Understanding Business Intelligence with Neural Networks James McCaffrey Microsoft Research Labs Wednesday, May 4, :15 – 4:00 PM Room Breakers CD
Agenda and Goals What is a neural network? How do neural networks work? How can you create a neural network? Q & A
training data independent variables / predictors / signals / attributes / features / X-values “the thing to classify (predict)” / label / dependent variable / Y What is a Neural Network?
How Do Neural Networks Work? 35 49,000 M Republican 0.62 inputhiddenoutput High weights and biases
How Do Neural Networks Work? 35 49,000 M Republican 0.62 inputhiddenoutput High encoding normalization weights and biases
Under the Hood ?? 1). (0.1)(4.0) + (0.2)(-5.0) + (0.3)(6.0) = 1.2 2) = 3.2 3). Activation(3.2) = ). ?? = Local output = perceptron weights and biases activation function
Activation Functions and Why You Don’t Care Logistic Sigmoid Output between [0, 1] y = 1.0 / (1.0 + e –x ) Hyperbolic Tangent Output between [-1, +1] y = tanh(x) = (e x – e -x ) / (e x + e -x ) Softmax Outputs between [0, 1] and sum to 1.0 y = (e -xi ) / Σ ( e -xj ) activation function
Training and Free Parameters – No Free Lunch Number of weights and bias values to determine: (n i * n h ) + (n h * n o ) + (n h + n o ) Ex: n i = 10, n h = 20, n o = 3 (10 * 20) + (20 * 3) + (20 + 3) = 283 back-propagation free parameters
Research vs. Reality back-propagation free parameters
Over-Fitting: The Biggest Challenge over-fitting error, accuracy
Four Ways to Actually Create a Neural Network Use an existing application tool (Weka, Azure ML) Hire a vendor company Use a library/API to create a custom system Create custom system from scratch
Weka
Azure Machine Learning
Neural Network from Scratch
Summary If your data can be put into a spreadsheet, then a neural network can be used to predict any of the columns. You don’t need to know how to implement a neural network, but you do need to know the 10 key vocabulary terms so you can communicate. The 10 Key Terms: perceptron, features, encoding, normalization, weights & biases, activation function, back-propagation, free parameters, over-fitting, error & accuracy. Neural network tools for Big Data are not quite ready for prime time.
Resources WEKA Step-by-Step: Technical Bible: ftp://ftp.sas.com/pub/neural/FAQ.html Python Implementation:
Thank You! Understanding Business Intelligence with Neural Networks James McCaffrey Microsoft Research Labs Wednesday, May 4, :15 – 4:00 PM Room Breakers CD