Ervin Hafter and AnneMarie Bonnel Department of Psychology
Ervin Hafter and Anne-Marie Bonnel Department of Psychology U. C. Berkeley “A role for memory in Shared Attention” "Theoretical and Experimental Approaches to Auditory and Visual Attention" Cold Spring Harbor April 20, 2008
Simplest definition of attention: a process inferred when responses on one task are affected by responding simultaneously to another
Two paradigms Detection S(+) S(0) Identification S(+) S( ) Energy detection model: compare stimuli in the signal epochs S(0) is the level of the standard.
S(+) λ S(0) A S(+) λ z B Area under the ROC S(0) 1. 00. 90. 80 A . 70. 60. 50 Predicts d’B > d’A B 0 0. 5 1. 0 1. 5 2. 0 2. 5 Signal Levels in z % correct in 2 AFC Probability density Ideal detection based on energy in the signal epochs.
But, Macmillan (w/tonal pedestal) and Bonnel et al. (w/visual pededestal) found d’detection > d’identification
Audio demo comparing detection to identification. Stock Market National Debt University salaries
Both attributed this to use of Transients
Thinking that, unlike the case with energy, responses to transients might be pre-attentive, Bonnel et al. (1992) tested in a dual task with independent stimuli on side-by-side LEDs.
Thinking that, unlike the case with energy, responses to transients might be pre-attentive, Bonnel et al. (1992) tested in a dual task with independent stimuli on side-by-side LEDs. Detection was comparable to performance found with S instructed to attend to only one LED; identification showed a tradeoff indicative of a shared attentional resource.
Thinking that, unlike the case with energy, responses to transients might be pre-attentive, Bonnel et al. (1992) tested in a dual task with independent stimuli on side-by-side LEDs. Detection was comparable to performance found with S instructed to attend to only one LED; identification showed a tradeoff indicative of a shared attentional resource. Caveat: “Perceptual grouping” in identification
Some changes in Berkeley Visual Auditory S(-) S(0) S(+) CR T z H 0 50 17 ms
Detection S(+) S(0) S( ) Identification S(+) S( )
Ideal energy detection λ S(0) S(+) λ B C S(0) S(-) A Area under the ROC S(+) -2 -1 0 +1 +2 λ λ z . 90. 80. 70 B A C . 60. 50 0 0. 5 1. 0 1. 5 2. 0 2. 5 Signal Levels in z λ % correct in 2 AFC Probability density S(0) 1. 00
Use the strictest possible criterion for asserting that there is no cost of shared attention: compare performance in the dual-task to that found in the separate single tasks.
Support for transient hypotheses: Attention-Operating Characteristic 1) detection > identification 2) No cost in detection 50%, 50% 3 80%, 20% d’audition 3) Tradeoff in accord with instructions in identification and are single tasks 20%, 80% 2 50%, 50% 80%, 20% 1 0 20%, 80% 1 2 d’vision 3
SI = d’dual d’single SIAOC Sharing Index (Audition) Sharing-Index Attentional Operating Characteristic (SIAOC) normalizes data from a dual task in terms of the single-task controls. 1. 0. 8. 6. 4. 2 0 . 2 . 4 . 6 . 8 1. 0 Sharing Index Vision)
SIAOC 1. 0 Sharing Index (Audition) These data clearly imply that detection of transience put no demand on shared attention, unlike discrimination of energy. . 8. 6. 4. 2. 2 . 4 . 6 . 8 1. 0 Sharing Index (Vision)
A more direct test of the idea that detecting transience (change per se) doesn’t require shared attention simply removes transients as information.
A more direct test of the idea that detecting transience (change per se) doesn’t require shared attention simply removes transients as information.
A classic ΔI/I Another classic ΔI/I Often called a reminder
Auditory single task no gap 4. 0 Same signal levels 3. 0 d’ 2. 0 1. 0 no 200 400 600 800 1000 1200 1400 1600 1800 gap Gap duration (ms) Energy detection
Dashes suggest absolute identification, i. e. comparisons to long-term, context-coded memory, rather than to a sensory trace of the reminder. 4. 0 Same signal levels 3. 0 d’ 2. 0 1. 0 200 400 600 800 1000 1200 1400 1600 1800 Gap duration (ms)
represent 500 -msec signals. SI-AOC “different” “same” “different” 1. 0. 8 “larger” “smaller” . 6. 4. 2 < 300 >. 2 . 6 1. 0. 4. 8 Sharing Index (Vision)
Test Duration {D} Integration Time {IT} Impact of increased duration. e. g. , , on apparent cost of attention. A or V {D} < {IT} d’ 1 Single task D 1 = {D}
Test Duration {D} Integration Time {IT} A or V {D} < {IT} d’ 1 Single task D 1 = {D} < {IT} Dual task D 2 = ½ {D} A d’ 2=. 707 d’ 1 V d’ 2=. 707 d’ 1
Test Duration {D} Integration Time {IT} A or V {D} > {IT}. d’ 1 Single task D 1 = {IT}
Test Duration {D} Integration Time {IT} A or V {D} > {IT}. Single task D 1 = {IT} d’ 1 {D} > {IT} . 707 d’ 1< d’ 2 ≤ d’ 1 Dual task ½{D} < {D 2} < {IT}
Test Duration {D} Integration Time {IT} A or V {D} > {2 IT}. d’ 1 Single task D 1 = {IT}
Test Duration {D} Integration Time {IT} A or V Single task D 1 = {IT} {D} > {2 IT}. d’ 1 {D} > {IT} d’ 2 Dual task {D 2} = {IT} *Here, time sharing produces the same result as no sharing.
