Quick help on JSL/c++ for client

Question

I need some quick help.  Traveling right now, call scheduled with client at 3:30 ET on 6/16.

Client asked for help with the following:

Working with neural nets, considering JMP Pro.  Would like to understand technical details behind test the model with new set of data using c++ code. I would like to understand if/how model can learn from new data as it arrives. Is model update is manual or automatic?

Rather cryptic question (at least to me), but I would appreciate any advice that anyone could provide in the short term.

Thanks in advance?

Nate_Riordan · Answer

For what its worth I see this kind of disconnect regularly where folks have an expectation that the generated code is going to create a dynamic or live model in contrast to a fixed model.

An analogy I like to think of is the facial recognition algorithms or any of these huge speech recognition algorithms are generated based on a massive data set and hardware for that matter.  The model is in principle built well with the test train validate mantra.  And finally at the end that robust model is shipped to the puny processors found in our handheld devices providing amazing results.  There is a lot of literature out there to this effect yet for some reason end users think a new model is being trained on the fly which is simply not true.

Hope that offers insight.

DonMcCormack · Answer

Jerry,

To add to what Jeff and Dan said, updating the model with new data isn't what the test set is intended for and runs the risk of overfitting. It may be used to determine if the model fit is inadequate (in which case it needs to be refit, as per Jeff and Dan) but shouldn't be used to update the model.

Don

Daniel_Valente · Answer

As Jeff has mentioned. JMP Pro includes the ability by way of the Formula Depot to output score code in the C (not C++) language. The key use-case here would be to bring in data into JMP Pro, build a model using a NN, publish the model to the Formula Depot, output the score code to C and then compile the C to a .dll to score new data outside of JMP using the same model.

You may want to share this Discovery paper that I co-authored last year:

https://community.jmp.com/t5/Discovery-Summit-2016/Scoring-Outside-the-Box/ta-p/22381