nov 15

or, and, xor

• pg 738 in text ...
• (mention typo for NOT)
• explain fig 20.17 and "bias"
• single layer "perceptron" in more detail :
  • "linear separator"
  • explain pg 741 picture
  • "soft threshold"
  • why XOR cannot be done with 1 layer
• mutiple layers
  • combine "soft threshold" to select nearly anything
  • be clear that real networks have *many* more dimensions
    • how many dimensions in digit recognition done Mon?
• several of the assignments ask you to discuss this

FANN

• one library to actually implement this stuff
• http://leenissen.dk/fann/
• in C ... but bindings to a host of other languages (Python, Perl, Squeak, Octave, ...)
• see tiny demo in fann
• also see article on FANN
• discuss optional coding assignment
  • ... and perhaps even start to set it up
• includes "cascade training"
  • one description of that is here: http://www.dtreg.com/cascade.htm

research

• Yann LeCun (NYU learning lab)
  • http://yann.lecun.com/exdb/publis/
  • "energy methods"
  • http://www.cs.nyu.edu/~yann/research/dave
• face recognition

convolutional methods

• define
• pg 267, "Pattern Recognition and Machine Learning"
• ... and mention some of the topics in (and level of) the book
• briefly touch on edge detection and "convolution" in general

recurrent networks

• http://www.willamette.edu/~gorr/classes/cs449/rnn1.html
• ... and following

projects

• http://www.willamette.edu/~gorr/classes/cs449/project.html

summary

• inspired by brains ...
  • but aren't particularly close (at least not yet)
• typically used for perception (image recognition especially)
  • but also other sorts of "fuzzy" patterns
• training phase; execution phase
• fairly standard technique these days
  • even commercialized, e.g. http://www.dtreg.com