FlorianCassayre / DeepLearningUtil

A toy library written in Java to better understand Artificial Neural Networks.
2 stars 0 forks source link

Support creation of Complex Networks #3

Open FlorianCassayre opened 7 years ago

FlorianCassayre commented 7 years ago

Currently the library only allows the creation of basic feedforward networks (essentially a stack of layers stored as a list). But some tasks requires a more complex architecture with mixed connections between layers. This implementation would need deep internal modifications, namely the class Network itself needs to be redesigned and potentially Layer aswell.

hugo4715 commented 7 years ago

Maybe supporting reccurent network could be usefull

FlorianCassayre commented 7 years ago

Yes, indeed (although these fit in another very different category). I was initially thinking about "networks in a network" such as Inception, U-Net and so on. Inception U-Net

FlorianCassayre commented 7 years ago

Basically two layers are needed: SplitLayer and ConcatLayer But implementing these will violate the rule where each layer must be independent from the others. Perhaps it's fine and I'm just over-engineering the whole thing.