Discovering and Predicting Patterns Using Neural Network Models

Presenter: Olivia Lippincott 

Neural Networks Overview and Case Studies

See how to:

  • Understand a neural network as a function of a set of derived inputs, called hidden nodes, that are nonlinear functions of the original inputs
  • Interpret Neural Network diagram inputs (factors) and outputs (responses)  8:30
  • Understand terms and how they apply to building Neural Networks (nodes, activation type, activation functions)
  • Understand types of activation functions  (TanH, Linear, Gaussian, Sigmoid, Identity and Radial) that transform a linear combination of the X variables at the nodes of hidden layers
  • Follow sequence for building a Neural Network model 21:20
    • Launch model (JMP vs. JMP Pro)
    • Specify # nodes in first and second layer
    • Use boosting to build a larger additive model fit on the scaled residuals of previous smaller models)
    • Specify Learning Rate (closer to 1 will run model faster, but may tend to overfit)
    • Transform covariates
    • Use Robust Fit for continuous output to minimize impact of outliers
    • Understand how JMP uses penalty methods and number of tours
    • Select Penalty Method (Squared, Absolute, No Penalty) to help avoid overfitting
    • Select number of Tours
  • Case study building a neural network for a continuous response 33:44
    • Use JMP validation to avoid overfitting
    • Compare Training set and Validation set values
    • Use Model Launch to change Neural Net settings
    • Use Actual by Predicted Plot and Profiler to evaluate how predictors impact results
    • Save formulas, publish Prediction Formula for Model Comparison
    • Save Fast Formulas and examine them in data table
    • Save to Formula Depot for comparison with other model results
  • Case study building a neural network for categorical response 39:28
    • Interpret Fit Statistics, Confusion Matrix, ROC Curves and Lift Curves
    • Understand why Transformation and Robust Fit options don't apply to categorical responses

Note: Q&A included at times 18:59, 19:49, 20:51 and 44:10


Back to Top