Episode 14: July 27, 2012
High School Science
by Lee Falin, PhD
Seventeen-year-old student Brittany Wenger recently made headlines by creating an artificial neural network that doctors can use to help diagnose whether or not a breast tumor is malignant. Brittany’s network is especially notable in that it out-performs several commercial efforts, a feat that helped her to win the grand prize in the Google Science Fair. But what exactly is a neural network, and how can one be used to diagnose breast cancer?
Sponsor: Squarespace.com. SquareSpace has launched a brand new content management system, making Squarespace faster and easier than ever to create a high quality site or blog. The New Squarespace features mobile responsive designs with automatic scaling to any size device, plus more than 50 new features. For a free trial and 10% off your first purchase on new accounts, go to squarespace.com/EVERYDAY
and use offer code EVERYDAY7
Real Neural Networks
Before we talk about artificial neural networks, let’s take a look at what they’re modeled after, real neural networks, like the one in the human brain.
A nerve cell, or neuron, has three main parts. The dendrites on one end are connected to the axon on the other end via a cell body. The dendrites receive signals from a variety of sources, including sensory organs such as your eyes, ears, and skin, as well as signals from other neurons.
Neurons are very complicated and aren’t completely understood yet, but the simple explanation is that when the neuron receives strong enough signals through its dendrites, it releases neurotransmitters from its axon. These neurotransmitters cross the space (called a synapse) between cells. The amount of neurotransmitters released varies depending on a few factors, including how often a particular synapse is used. Scientists believe that repeated learning strengthens particular synapses, making them stronger producers of neurotransmitters.
No Soup For You
An artificial neural network works very similarly. Imagine that one day you’re feeling nostalgic and you decide to try and recreate your mother’s homemade chicken noodle soup. You remember exactly which ingredients she used, but you don’t remember the amounts very well. You do know exactly what it is supposed to taste like though.
So you set to work, adding what you think is the proper amount of each ingredient. Then you taste the soup and realize that you put too much salt in and not enough chicken. So you start over and adjust the amounts of each ingredient. You keep repeating this process until you get the exact flavor of soup that you’re looking for.
While this is an extremely inefficient way to make soup, this is a very efficient way for training a neural network.
Neural Network Training
Neural networks consist of a set of inputs, some mathematics in the middle, and one or more outputs. If you tell the network what the inputs are and what output you expect, it can continually adjust the mathematics part in the middle until it always gets the right answer for a given set of inputs. This is called “training the network.”
Going back to our soup example, once you’ve figured out the recipe to your mom’s soup, you could see a pile of ingredients and know immediately whether or not those ingredients would result in your mother’s soup.
Similarly, once a neural network has been trained using a set of known inputs and outputs, it should be able to take new inputs and get the correct outputs, even if you don’t know the answer ahead of time.
Neural Networks and Breast Cancer
Budding scientist Brittany created a neural network where doctors could provide various characteristics of a biopsy as inputs. She then trained it by telling it a set of tumor characteristics that were known to belong to tumors that turned out to be malignant and those that turned out to be benign. The network then adjusted its mathematics until it could correctly predict the results of each test case.
The neural network can now use the mathematics it developed during its training phase to determine whether or not the characteristics of an unknown are likely to mean that the tumor is malignant or benign.
Now you not only understand the worst way in the world to make soup, but also how neural networks function and, more importantly, how they can be used by doctors to help diagnose major conditions like breast cancer.
If you liked today’s episode, you can become a fan of Everyday Einstein on Facebook or follow me on Twitter. If you have a question that you’d like to see on a future episode, send me an email at firstname.lastname@example.org
Neuron photo from Shutterstock