freeze some weights inside of a fullconnected layer

Is it posssible to freeze only apart of the weights within a fully connected layer?

That means some number of weights shall be learnable and the rest number of weight inside of this layer shall be not learnable.

Answer 1

[Matt J]

It is possible with a custom layer. See this recent, related thread,

https://www.mathworks.com/matlabcentral/answers/2174601-how-to-freeze-specific-neuron-weights-in-a-feedforward-network-during-training-process?s_tid=srchtitle

[Armin]

Thanks a lot