PSY370H1 Lecture : LEC 10 – Creativity and How We Learn Nov 19 2009
PSY370: Thinking and Reasoning
LEC 10 – Creativity & How we Learn
Nov, 19th, 2009
How do neuronetworks learn?
! parallel process
! the weight of connections are altered as we learn
! this is done through back propagation of error
! target value – performance value = error
! learning algorithm use statistical contribution of each connection to assign blame for the error
! then the weight of connection is altered slightly to the degree to which it had blame, you do not
turn it all off, because the blame calculation is very probabilistic
! this is back propagation of error, and we do this many times until the network is able to do the
task well
! but this theory requires the ability to independently learn, because it is presupposing the thing it
is trying to explain
! presumption that we know the target value
How do we get unsupervised learning?
Gefford Hinton: Internalization
Terms used below are not from Hinto but from Vervaeke
! network break down into 2 parts
! letter for parts, number for stages
A
B
1 [Wake]
weak reinforcement of world,
because it does not know how big
the sample is for the population
WORLD
2 (a)
[Sleep]
B's sample act as A's
Population.
B will provide back propagation
for A, because it has the target
value of what the world is in A's
eyes
2 (b)
A is really good at solving
the problem, and it teaches
B how to model now.
1
Go back to stage 1 now,
the loop occurs again,
indefinitely until B has a
good model of the world
! but the criticism might be that the machine will never truly achieve a real representation of the
world but really just a good model
! bad criticism b/c humans don't have that either (we are not God, and its unfair to ask the
machine to be one)
Ito et. al
! cerebellum is structurally dense
www.notesolution.com
Document Summary
Lec 10 creativity & how we learn. but this theory requires the ability to independently learn, because it is presupposing the thing it is trying to explain. presumption that we know the target value. Terms used below are not from hinto but from vervaeke. network break down into 2 parts letter for parts, number for stages. B weak reinforcement of world, because it does not know how big the sample is for the population. B will provide back propagation for a, because it has the target value of what the world is in a"s eyes. A is really good at solving the problem, and it teaches. Go back to stage 1 now, the loop occurs again, indefinitely until b has a good model of the world. but the criticism might be that the machine will never truly achieve a real representation of the world but really just a good model.
