The pre-main-post parallel branch networks structure on intermediate orthogonal omnitoken has vast potentials and possibilities, which I think many you guys may have not realized yet, so I’d like to provide a little hints I thought about below.
First, fixed-dynamic mixed networks.
The different parallel pre/post branch networks in this premainpost networks on intermediate omnitokens can be not only for different modalities but also for different fuctionalities of ONE modality or different modalities.
For example, you can combine fixed networks with dynamic networks after training by this premainpost networks on intermediate omnitoken. You can design some branches in pre/post networks and the main network as fixed after training to preserve the best outcome and accuracy from training, and at the same time you can make some other branches as dynamic to be able to be learning or training during working or inference, like applying liquid network concept on these dynamic learning branches.
And in fact, you can design as many different functionalities as you can imagine in this premainpost networks on intermediate omnitoken.
Second, iterated nesting networks.
This premainpost networks on omnitoken can be iterated nesting struture like a multiple layers neural network, in which each pre or post branch itself can contain its own submain network and subpre/subpost networks.
These are what I can think about for now, but In fact, the limit for this premainpost networks seem only the imagination of you!
Be First to Comment