This may be old news, but Brittany Wenger–the 17-year-old who snagged the top prize at this year’s Google Science Fair–is a boss. She built an artificial neural network that can detect breast cancer to 99.1% accuracy over 7.6 million trials. That’s 4.97% better than commercial networks for detecting breast cancer, and it’s less invasive. She’s one of those girls that makes me look in the mirror and say: “what are you doing with your life?”
Anyway. Artificial neural networks are touted as programs that think like brains, except better. What does that mean? It means they are capable of processing and interpreting huge amounts of data and finding patterns that are too complex for us mere mortals to find. This is sort of what the brain does (and I’m reducing big time here): it receives sensory input and processes that data, identifying and categorizing and pattern-recognizing all over those sensory inputs to make sense of them. In other words: it decides what to tell you to do based on the information it pulls out of the data. Wenger “taught the computer how to diagnose breast cancer,” meaning she built it to take an input and give a “right answer”–namely, benign or malignant tumor–as output.
This is what’s called a classification problem: to produce a discrete output–for example 0 or 1, yes or no, or benign or malignant–from the input. The program is first fed a huge data set with the “correct answers.” The program analyzes the data set, assesses the features, and fits a line to the data to use as a guide for future predictions. Then it is fed a new data set and uses what it learned from the first data set to make its own predictions. By receiving feedback on when it gets the answer right or wrong, the program’s accuracy increases with each prediction. This is classic supervised machine learning.
To achieve the accuracy that Wenger achieved requires really stellar programming.
I do need to say one thing: artificial neural networks do not artificial brains make. They are a major step in the development of AI, as are many advances in machine learning, but we still need to know more about what the brain does. For instance, how do we write an algorithm that will draw a line of equal finesse and flexibility to that of the brain? We’re not there yet but we’re getting there!