PSYC 325 Lecture Notes - Lecture 6: Discrimination Learning, Broccoli, Collaborative Filtering
33 views3 pages
Psyc 325 ch6 generalization and discrimination learning
Generalization = transfer of past learning to new situation and problem
*core issue= find an appropriate balance btw specificity .
Discrimination = the perception of difference btw stimuli
-collaborative filtering: auto filter use in amazon based on info from large no of other people past behavior and
create a very detailed stereotype of you.
When similar stimuli predict similar outcomes
Eg: broccoli and cauliflower = both nasty
-trained pigeon respond to yellow light for food reinforcement , they also respond to light similar to yellow. As
the color grew increasingly different, the respond decrease.
*generalization gradient= curve showing how change in the physical property of stimuli correspond to
changes in responding. Measure person perception of similarity in that if 2 stimuli perceived as highly
similar, there will be significant generalization btw them
-challenge of generalization= consequential region: to identity the set of all stimuli that have the same
consequence as the training stimulus.
-generalization in pigeon is their best estimate of the probability that novel stimuli will have the same
consequence as the training stimulus. Attempt to predict, based on past experience, the likelihood it will be the
-application of the RW model to pigeon operant paradigm? 5 input nodes needed for each of the 5 discrete
color (discrete component representation: each possible stimulus is represented by its own unique node)
how it works? First trained to respond to yellow light and will activate the yellow input node. At the end of the
training the weight from the yellow input node is strong. Later when new yellow orange stimuli presented, it
activate the yellow orange node. However since it has never been activate before, it not strengthened and will not
cause activation . therefore no response despite the similarity to yellow. => this model will produce strong
response to yellow, but not other color. Not a smooth generalization gradient => WRONG.
Critique: RW model only useful for understand and describing , predicting how organism learn about highly
dissimilar stimuli (tone and a light ) but not work well with similar stimuli
-distributed representation: stimulus are presented by overlapping sets of nodes or stimulus elements.
Similarities emerge naturally from the fact that 2 similar stimuli activate elements belong to both sets. Therefore,
what is learned about one stimulus will transfer or generalize to other stimuli that activate some of the same
* this network has 3 layers of nodes and 2 layers of weight
how it works? Net work is trained to respond to yellow and presentation will activate the correspond input node
and will activate the 3 nodes in the internal representation, which connect to the output node. When the yellow is
repeatedly presented with reward, weight from 3 internal node strengthened. Next when tested with yellow
orange, yellow and yellow orange share overlapping 2 internal nodes, then some response activation is produced
at the output. Next, more distinct color presented will not trigger strong response because only 1 node/ no node
*topographic representation: nodes responding to physically similar stimuli ( yellow and yellow orange) are
placed next to each other. The degree of overlap btw representation of 2 stimuli reflect their physical similarity
*it produce a generalization gradient with peak responding to the trained stimulus yellow and decrease response
for other stimuli. Smooth gradient.
When similar stimuli predict different outcome
Eg: green broccoli =nasty, cauliflower =yummy
*learn to discriminate btw broccoli and cauliflower despite their similar physical appearance because they led to
find more resources at oneclass.com
find more resources at oneclass.com