Week three - pattern recognition

Top down processing

• seeing the whole thing followed by parts

• prior experience and expectations shape this processing 

  • ex experienced readers use top down processing when reading (don’t read letter by letter)

Bottom up processing

seeing the parts and then putting the whole thing together

• inexperienced readers who read letter by letter are using this processing

Apperceptive Agnosia

• only perceptual 

• can see it through feature detectors but cannot recognize it 

• ex. person asked to draw a cup, know what a cup is and what it looks like but cannot recognize or draw the object

Associative agnosia 

can tell all the parts of an objects but cannot put a name to it 

Prosopagnosia

• cannot recognize faces

• they may know a person through voice and know what the person looks like internally, but when seeing them in person they would not recognize

Template theoretical approach

in order to recognize an object ,we need an exact template

• if an object changes, rotates, transforms, our template does so as well

• problem is we do not have templates for novel objects, so how can we recognize them?

Feature Analysis

since certain neurons respond to certain features, we use all of our features to put an object together 

Pop out affect

when you are looking for an object in a lot of objects and it has one feature that is able to pop out so we immediately recognize it with no effort

Conjunction search

when we are searching through more than one feature, it takes ore effort to recognize it

example of feature analysis 

finding a tilted line in a bunch of vertical lines 

• its harder to find a vertical line in a bunch of tilted lines because we do not have the feature for vertical lines but we do have a feature for tilted lines

GRAPH OF FEATURE ANALYSIS VISUAL SEARCHES

X AXIS: # of features to search for

Y AXIS: Reaction time

the two lines will not be parallel

Componential/ Structure Approach

• Biederman RBC theory AKA geon theory

• only 36 basic shapes are important for us to recognize object

• Problems: difficult to distinguish objects based just off aeons because different objects can be made up of the same basic neons

  • we also process faces holistically and not by its geons

Thatcher Illusion of Face Processing 

• if we see faces upside-down we see them as normal but once they are right side up we see that they are not normal 

Feature detectors and word recognition

feature detectors have:

• activation levels required to reach threshold

• response threshold the point at which detectors will fire to recognize the word

• like bucket, as input fills the bucket it gets closer to threshold

What determines detectors starting activation level?

• word frequency: cat will have a lower threshold than quail

• If detector was fired recently: if a word is previously primed, it will take less activation to read the word again

What makes word recognition hard or easy? 

familiarity and frequency of the word in your life 

Word Superiority effect

the idea that its easier to recognize a letter as part of a word rather than on its own, this does not work when there is well formedness

Well formedness

if a word looks good and is possible in the English language we can read and recognize it as a real word (ex. GLAKE)

Making errors when reading 

when we are reading we are primed by the context of what we are reading and know what we are trying to say/read. this is why we are so bad at proof reading

What is over regularizing errors an example of

top down processing - we are influenced by our prior knowledge of words and expectations of what it should be, so we read words that are misspelt easily because our top down processing fills in the gaps

Feature Nets

a theory of how our brains layer detectors

• there are layers of feature nets: pieces of letters, letters, word and each layer gets larger objects for recognition. Once threshold is reached, the feature net is activated and the word is recognized.

Explain how feature nets work

1. when we look at a word, our feature detectors activate in response to basic shapes like lines, curves, edges of letters

2. activations spread to the next level, triggering letter detectors activating whole letters

3. word detectors integrate the letters to make up the word

4. we move to the next layer when our feature detectors reach threshold 

How do feature detectors use bottom-up and top-down processing

• feature detectors use bottom up processing because our detectors for parts of the letters are activated before we can recognize the whole letter

• feature detectors use top down processing because context, meaning, or previous knowledge can support the recognition of the word and make it faster

What can feature detectors not explain

wellformedness : if we have feature detectors for word recognition, why can we also easily recognize words that are not real? how is our threshold reached in that scenario?