A User's Guide to Bubbles Frédéric Gosselin Philippe G. Schyns
Contents
1. P G 2001 Bubbles A technique to reveal the use of information in recognition Vision Research 41 p 2261 2271 Gosselin F and Schyns P G 2002 RAP a new framework for visual categorization Trends in Cognitive Science 6 p 70 77 Hubel D H 1988 Eye Brain and Vision New York Scientific American Library Humphreys K Gosselin F Schyns P G Kaufman J and Johnson M H Do 7 month old infants use the same information as adults to process facial identity International Conference on Infant Studies Toronto Canada 18 21 avril 2002 Jentzsch I Gosselin F Schweinberger S R and Schyns P G Applying Bubbles to Understand the Face Information Driving Event Related Brain Potentials ECVP Kobatabe E amp Tanaka K 1994 Neuronal selectivities to complex object features in the ventral visual pathway of the macaque cerebral cortex Journal of Neurophysiology 71 856 867 Leclerc J and Gosselin F 2002 Adaptative Bubbles ECVP Leder H and Bruce V 2000 When inverse faces are recognized The role of configural information in face recognition Quarterly Journal of Experimental Psychology A 53 513 536 Moscovitch M Behrmann M and Winocur G 1997 What is special about face recognition Nineteen experiments on a person withvisual object agnosia and dyslexia but normal face recognition Journal of Cognitive Neuroscience 9 555 604 Nielsen K J Rainer G Brucklacher2. and Gosselin F 2002 Use of frequency informaion at the basic and the subordinate level in animal categorization Vision Sciences Schyns P G Jentzsch I Johnson M Schweinberger S R amp Gosselin F 2002 A principled method for determining the functionality of brain responses Submitted Searcy J H amp Bartlett J C 1996 Inversion and processing of component and spatial relational information in faces Journal of experimental Psychology Human Perception and Performance 22 904 915 Sigala N amp Logothetis N 2002 Visual categorization shapes feature selectivity in the primate temporal cortex Nature 415 318 320 Simoncelli E P 1999 Image and Multi scale Pyramid Tools Computer software NewYork Author Tanaka J W and Farah M J 1993 Parts and wholes in face recognition Quarterly Journal of Experimental Psychology 46A 225 245 Tinbergen N and Perdeck A C 1950 On the stimulus situation releasing the begging response in the newly hatched herring gull chick Larus argentatus Pont Behaviour 3 1 39 Valentine T and Bruce V 1986 Recognizing familiar faces the role of distinctiveness and familiarity Canadian Journal of Psychology 40 300 305 Vinette C and Gosselin F 2002 Spatio temporal use of information in face recognition Vision Sciences Vogels R 1999 Categorization of complex visual images by rhesus monkeys Part 2 single cell study European Journal o
3. and the response s to measure Applying these to the FIE we revealed that differences in local face information use the eyes nose and mouth represented at the scales between 90 and 22 5 cycles per face determined the effect This subspace at least in the visual realm takes the shape of a diagnostic filtering function that can be applied to render the effective stimulus of the task There are obvious shortcomings to Bubbles The first main shortcoming is the combinatorial explosion in number of trials that are required to exhaustively explore the search space Practically Bubbles is tractable within low dimensional search spaces and users are advised to restrict the dimensionality of the search space to be as low as possible High dimensional problems are made tractable with heuristics that guide the search towards regions of more promising solutions leaving aside less promising regions of the space Heuristic search can be performed with adaptive statistical sampling or their implementations However any heuristic search introduces biases resulting in a trade off between speed and suboptimal solutions i e local minima In any case the number of trials in Bubbles will need to be reduced to apply the techniques to brain damaged patients children or when learning is itself a factor of the experiment To illustrate with a Gaussian of sigma 10 pixels and a 256 x 256 pixels image the search space comprises minimally 25 x 25 different
4. at a given level e g Gosselin amp Schyns 2001 Schyns Bonnar amp Gosselin 2002 or be kept constant throughout the experiment 1 e Gibson Gosselin Wasserman amp Schyns 2002 Jentzsch Gosselin Schweinberger amp Schyns 2002 The technique will work so long as performance is between floor and ceiling The advantage of adjusting bubble numbers to equate performance is that Bubbles solutions are comparable In the FIE example we maintained categorization of sampled face information at about 75 correct by adjusting the number of bubbles using a gradient descent algorithm on a trial per trial basis The initial bubble number resulted from an informed guess i e between 50 and 60 bubbles for a first session and we let the gradient descent algorithm take over and adjust the bubble number to maintain performance at 75 4 The task At this stage the sampling procedure has been fully specified The final decision is that of the task We have explored a variety of face categorizations in humans and animals Gibson Gosselin Wasserman amp Schyns 2002 Gosselin amp Schyns 2001 Jentzsch Gosselin Schweinberger and Schyns 2002 O Donnell Gosselin amp Schyns 2002 Schyns Bonnar amp Gosselin 2002 Vinette amp Gosselin 2002 basic and subordinate categorizations of models of animals Schyns amp Gosselin 2002 and disciminations of an ambiguous painting by Dali Bonnar Gosselin amp Schyns 2002 In the FIE e
