fixed-dynamic mixed networks and branch networks model

Last updated on January 29, 2026

The model based on intermediate omintoken can be called as branch networks model, in which each of the networks before or after the main network is a branch network, and each of all branch networks is a neural network and the main network is also a neural network.

The fixed-dynamic mixed networks is based on this branch networks model.

The fixed-dynamic mixed networks has the some branches in pre/post networks and the main network as fixed (weights) after training to preserve the best outcome and accuracy from training, and at the same time you can make some other branches as dynamic (weights) to be able to be learning or training during working or inference.

And also, in the first layer or first hidden layer of the main network, you can make the weights of the columns dedicated to the dynamic branches are dynamic too. And thus the weights in the neural networks of dynamic branches and the weights in the columns of first layer of the main network dedicated to the dynamic braches are dynamic which can be trained or changable after training or in inference, and other weights of the mian network and the weights in the neural networks of fixed branches are fixed or unchangable after training or in inference .

A fixed-dynamic mixed networks is a branch networks model in which there are multiple branch networks before the main network, wherein a branch network the weights of which are fixed or unchangable after training or in inference is a fixed branch, and wherein a branch network the weights of which are dynamic or changable after training or in inference is a dynamic branch.