Supporting Jacobian calculation in MATLAB

AI & Data Science Deep Learning (Training & Inference)
1

In MATLAB (2013b/2014a) I am making use of the Neural Network Toolbox. There there is provision for making use of a GPU to train and simulate a neural network.

Looking through the help (see link supplied), however, there appears to be a caveat, associated with using the levenberg marquardt training method (denoted trainlm within MATLAB). I have put the caveat in quotes as extracted from the help file:

"If the network has the default training function trainlm, you see a warning that GPU calculations do not support Jacobian training, only gradient training. "

I would like to persist with trainlm and so this leads me to ask, can GPU calculations be made to support
Jacobian training?

2

https://www.astro.rug.nl/~yatawatta/sagecal.pdf is one resource which I am scanning through

3

Hi. I wondered whether you were able to progress in this endeavor. I find myself in a similar situation : trainlm is the best training algorithm for my application, but I am frustrated by the long training times.

4

Hi! Have you solved the issue, or it is still impossible to use GPU for training LM or BR?