Neural Networks

1
Neural Networks

• My name is Burleson

2
Neural Networks vs Conventional Computing

• Programming is broken into small, unambiguous
  steps
• Algorithms must be already known and understood
  in order to obtain a result
• Errors in results are always software or hardware
  based

• Neural networks are trained
• Can obtain results when there is no solution
• Errors are based on faulty training data sets

3
General Structure

• Composed of nodes arranged in layers (or evil
  lairs, depending)
• Each node can take different input or the same
  input and output various results
• Unique patterns among nodes determine an end
  result of the execution
• Back propagation can allow a network to train
  itself by giving feedback to each node

4
Three Types of Learning

• Supervised learning involves a human training the
  network repeatedly in order to improve accuracy
  (military facial recognition)
• Unsupervised learning can be done by the computer
  where certain cost functions aim to be minimized
  (image compression)
• Reinforcement learning is based on the computer
  building its own data and extracts reliable
  knowledge based on repetition of events (game
  intelligence)

5
Potential Problems

• Over training can lead to the network not being
  able to identify new input correctly if the
  network it too dependant on the original training
  data
• Generalization can lead to false positives if the
  training set was too diverse or the error
  correction was not strict enough
• Error correction is touchy - steps can be taken
  in large or small increments with smaller steps
  being more accurate but slower and larger steps
  doing the other (like, duh, for sure)

6
Explanation of a Simple Network

• I have no idea, but luckily for us, this site can
  explain (please pull out your binoculars for the
  remainder of the presentation)
• http//www.doc.ic.ac.uk/nd/surprise_96/journal/vo
  l4/cs11/report.htmlAn20engineering20approach

7
Interesting Sites

• http//rodin.wustl.edu/cosc/doc3.html
• http//www.20q.net/