1. Searching the truth: Visual search for abstract, well-learned objects Denis Cousineau, Université de Montréal This talk will be available at www.mapageweb.umontreal.ca/cousined

2. How do we find a target?

3. 3 Visual search: a basic proficiency… very little understood…

4. 4 Two models of visual search…  Serial search:  The famous 2 : 1 ratio of mean slopes;  Based on the MEAN response times;  Parallel search  Flat performance.  Unlimited capacity

5. 5 Some problems with these models…  This dichotomy difficult to conciliate with progressive transitions  Mean performances are little diagnostic  Mimicking (Townsend, 1990)  Standard deviations can also be mimicked…  2:1 ratio depends heavily on the stopping rule  How do we stop searching?

6. 6 Standard model: Serial Self-Terminating Search (SSTS) Get ready Implicitly: a Random-Order visual search model

7. Experiment 1

8. 8 Methodology: Visual search task  34 sessions of training; 10 sessions of test, 4 subjects, consistent mapping:  Targets:Distractors:  Targets had to be learned; * Fixation point Test display Reaction time measured since stimulus presentation Circles indicating where the stimuli will appear

9. 9 Mean results  A seems to be perfectly serial; B is the least “serial”  Yet, we will see that  B is nearly identical to A  None of them are random-order serial

10. 10 Results of Target-present RT distributions  A and B are the most similar!

11. 11 Modeling the modes of the distributions  The D =1 condition could be modeled with a normal distribution with parameters ;  The D = 2 condition should be the same as the D = 1 condition except shifted by and variance doubled;  In general, the distributions have parameters  The modes are pooled: a “mixture of distribution” -With parameter according to SSTS -With free mixture parameter  unrestricted model

12. 12 Results of Target-present RT distributions  For all participants, the mixture parameters are not equal to 1/D.  The last mode is underrepresented. Errors?

13. 13 Results of Target-absent RT distributions  B perform early termination  A does not, yet her ps are not equal!  C does this too often compared to his error rate

14. 14 In sum 1.Regarding the exhaustivity prediction:  The participants sometimes stop earlier than predicted by an exhaustive search  This predicts errors, but too many errors are predicted.  Regarding the random-order prediction:  The participants are serial…  …but they are not random  Seriality is one process going on, but there must be a second process which aims at biasing the search itinerary so that targets will be visited earlier than by chance.

15. 15 A new model of visual search: m-Sr-STS  The Mostly Serial, Roughly Self-Terminating Search  Essentially a two-stage model (Chun & Wolfe, 1996, Wolfe, 1994, Cousineau & Larochelle, 2004). The pre-attentive module outputs probabilities

16. 16 Yet, there is still some magic left…  Unbeknownst to the participants  was diagnostic:was irrelevant:  The pre-attentive module could drive attention on the stimuli having those conjunctions of features  A parallel search for conjunctions  It should be an impossible feat according to Treisman (1980), Wolfe (1994) and many others.

17. 17 Let’s concentrate on the decision mechanism  The Mostly Serial, Roughly Self-Terminating Search The pre-attentive module outputs probabilities  What is “Recognizing a target”?  How does cycling occurs?

18. Experiment 2

19. 19 Methodology: Same-different task  Well-trained participants (10 hours to reach asymptote then 5 hours of testing).  The display size D is fixed at 1;  The stimuli are varying in complexity C, e.g.

20. 20 Mean “Same” response times  Saying “Same”  is very fast  affected by C (20 ms/spike)  Linearity is not found using characters instead of complex stimuli  Parallel, limited-capacity models complies with such results  e.g. a template matching process?

21. 21 Mean “Different” response times  A main effect of the number of differences but no effect of complexity!  Suggests that responding “Different” requires the localization of at least one difference.  Parallel search for a difference benefits from the presence of many differences

22. 22 The Revised possible explanation  There might be two distinct processes:  one for confirming the sameness,  one for establishing the “differenceness”  How do they relate to one another? In succession?

23. 23 Slow “Same” vs. fast “Different” in the C = 4 condition  The two conditions are very close (mean difference of 13 ms). Do they follow in time?  Again, let’s look at distributions

24. 24 Distributions of RT in Same and (very) Different responses at C = 4  The slow “Different” responses are faster (by 4 ms) than the slow “Same” responses.  One process cannot operate *after* the other.

25. 25 Revised revised-architecture  “No” may not be an option for a neural decision mechanism…

26. In conclusion…

27. 27 Visual search is a proficiency (1/2) Proficiencies are an amalgam of processes  Parallel pre attention process outputs probabilities  Serial deployment of central attention  Stopping rule which can end prematurely  Unitary (template matching?) recognition process  Unitary (find-a-difference) rejection process  In sum, the SSTS architecture was all wrong.

28. 28 Visual search is a proficiency (1/2) Processes are univoque (from french: One and only one meaning, one and only one semantic content, but also one and only one voice)  As an example  If a “not-face” is presented to a face recognition module, does it “knows” that it is not a face, or does it remains “silent” by omitting to respond…  What would be a brain which detects objects (of many kind) and their negation? what would be the EEG of such a system?  Negation is not part of the neural process toolbox  it is not “To be or not to be” but “To be and to un-be”  “NO” branches should be forbidden in psychology.

29. 29 Methodological consideration  Distribution analyses rocks!  Mean results can be interpreted in so many ways that they cannot reject any model at all.  We have been stuck with a fruitless dichotomy for over 40 years because we were unable to make the data speak.  Anyone with a serious model should implement it using distributions or remain quiet  Distribution modeling and testing is not difficult (it can be learned in 3 hours).  as long as you know matlab or mathematica…

30. Thank you. This talk will be available at www.mapageweb.umontreal.ca/cousined