5. level However to the extent that this information is processed somewhere between the input and the response it has interesting implications for psychological processing To illustrate consider the high level task of face categorization and its underlying face features One interesting property of Bubbles is that the researcher can set up a search space that subsumes that of the assumed categorization features e g the eyes the nose and the mouth For example the 3D search space discussed earlier 2D image x spatial scales formally represents any face feature as a linear combination of information from the scales Consequently task specific face features can emerge in the Bubbles solution from the use of information at one or several of these scales see Gosselin amp Schyns 2001 Schyns et al 2002 for examples Thus while not applying the search directly to the features but in a space that represents these features Bubbles can reveal the subspace in which important features are actually represented It is in such spaces that Bubbles solutions tend to be most interesting To illustrate some would argue that the subspace in which important features are represented is in fact the information subspace to which attention is allocated Earlier we argued that this diagnostic subspace could be used as a diagnostic filtering function to reveal the effective stimulus Future research will need to characterize this filtering function en
6. locations to visit and the solution should converge within less than 500 trials If 5 scales are added as a third dimension the solution converges within about 5000 trials There is little doubt that significant learning occurs during these 5000 trials We are currently developing several heuristic searches to minimize these numbers e g Leclerc amp Gosselin 2002 A second shortcoming of Bubbles the relationship between the scale and geometry of the sample and the scale of the solution The scale and geometry of the sample impose biases on the search space If the scale is too small with respect to that of the solution important information will not be revealed within one sample and the same situation occurs when the geometry of the punch hole sampling unit does not fit that of the solution Remember that the Bubbles algorithm adaptively adjusts the number of bubbles to maintain the observer at a given performance criterion e g 75 correct Thus a higher sampling density leading to sample overlap can partially overcome the problems just discussed However there is always the possibility that observers will adopt strategies that enable performance to criterion but are nevertheless atypical of natural human categorization At a more theoretical level one could ask the question of What is the information revealed by the Bubbles algorithm The safe response is the information required to drive a response at a given performance
7. or a specific function e g a beach deck chair or texture detector In the absence of a principled method the specificity of the response e g to the beach is determined by contrast with responses from carefully chosen contrast categories e g roads cities mountains fields and so forth and informal hypotheses tested Unfortunately a dense correlative structure exists in the low level visual properties of category members e g luminance energy main directions of orientation spatial frequency composition and so forth only a small subset of which can be controlled with a finite number of carefully chosen contrast categories Consequently the specificity of the brain or behavioral response might be due to incidental input statistics and not to the category per se Schyns Jentzsch Johnson Schweinberger amp Gosselin 2002 In this article we present Bubbles Gosselin amp Schyns 2001 a method designed to solve the problem of finding the effective stimulus in the complex search spaces that are characteristic of visual categorization From the outset it is important to stress that the method can be scaled down and applied to simpler search spaces However originality of Bubbles is that it can handle search spaces that have so far proven to be elusive e g the information responsible for face recognition Gosselin amp Schyns 2001 Schyns Bonnar amp Gosselin 2002 scene recognition Nielsen Rainer Brucklacher amp
