Are there neural network algorithms with a variable number of incoming channels?

Question

Programming

In a nutshell, there are certain types of input channels, the number of inputs in each type of channel should change without rebuilding the network or with minimal rebuilding (no retraining). There must be a certain number of output channels. Training is not a problem. How many read did not find whether there are such algorithms, everywhere the network configuration is configured when the network is created and cannot be changed. Maybe someone knows a similar algorithm?

Answer 1

As for any official solution - I unfortunately do not know if there is such a thing.
However, I have one question:
Why do you have a different number of input channels?
If these are logically different entities , then obviously it is worth making separate models for them, and then somehow working with their results.
If these are logically identical entities, then I would also consider the absence of any information as information. Perhaps you can try to fill them with default values ​​that are different from all possible ones, when the information is available, and train the model in the usual way.

Answer 2

See RNN or CNN, the choice depends on the nature of the data.