Neural network and FDE

Question

Hello,

we've been in touch with a customer who will evaluate JMP Pro for Neural network and FDE. He had very precised questions such as

1) Neural network: 1a) how JMP Pro compare to Keras and Tensorflow. 1b) can we change connections between nodes / layers 1c) is bootstrap forest available (in addition of boosting) 1d) can we quantify uncertainties.

2) FDE: 2a) is there a spline regression, 2b) how works dimension reduction in FDE, 2c) FPCA Clustering

As a comment he is mathematician ...

Thank you for the help.

Emmanuel

Phil_Kay · Accepted Answer

Hi Emmanuel,

Some of these are big questions. I am not an expert on any of this. But here goes...

1a. I don't know these tools well but I think you should not compare them to JMP Pro. Keras and Tensorflow are specialised tools for NNs. NNs in JMP is just one platform in a general statistical discovery tool for scientists and engineers.

1b. Do they want to have the ability to edit individual connections between nodes? I wonder why they want to do that. No, I do not think that would be possible. If they want to define a specific "path" for how predictors should be connected, that sounds more like SEM.

1c. I'm not sure what they mean. A bootstrap forest of NNs? You could save out different NN models and then aggregate them as predictors using BF. I think you need to ask for more detail. Why do they want to do this?

1d. You can bootstrap point predictions from a NN to generate confidence intervals around that prediction but I don't think there is a good way to construct overall uncertainty intervals.

2a. I am not clear what they are asking. Spline fitting is part of the FDE process. You need to ask for more detail about what they want to do with spline regression.

2b. How long do you have :)  ? I would search for one of the presentations from Chris Gotwalt and/or Ryan Parker for the statistical details about dimension reduction.

2c. You can save out the FPCs and use the clustering platforms. I assume they want to find clusters of similar functions. Yes, no problem. They can do this.

This sounds to me like someone that is trying to find excuses for why JMP Pro will not work for them and they will have to stay with the existing toolset that they like. I could be wrong but I would be careful about investing too much time on this.

MikeD_Anderson · Accepted Answer

I hate to be the one to say this, but they need to just RTFM (read the freaking manual...). A lot of the questions are addressed in the stats sections of the docs.  Chris G. also did a white paper (available in the community) on what our NN does.  It might be a little out of date, but I think the fundamentals are still sound.

So, I'd just say, "have a look at the documentation, it's online and fully searchable. There's a great white paper by our Chief Data Scientist on the community - here's the link.  Have a look at those and let me know if you have any questions they don't answer."

Best,

M

Jason_Wiggins · Accepted Answer

Keras is a high level Python wrapper for TensorFlow. There are deep learning NN capabilities in TensorFlow  that JMP does not have but I believe there is overlap. My exposure to TensorFlow  was during a project where we were using a pretrained convolutional neural net (CNN) to automatically identify defects in manufacturing parts from images. Even if JMP had CNN capabilities I don't think we would have been able to get results quick enough for real time monitoring that we were getting in with TensorFlow. I personally don't believe we need to compete at this level. I have interacted with a few data scientists recently who are using Keras for NN models they believe could be generated in JMP. They like how quick it is to generate a model in JMP vs scripting in python. They also like the idea that people who don't have data science and scripting backgrounds can participate in modeling efforts. I believe this is where we can compete and win.