8. position of their main features 1 e the x y locations of the eyes nose mouth chin cheeks and forehead are roughly similar across face pictures This is necessary because the selected search space is not invariant to translation in the image plane Similarly we are assuming that the faces in the pictures will have the same size because the search space is not invariant to scale changes Note that these constraints on the search space are not constraints on the technique itself It is possible to set up translation invariant search spaces see Schyns amp Gosselin 2002 for a Fourier implementation and it is also possible to set up scale invariant search spaces However the experimental question the nature of face features that determine FIE suggested a search space where the location of face features would be known In our research we have used a variety of stimulus generation space ranging from the 2D image plane Gosselin amp Schyns 2001 Gibson Gosselin Wasserman amp Schyns 2002 Schyns Jentzsch Schweinberger Johnson and Gosselin 2002 O Donnell Gosselin amp Schyns 2002 a 3D plane identical to the one used here 2D image x spatial scales Bonnar Gosselin amp Schyns 2002 Gosselin amp Schyns 2001 Schyns Bonnar amp Gosselin 2002 a translation invariant 1D scale space Schyns amp Gosselin 2002 and a 3D space comprising the standard 2D image plane and time Vinette amp Gosselin 2002 From the
9. A User s Guide to Bubbles Fr d ric Gosselin Philippe G Schyns Correspondence concerning this article should be addressed to Fr d ric Gosselin D partement de psychologie Universit de Montr al C P 6128 Succursale centre ville Montr al Qu bec Canada H3C 3J7 or to Philippe Schyns Department of Psychology University of Glasgow 58 Hillhead Street Glasgow Scotland United Kingdom G12 8QB Electronic mail may be sent via the Internet to frederic gosselin umontreal ca or to philippe psy gla ac uk Abstract This article provides a user s guide to Bubbles a technique that reveals the information that drives a measurable response We illustrate the technique with a complete example the Face Inversion Effect and discuss the six basic decisions that must be made to set up a Bubbles experiment i e the stimulus set the generation space the bubbles the task the group of observers and the response We describe methods to analyze the data and provide practical advice for the researcher intending to use the technique A User s Guide to Bubbles The herring gull chick begs for food by pecking at its mother s beak In a seminal experiment Nobel prize winning ethologist Nikko Tinbergen and co worker Tinbergen amp Perdeck 1950 sought to discover the stimulus that maximized this response This enterprise led to the remarkable discovery of the super stimulus An artificial stimulus that evokes a stronger
10. Logothetis 2002 or the perception of complex figures Bonnar Gosselin amp Schyns 2002 The article is organized as a user s guide First we introduce a typical research problem never before addressed with Bubbles the Face Inversion Effect We then discuss the six main decisions that must be made to set up a Bubbles experiment discussing critical issues with examples from our own research The problem The Face Inversion Effect In a seminal article Yin 1969 reported that the recognition of face pictures was disproportionately affected by a 180 deg rotation in the image plane from the normal upright viewing condition This phenomenon is now commonly called the Face Inversion Effect FIE Since then the FIE has been replicated in multiple experimental situations e g Carey Diamond amp Woods 1980 Philips amp Rawles 1979 Scapinello amp Yarmey 1970 Carey amp Diamond 1977 Diamond amp Carey 1986 Freire et al 2000 Leder amp Bruce 2000 Scapinello amp Yarmey 1970 Tanaka amp Farah 1993 Valentine amp Bruce 1986 Yarmey 1971 There is now agreement amongst most face recognition researchers that the FIE does not arise from long term memory interferences but instead from a greater difficulty to perceptually encode inverted face information e g Farah et al 1998 Moscovitch Berhmann amp Winocur 1997 Phelps amp Roberts 1994 Searcy amp Bartlett 1996 Freire et al 2000 Therefore recent stud
11. V and Logothetis N K 2002 Studying the representation of natural images with the use of behavioural reverse correlation ECVP Perrett D I Mistlin A J amp Chitty A J 1987 Visual neurons responsive to faces Trends in Neuroscience 10 358 364 O Donnell C Schyns P G and Gosselin F 2002 The acquisition of facial expertise and how that mediates the information utilized to recognize the face Vision Sciences Oshawa I DeAngelis G C and Freeman R D 1990 Stereoscopic depth discrimination in the visual cortex neurons ideally suited as disparity detectors Science 249 4972 1037 1041 Phelps M T and Roberts W A 1994 Memory for pictures of upright and inverted primate faces in humans squirrel monkeys and pigeons Journal of Comparative Psychology 108 114 125 Philips R J and Rawles R E 1979 Recognition of upright and inverted faces a correlational study Perception 43 39 56 Rossion B and Gauthier I in press How does the brain process upright and inverted faces Behavioral and Cognitive Neuroscience Reviews Scapinello K F and Yarmey A D 1970 The role of familiarity and orientation in immediate and delayed recognition of pictorial stimuli Psychonomic Science 21 329 331 18 555 586 Schyns P G Bonnar L and Gosselin F 2002 Show me the features Understanding recognition from the use of visual information Psychological Science 13 402 409 Schyns P G
