COMP9444 Neural Networks

Solutions to Exercises on Temporal Processing 2

  1. For each network described in the section of the course on temporal processing, classify the input and output data representations used according to whether they are:

    • binary encoded
    • one-of-many-choices/winner-take-all
    • thermometer coded
    • grey-scale values
    • localist versus distributed
    • microfeatural

    N.B. The networks can and will have a combination of these attributes.

    Answer:

    NetTalk:
    input representationoutput representation
    √ one-of-many-choices/winner-take-all
    √ localist
    		 √ one-of-many-choices/winner-take-all
    √ localist

    TDNNs:
    input representationoutput representation
    √ grey-scale values
    √ localist
    √ microfeatural
    		 √ grey-scale values
    √ localist

    8-5-8 encoder:
    input representationoutput representation
    √ one-of-many-choices/winner-take-all
    √ localist
    		 √ one-of-many-choices/winner-take-all
    √ localist

    Jordan shift register:
    input representationoutput representation
    √ I guess this one is "reinforced one-of-many-choices"
    √ localist
    		 √ one-of-many-choices/winner-take-all
    √ localist
    Note that the .data and .teach files say "distributed". We could have used "localist" .data and .teach files for these, but it is easier for a human to understand what tlearn refers to as "distributed".

    Sequential XOR:
    Prediction tasks inevitably all have the output representation the same as the input representation.
    input representationoutput representation
    √ one-of-many-choices/winner-take-all
    √ localist
    		 √ one-of-many-choices/winner-take-all
    √ localist

    badiiguuu net:
    input representationoutput representation
    √ distributed
    √ microfeatural
    		 √ distributed
    √ microfeatural

    manyyearsagoaboyandgirl... (concept of a word):
    input representationoutput representation
    √ binary encoded
    √ localist versus distributed
    		 √ binary encoded
    √ localist versus distributed

    monster eat cookie (discovering lexical classes):
    input representationoutput representation
    √ one-of-many-choices/winner-take-all
    √ localist
    		 √ one-of-many-choices/winner-take-all
    √ localist

    starting small:
    Not really specified in lectures, but ...
    input representationoutput representation
    √ one-of-many-choices/winner-take-all
    √ localist
    		 √ one-of-many-choices/winner-take-all
    √ localist

    graphotactic prediction (Tower nets):
    input representationoutput representation
    √ one-of-many-choices/winner-take-all
    √ localist
    		 √ one-of-many-choices/winner-take-all
    √ localist

    backprop-through-time shift register:
    input representationoutput representation
    √ binary encoded
    √ distributed
    		 √ binary encoded
    √ distributed

Copyright © Bill Wilson, 2008.
Bill Wilson's contact info

UNSW's CRICOS Provider No. is 00098G