McClelland Rogers 2003 Review PDP Parallel Semantic Cognition
<p>Nov 15, 2009 9:07 PM
 
<b>COHERENT VARIATION</b> CARVES <b>CONCEPTUALIZED</b> NATURE AT ITS JOINTS
 
Nature Neuroscience Reviews  2003  April
 
<b>310 Ross Quillian hierarchy from the sixties.  Taxonomic hierarchy...  Collins and Quillian 1969.</b>
 
1975 Warrington described semantic dementia...  lose detail first and progressive loss as you ascend the tree.
 
<b>Children seem to learn genl concepts first and then make finer and finer distinctions</b>.
 
However some debate abt whether kids also make a <b>perceptual-to-conceptual shift</b>.
 
Warrington suggested the development follows Quillians hierarchy down and disintegration starts at the
bottom and works its way up.
 
Graded categoy membership is important.  TYPICALITY is important.
 
Subjects verify  robin is a bird   faster than checken is a bird.
 
So, in 1970s several alternatives to Collins and Quillian arose to capture graded membership.
Explicit Bayes approach to categories.
If a dog has an omentum does a duck have one too.
 
They describe PDP models that capture important properties.
 
313  Box 1:  How Learning Takes Place ina FEEDFORWARD PDP network
 
I = input;  H = hidden;  O = output
 
No recurrent loops or within layer loops.
 
LEARNING depends of the DIFFERENCE btwn the TARGET and the obtained activations of the output units.
Error = (Observed  -  Expected).  Adjust weights to reduce the rror E sum of squares of difference.
 
 
 
 
Coherent Covariation  is critical.
 
 
317
 
 
318
 
 
<b>319:  Theory Theory ...</b>  states that there are some types of knowledge that are much easier to learn
because of innate tendencies.  We then use central causal beliefs to organize learning that knowledge.
Babies are like little scientists - they have causal theories of the world that guide thier acquisition of
knowledge.
 
   <b> Jay suggest instead that coherent covariation can do it all.</b>
<b>Properties that covary coherently come to be treated as central..</b>. even tho objects may differ in
other respects.  Salient properties are flying and having wings in birds.
 
    Consider  Has wings  vs  IS yellow.
(read para at top of second column ... it is crucial!)
 
Having WINGS coherently COVARIES with the other properties that canary shares with robins...
animal, move, fly, has skin, feathers    ...  therefore, the connection weights are mutually re-inforcing.
The model's sensitivity to coherent covariation explains why some sets of properties are
easier to learn and remember than others ... and they are central to category membership.
 
 
 
<b>320  PDP model</b> generalizes difft properties differently because their importance differs depending
on the domain.  eg  with TOYS, shape is important;  with foods, color is important.
 
Btwn  TREE and plants... size is crucial;  but does not covary with fish and birds.
 
Similarly,  being YELLOW is simply tacked on to flowers and birds and is not an essential feature ...
it does not coherently vary.
 
Carey... a theory-theory advocate  emphasizes evolution or change in children's concepts as
they move from superficial similariteis (appearances/percepts)  to  conceptual distinctions.
But Jay says same thing happens with PDP.
If the model has more frquent exposure to appearance items than it learns based on that.
However, later as it acquires info with coherent covariation abt HAS, CAN, and ISA info that
comes to dominate learning.
 
 
<b>321  To learn causal sequences, just break it into series of steps and learn to predict consequences</b>
from previous steps.  EG  remote control to control TV - point at TV, needs batteries, press button.
Generalize to other similar devices based shared properties that COVARY and not to color of
remote control etc that do not.
 
<b>321  In summary.... Theory-Theory not required;  the very simple model of Rumelhart network can capture</b>
many of the relevant findings.   Causal info may be learned thru event sequences.
Optimistic that these models will work for the full range of findings.</p>