Module 4: Neural Network - seeking advice

[Jeff_RNL]

Hi,

I am currently taking Module 4: Advanced Predictive Modeling and looking for advice from anyone who has completed that course and the associated certification exam. Whereas the logistic regression section had a lot of explanations and justifications for why certain techniques were used, the neural network section primarily seems to be assessing how well I can duplicate the instructor's steps in Enterprise Miner with very little explanation. Additionally, the neural network course has no end-of-chapter quizzes to assess whether or not key points have been learned.

I'm looking for suggestions for how to supplement my neural network education in preparation for the exam.

Thanks!

[RobertBlanchard · Accepted solution]

Hey Jeff,

First, congratulations on your journey to becoming a SAS Certified Data Scientist.

I understand your frustration with the design of the course, in that certain actions are performed in some demonstrations before the related content is described - I sympathies. 

The exam content was created with the neural network course in mind.  Therefore, I would recommend focusing most of your time studying the course, as opposed to seeking supplemental material.  I suggest you rework through the course and make sure you understand why actions are being performed.  For example, in the Compositional Data demonstration (ch. 2), be sure you know why the softmax activation function was used in the output layer... and why the cross-entropy error function is used.  If you are unsure as to why these options are being chosen, then navigate to the activation function and error function sections - investigate and become comfortable with why those option values were chosen.

I would also highly recommend taking the practice exam after you have become comfortable with the course material.  

All of the above stated, a practitioner that has used neural networks for years, or a researcher that focuses on neural networks should be able to pass the exam without taking the course.  If you would like to dive deeper into subject areas as preparation for the exam, I would recommend the reference section in the appendix of the course.  I'm directing your attention to the reference section as opposed to the broader internet because the course was built from the works of the referenced authors/researchers.  And again, the exam content was created with the neural network course in mind...  

A closing thought - I think you'll do well on the neural network section of the exam.  Some of my past SAS Academy for Data Science students expressed that they scored highest on the neural network section.  

I'm here to help if you have other questions.

All the best and good luck,

  Robert W. Blanchard

[Jeff_RNL]

Thanks so much for your response and advice.

I'm curious about something...the Neural Network module has a pretty good balance of models built with the Enterprise Miner Neural Network node as well as models built in a code node using PROC NEURAL. Which approach do you prefer in practice? 

Jeff

[RobertBlanchard]

Hey Jeff,

That's a great question...  If time is limited and I need a quick model, then I typically run several variations of neural network node and use a model comparison node to select a champion model.  If I have the time, then I prefer to use PROC NEURAL because I have greater control over how the model is designed.  For example, I can control the feedforward connections in the neural network, I can add additional hidden layers, train using a variable learning rate schedule, etc..

I'd recommend taking the time to learn to code using PROC NEURAL.  You'll likely have to tackle problems that do not conform or fit into a nice, neat box - You'll need the coding skills to create a customized solution.  Also, SAS Viya's PROC NNET is closely modeled after PROC NEURAL.  So learning one will make it very easy to learn the other.  Two birds as they say.. 

Best,

  Robert W. Blanchard