by Dr. Steve Poulin

Last week we discussed what makes artificial intelligence…artificial intelligence (AI).  Over the past several years AI has grown increasingly popular, helped along by the mystique of its name.  AI is one technique of many in the world of predictive analytics, though.  While last week we talked about what artificial intelligence is, we didn’t go into as much detail about how it works.

In our previous post, we mentioned that some predictive analytics techniques have histories spanning back to the 19th century.  As it turns out, AI has a long history as well, tracing its roots back to the 1940s, though renewed interest in the field came in the early 1980s.  As the name implies, the use of artificial intelligence for predictive analytics began as an attempt to mimic the way our brains acquire knowledge with computers.  This process is often described as machine learning.

Given that learning occurs when neurons in our brains acquire and share information, developers used the phrase neural networks to describe the way their computer algorithms were learning how to predict an outcome.  At a high level, we can break down this process into a few simple steps:

1)    Determine your inputs:
Before anything else, we must determine the information that should be selected as inputs to the neural network.  Broadly in predictive analytics we would call these the predictor fields.  In the context of a neural network, though, we would call these the neurons.

2)    Assign random weights:
The neural network process begins by assigning random weights to the neurons.

3)    Train the data:
These weights are adjusted through a training phase based on their ability to correctly predict the known outcome for each test subject.  The weights are changed frequently in a hidden layer with the goal of improving the accuracy of the predictions.  This process of adjustment is known as back propagation, and continues iteratively until the predictive accuracy of the weights stops improving sufficiently.

4)    Utilize the model:
Like the coefficients from a regression analysis, the final weights can be used to make predictions about new "test" subjects when the same information is available for them as it was for the training data.  In addition, we could use the model to rank the influence of each predictor.

This iterative process is what allows us identify complex patterns in data.  While artificial intelligence is a powerful tool, we must understand its strengths and weaknesses.  Next week we’ll dissect AI, taking a deeper look into its pros and cons.

AuthorRyan Harrington