There is an option to keep specific layers' learning rates low so you can fix them as they are.
for example
fullyConnectedLayer(<outputsize>, 'WeightLearnRateFactor', 0, 'BiasLearnRateFactor', 0)
This way, you would multiply zero to the global learning rate, which is set via trainingOptions function, and thus the learning rates of the weights in the fully-connected-layer are set as zero.
