Transfer Learning shows no enhancements in runtime

When configuring my network for transfer learning, to "freeze" the layers I am setting the 'WeightLearnRateFactor' and 'BiasLearnRateFactor' to 0, which from my understanding is Matlab's intended implementation of transfer learning. However, given the same number of iterations and epochs, training my network takes the same amount of time as it would during standard training. Are there any additional steps that must be taken to implement transfer learning in a custom neural network.

[Catalytic]

I don't know if it is such a good idea to freeze the transfered layers. The doc page on transfer learning recommends conversely to increase the learn rate factors on the new layers.

Answer 1

[Matt J]

If you run the same number of iterations and epochs with the same data, then naturally it will take the same amount of time. Run fewer epochs or run with fewer training data samples.