A still stronger test forces the subject to use context-coded memory by simply removing the reminder. “other” “standard” “other” “large” “small”
It seems clear that responses based on context-coded memory were limited by sharing of an attentional resource. Sharing Index (Audition) 1. 0. 8. 6. 4. 2. 2 . 6 1. 0. 4. 8 Sharing Index (Vision)
Perhaps the reason that use of change per se did not provoke a cost of sharing is that it was done in sensory trace (? rehearsal? ) memory?
Subjects can be forced to use sensory-trace memory by roving the standard from trial to trial. Trial Rove Standard Test Correct response 1 ‘louder’ 2 ‘softer’ 3 4 ‘softer’ ‘louder’
Unlike the case with context-coding, performance fell as the ephemeral sensory-trace faded over time. 3. 0 2. 5 2. 0 da 1. 5 Fixed levels Roved levels 1. 0 0. 5 *Need higher signals in roving 1 2 3 4 5 6 GAP Duration (sec) 7 8 Auditory Identification 50 -ms reminders and signals
Most intriguing is that despite very poor performance, especially with long delays, there was no cost of sharing. rove <Gap (ms)> Gap (ms) 310 510 1022 8350 Sharing Index (Audition) 1. 0. 8. 6. 4. 2. 2. 6 1. 0. 4. 8 Sharing Index (Vision)
In typical, everyday life, we label sensory stimuli on the basis of comparisons to long-term memory. Test Sensory/ Neural Compare Experience Context-coded, long-term memory loud, dim, green, hot, salty, etc.
When presented with an adjacent standard, the response may be to Reminder T 1 Sensory/ Neural SOA Test Sensory/ Neural T 2 Compare Experience Context-coded, long-term memory loud, dim, green, hot, salty, etc.
When presented with an adjacent standard, the response may be to simply ignore it, labeling the test in accord with long term memory. Reminder T 1 Sensory/ Neural SOA Test Sensory/ Neural T 2 Compare Experience Context-coded, long-term memory loud, dim, green, hot, salty, etc.
Without a reliable context-coded memory, S must compare the test to the reminder the in sensory trace memory. Rehearsal Memory Reminder T 1 Sensory/ Neural SOA Test Compare Sensory/ Neural T 2 Compare Experience Context-coded, long-term memory louder, dimmer, greener, hotter, saltier, etc. . loud, dim, green, hot, salty, etc.
Our audio/video dual-task shows these comparisons to be independent, i. e. , no cost of sharing. Rehearsal Memory Reminder T 1 Sensory/ Neural SOA Test Compare Sensory/ Neural T 2 Compare Experience Context-coded, long-term memory Conversely, these comparisons were limited by a shared attentional resource. louder, dimmer, greener, hotter, saltier, etc. . “loud, dim, green, hot, salty, etc. ”
Okay, so comparisons to the sensory trace memory produced no cost of shared attention. What in the world is trace memory?
Okay, so comparisons to the sensory trace memory produced no cost of shared attention. What in the world is trace memory? Recently, we’ve approached this in terms of Weber’s Law. Bringing the lab up to 1834.
Weber’s Law I=k. I In signal detection terms, k can be treated as a multiplicative noise
Identification Without Roving: performance based on long-term, labeled memory produce a constant Weber fraction 3 Threshold 10 log (ΔI/I) Ped = Sig = 50 ms Gap = 1 s 1 -1 -3 -5 -7 55 60 65 70 Pedestal (d. B) 75
What happens when comparisons are to a roved standard? Thresholds go up. But in what way? To answer this, we parse the data in terms of the individual standards, analyzing performance separately for each pedestal level.
No Rove Labeled memory Rove Trace memory Threshold 10 log (ΔI/I ) 3 1 -1 -3 -5 -7 55 60 65 Pedestal 70 75
The change in slope implies a second, additive noise, c. 3 Threshold (d. B) 1 I=k. I+c -1 mul tipl icat i -3 ve + add itiv e no -5 ise multiplicative noise -7 55 60 65 70 Pedestal (d. B) 75
6 4 2 0 -2 -4 -6 55 60 65 70 75
ΔI = k. I + c What is c? Perhaps it is a decision noise associated with responding to the stimulus in trace memory? Maybe it is simply the result of decay in the trace that makes that adds noise to the remembered amplitude code.
Our next plan is to go into f. MRI in search of sensory rehearsal. Wish us luck.
- Slides: 50