Networks Beyond Modeling: Google Brain and Deep Learning
In this course we study networks which we can model mathematically and we learn the basic tools which allow us to reason about networks. Often this modeling leads us to draw some networks on paper. Limited by the size of the average piece of paper, this leads students to think that a majority of networks contain less than 20 nodes and a similar number of edges. This isn’t true: most exciting networks are exponentially larger than the ones we can draw. For example, Facebook has over 1 billion users, and we know that their friendships form an interesting social network (which Facebook experiments on regularly). The Internet itself is another beautiful network to think about. It contains around 2.4 billion pages, and with the links between them we have a complex network to study.
But are also some things in the physical world which may too large to effectively study as a network. A great example is the human brain. Sure, the brain is physically small, but it’s an extremely complex organ. The brain can be viewed as a large network of neurons which all need to send signals to each other. Can we model this network of neurons to help understand the brain? Probably not! There are around 100 billion neurons in the human brain. That’s two orders of magnitude bigger than the number of users on Facebook. If we think about the technical prowess required to keep Facebook’s social network running, we have no chance of accurately modeling a brain with computers. And this is one reason that the brain is still a huge mystery for scientists.
We may not be able to understand exactly how the brain operates as a network, but computer scientists have been successful in creating networks that mimic the brain. There are computational models known as artificial neural networks which use digital neurons to compute values from a series of inputs. These artificial neural networks then “learn” about their inputs over time and can solve problems that traditional computer programs find very difficult.
One company that has really embraced artificial neural networks is Google. They’ve used them to produce an internal tool called Google Brain. Google Brain uses a technique called deep learning and is being used to solve a large number of machine learning problems at Google. Google Brain is being used now in Google Maps to identify street addresses, in Android to improve voice recognition, and by another 30+ teams at Google. It’s effective because it attempts to learn in the same way that a human would…and humans are fantastic learners! So while modeling the human brain may be “beyond the scope of this class,” huge advances have been made possible by studying it.
Sources
Pages in Internet: http://www.worldwidewebsize.com/
Users on Facebook: http://thenextweb.com/facebook/2014/01/29/facebook-passes-1-23-billion-monthly-active-users-945-million-mobile-users-757-million-daily-users/
Neurons in Brain: https://faculty.washington.edu/chudler/what.html
Google Brain: http://www.wired.com/2014/07/google_brain/