing neural networks, without any convolutionals layers, full dense only. They work pretty well but i have to start initialized weights with zero mean and a standard deviation of the squared root of 1/(size of input).
So i wonder if it is possible to start with pretrained weights. Searching the web has made me believe that conv layer are a requirement for this, because on every site, they only replace/not freeze the full dense layers behind the conv. layers.
I am not stuck on SNNs, if other (not conv) networks are better suited for transferlearning, I would switch, because I see a big advantage for my application in transferlearning.