Neural Networks
In class, we discussed a simple Pagerank algorithm. The intent of that iterative algorithm is to converge on some set of ‘page ranks’ by propagating node’s values to each other. Neural networks are in some ways the inverse of that method – instead of converging on some value of interest, often Artificial Neural Networks (ANNs) are designed to converge on a specific output using very similar techniques. In fact, ANNs can be designed to produce different outputs for different inputs. That process can be extremely powerful, capable of voice recognition, facial recognition, and controlling complex robots.
A simple example is described in a UVM class’ assignment: given three random numbers between 1 and 0 as inputs, recombine them to an arbitrary sequence, for example: 1 0 1
A neural network which accomplishes that task might look like this:
Each red connection has weight 1 while blue connections have weight 0. Each nodes value is the sum of the nodes that point to it times the arrow’s weights.
This is an extremely basic neural network. More complex ones might take pixel colors as their inputs, and output a code identify what object is featured in the image, or who is in the image. Voice recognition is also often implemented with neural networks, where sound files are the input and words are the output.
Neural networks are extremely powerful, as evidenced by the adoption in many CS fields as a fast, efficient form of data transformation.
http://www.uvm.edu/~ludobots/index.php/Main/ER2014
https://www.ll.mit.edu/publications/journal/pdf/vol01_no1/1.1.7.neuralnetworks.pdf