Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT)...
Transcript of Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT)...
![Page 1: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/1.jpg)
Perception• Bottom-up perception
– Driven by stimulus
• Top-down perception– Driven by expectations
• Interactive– Driven by both at the same time
![Page 2: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/2.jpg)
Knowing where an object is can make anotherwise invisible object appear
![Page 3: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/3.jpg)
![Page 4: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/4.jpg)
![Page 5: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/5.jpg)
![Page 6: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/6.jpg)
![Page 7: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/7.jpg)
![Page 8: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/8.jpg)
![Page 9: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/9.jpg)
![Page 10: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/10.jpg)
Matisse
![Page 11: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/11.jpg)
![Page 12: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/12.jpg)
Units of Perception• What are the basic features of perception?• Evidence for chunks
– Context effects (dependencies between parts)• ABC/12 13 14• Fruit faces (Palmer, 1982)• Part in whole versus part judgments (Tanak & Farah, 1993)
– Lack of complexity effects (Goldstone, 2000; Shiffrin & Lightfoot, 1997)– Interference between parts
• Garner interference: can’t ignore irrelevant dimensions (Garner, 1974)– Inversion cost (Gauthier & Tarr, 1997)
• Upright recognition much easier than inverted for familiar configurations– Blindness to parts within chunks (Tao, Healy, & Bourne, 1997)
• proofreading example
• Chunks as modules– Large within-chunk dependencies, small between-chunk dependencies
![Page 13: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/13.jpg)
![Page 14: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/14.jpg)
![Page 15: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/15.jpg)
Archimbaldo
![Page 16: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/16.jpg)
Finished files are the result of years of scientificstudy combined with the experience of manyyears.
ProofreadingCount the number of “F”s:
Tend to miss letters when the word they occur inis highly frequent.Automatic “gluing” of letter to its word
![Page 17: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/17.jpg)
Brightness and Size are SEPARABLE Dimensions
![Page 18: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/18.jpg)
Brightness and Saturation are INTEGRAL Dimensions
![Page 19: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/19.jpg)
Word Superiority Effect
WORKXXXX ---K Or ? ---D
K ---K Or ? ---D
Subjects are more likely to choose the correct letter when itis in the context of a word than when it is isolated.
Context improves people’s sensitivity, not just bias
XXXX
![Page 20: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/20.jpg)
Word Superiority Effect
DXXXX…… CARDXXXX……
D or T? D or T?
![Page 21: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/21.jpg)
Aspects of the Word Superiority Effect (WSE)• Letters better identified in words than in non-words or by
themselves• Words as perceptual units• Sensitivity, not bias, effect
– Bias to give respond with letter than would form a real word cannot explainWSE because both letter choices form a real word
• Pseudo-word superiority effect too– E in MAVE is better identified than E in VMAE
• Pattern mask is important for WSE
![Page 22: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/22.jpg)
Interactive Activation ModelMcClelland & Rumelhart (1981)
![Page 23: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/23.jpg)
Interactive Activation Model (IAT)• Cascading activation
– Feature-level processing not complete before higher-levels start– Top-down and bottom-up is not viciously circular– Contrast to standard information processing
• Architecture– Feature, Letter, and Word level units– Activation between levels, inhibition within levels– Lateral inhibition
• Good for creating discrete edges, category, decisions• Digitalization
• Processing: flow of activation/inhibition alonglinks
![Page 24: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/24.jpg)
Phenomena Explained by IAT• Word superiority effect (WSE)• WSE strongest with pattern mask• Pseudo-word effect• Rich-get-richer effect• Gang effect
![Page 25: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/25.jpg)
![Page 26: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/26.jpg)
![Page 27: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/27.jpg)
![Page 28: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/28.jpg)
†
ni t( ) = a ije j t( )j
 - g ikik t( )k
Â
n= net input, a=weight of excitatory input, e =activation ofincoming excitatory node.g=weight of inhibitory input, i = activation of incominginhibitory node.
Net input to node is based on consistent and inconsistentinputs.
![Page 29: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/29.jpg)
†
Ei t( ) = ni t( ) M - ai t( )( )if ni t( ) > 0Ei = effect on node i, M = maximum activation possible, ai =activation of node i.Effect on node is based on input, but has a ceiling at M. The closer the current activity ofthe node is to M, the less the effect of positive input will be.
†
Ei t( ) = ni t( ) ai t( ) - m( )if ni t( ) < 0m = minimum activation possible.If input is negative, then the floor is at m. If currentactivity is already at floor, then input has no effect.
![Page 30: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/30.jpg)
†
ai t + Dt( ) = ai t( ) -q i ai t( ) - ri( ) + Ei t( )q=rate of decay, r = resting level of unitNew activity is based on old activity, and decay to a restinglevel, and the effect of the input to the node.
†
a i t( ) = ai x( )-•
t
Ú e-( t-x)rdx
a-bar = running average of activation, decay of old information.A cumulative average across time of a unit's activity will be its strength.More recent activity levels matter more than older activity levels, and thedecay rate of old information is based on r.
![Page 31: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/31.jpg)
†
si t( ) = ema i t( )
Si = response strength of unit i, m=steepness of functionrelating activation to response.
Exponential functions serve to emphasize differencesbetween larger quantities, which is important becauseactivations are capped at 1. The difference between .8 and.9 should be greater than between .7 and .8.
![Page 32: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/32.jpg)
†
p Ri ,t( ) =si t( )
sj t( )j Œ LÂ
P(Ri,t) = probability of responding with unit i's responseL = set of nodes at same level as i.
Luce choice rule: if you have N alternatives that aremutually exclusive (can only do one of them), then this ruleassures that the probabilities will add up to one, and theprobability of making a response is based on its relativestrength.
![Page 33: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/33.jpg)
![Page 34: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/34.jpg)
![Page 35: Perception •Bottom-up perception - Indiana University · Interactive Activation Model (IAT) •Cascading activation –Feature-level processing not complete before higher-levels](https://reader030.fdocuments.us/reader030/viewer/2022040913/5e8840f81505ee67f704dc4f/html5/thumbnails/35.jpg)