theory - number of stages between the presentation of word on page and muscles moving to speak
priming experiment � chance of recognising the target word (shown for longer) is greater if the subject is primed, by being shown either similar letters (letter level) or word in similar semantic category (word level)
� word on page
� pattern falls on retina
� letter recognition
� word recognition
� production of a motor plan
� muscles move
� word produced
�as if� explanation - no idea what�s going on in the brain, but it�s as if the information excitation is flowing from here to there �
is that useful?
or is it only useful if I can show you what�s going on at the neuronal level
1. describe the behvaviour
processing one word speds processing of related words
2. produce an abstract model
can reproduce the behaviour � nodes representing words with activation passing between them
i.e. activation speeds processing
3. identify the brain mechanism actually implementing the behaviour
e.g. voluntary action <= dopamine flowing in the basal ganglia
building a computational model demonstrates that the explanation adequately explains the phenomenon � concretises the claim
leads � connectionist modelling (central to many areas)
connectionism = sophisticated + developed version of interactive activationist model
activation flows from one level to another or within a level � either excitatory (pointed triangle arrow) or inhibitory (blob arrow)
trying to get from the outside world to the model�s knowledge of a domain
(Rumelhart & McClelland, 1981)
the presence of a feature � excite/inhibit the presence of each letter
neuronally-inspired � its basic element leads back (conceptually mimic) some aspects of the way the brain works
neurons are all about integrating information
connections are very cheap and profligate
always/never been the case that there�s a correlation between horizontal
classical conditioning (fundamental aspect of the way the NS works) - events in the outside world and internal states of knowledge
in this model, letter-level and word-level information are different
at a conceptual level,
1. sums up info � some exc, some inhib � if > threshold, then it fires
the model is not completely arbitary � based loosely on the brain
the threshold for �T' is such that it would only fire if both the mid-vertical bar and the top-horizontal bar were present, which would only be > threshold if there were no inhibitions present (e.g. diagonal bars)
info proc = a very large number of feature detectors operating in parallel � trying to provide evidence about the state of the world to a system which works out on the basis of how the outside world�s been in the past
could a system evolved the make certain survival-based decisions about the outside world learn to read?
the model takes all the four letter words in English with frequency > 2 per million (1179) � obviously an over-simplification, but at least realistic in scope
the model really works � if forces the modeller to concretise/make explicit what is happening
can test to see whether it matches the RT (reaction time) data for normals
2. system operates by constraint satisfaction
the node which has most things consistent with it (i.e. the largest number of exc inputs) = the most active
the decisions are the best bet about the input given all available sources of information
this is how the brain works
doesn�t require certainty
mutual inhibition then suppresses all alternatives
if 2 paths which are almost identical, then the slightly more exc N inhibits the other one more, which inhibs the 1st slightly less
positive feedback � ensures that come to decisions quickly
but: if it was just noise, then it would magnify the noise (inherent instability)
what actually happens if given a degraded stimulus?
e.g. could be worK or worR or worD?
so although the evidence for K and R is equal at the letter-level
the activations for the diff words at word-level passes back down to the letter-level to further activate the �K� detector
we always work in a noisy environment � e.g. listening in a crowded room or reading a hand-written script
(Reicher 1969)
how does the model relate to normal data?
present (i) a word (�work�) or (ii) a letter (�***k�), followed by mask, so that identification = 50%
ask whether the letter was a �D� or a �K�? (can�the answer by guessing a word)
result: subjects do better in (i) than (ii)
inference: easier to detect a letter in a word than in isolation
how can you know about a word before you know about the letters of which it is composed?
the model shows that knowledge about a word influences knowledge about the letters which make it up
model shows that it makes snese to see the perception of words as involving independent stages of letter and word ID, confirming the evidence from both patients & normals
converging
operations
read a word aloud
more common words �/span> � quickly read it (frequency effect)
the NS processes things more quickly if it�s seen them before
if seen recently, � quickly (repetition effect)
the NS finds it easier to access information which it has just accessed
interaction � the effect of repetition has a larger effect on low-freq words than high-freq words
name a picture
the more common the word, the quicker
if you repeat the pic, repeat it quicker the 2nd time
inference: information in the brain is not stored like a dictionary (having looked up a word many times before does not make it easier to find again)
memory is adaptive to the use made of it
information used frequently/recently is easier to get at
high freq words have lwer mean naming latency than low freq words
interacts with regularity � �have� is irregular to pronounce
exceptionally low freq word which is irregular has much higher mean latency than high freq irregular word
subtle diffces in speed of processing �/span> inferences about the brain mech
if something seen frequently � stored so easier to access (freqency effect)
if the brian has a lot of similar xps, benefits from them (regularity effect) - asumes new xps will be like them
the way the brain stores information is diff to how it is sotred in phys media
possible stages between seeing a word and pronouncing it
computation model can be built in which flow of activation between nodes representing what the sys knows about current stim/past xp
can that give us explanation for freq, repetn and regy effects?
it can do that
varying the strength of the connection � can model the frequency effect (frequent use builds strong connections and activation in nodes with strong connections builds up faster)
time-lapsed � recency
how quick the route � regularity
does the fact that we have a computn model which gives the correct answer mean that we know any more about the way the brain does it?
no, because we still don�the know anything about how the brian does it
yes, because it�s brain-like and replicates results