12. able bridges to be erected between cognition attention and perception References Ahumada A J Jr 1996 Perceptual classification images from Vernier acuity masked by noise Perception Supplement 26 18 Ahumada A J and Lovell J 1971 Stimulus features in signal detection Journal of the Acoustical Society of America 49 1751 1756 Bonnar L Gosselin F and Schyns P G 2002 Understanding Dali s Slave Market with the Disappearing Bust of Voltaire A case study in the scale information driving perception Perception 31 p 683 691 Carey S and Diamond R 1977 From piecemeal to configurational representation of faces Science 195 312 314 Carey S Diamond R and Woods B 1980 The development of race recognition a maturational component Developmental Psychology 16 257 269 Diamond R and Carey S 1986 Why Faces Are and Are Not Special An Effect of Expertise Journal of exp rimental Psychology General 115 107 117 Farah M J Wilson K D Drain M and Tanaka J N 1998 What is special about face perception Psychological Review 105 482 498 Freire A Lee K and Symons L A 2000 The face inversion effect as a deficit in the encoding of configural information Direct evidence Perception 29 159 170 Gibson B M Gosselin F Wasserman E A and Schyns P G 2002 Pigeons use specific and consistent features to discriminate human faces Submitted Gosselin F and Schyns
13. discussion above it should be clear that the number of dimensions making up the stimulus generation space is critical to the number of trials required to reach a stable Bubbles solution Generally speaking to visit each point of a search space there is a combinatorial explosion of steps with the increasing number of dimensions Note however that if the dimensions of the search can be collapsed for the analyses then the search space itself can be large For example one could decide that spatial scales are after all not that important for FIE collapse the data along this dimension and analyze feature use in the 2D image plane 3 The bubbles At this stage two important decisions have been made and the search can almost begin In the search information is sampled from the set up space and the next decision to make concerns the unit of sampling This unit depends on a number of factors including the stimuli the nature of the search space and the task to be performed To bring the observer away from ceiling relevant information must sometimes be sampled but sometimes not sampled The parameters of the sampling unit must be adjusted to ensure this modulation of performance A first parameter is the geometry of the sampling unit An information sample is effectively a cut in the search space Sampling unit with different punch hole geometries will change the information sampled and displayed to the observer Our research has mostly us
14. ed a Gaussian shaped geometries either in 2D Gibson Gosselin Wasserman amp Schyns 2002 Gosselin amp Schyns 2001 O Donnell Gosselin amp Schyns 2002 Nielsen Rainer Brucklacher amp Logothetis 2002 Schyns Jentzsch Schweinberger Johnson amp Gosselin 2002 or in 3D Bonnar Gosselin amp Schyns 2002 Gosselin amp Schyns 2001 Schyns Bonnar amp Gosselin 2002 This choice was motivated by two main factors Gaussians functions produce a smooth cut producing a sample that does not introduce hard edge artifacts without orientation biases of the sampled information i e a Gaussian is circularly symmetric A different search space could require geometries other than Gaussians For example if orientation information was searched as an independent dimension the sampling unit would need to introduce orientation biases For example a Gabor function could be designed to sample information at several orientations e g 0 45 90 and 135 deg More abstract geometries can also be used when the search space is itself abstract For example in Schyns and Gosselin 2002 the bubble was a dot sampling Fourier coefficients in a Fourier Transform search space Another important parameter of the sampling unit is its resolution The resolution is largely determined by considering the scale of the stimulus and the expected resolution of the relevant information for the task at hand To illustrate we know that the eyes the mouth a
15. eep all the sampled information of each trial to be able to do more detailed analysis such as the conjunctive use of information see Schyns et al 2002 constant viewing distance Stimuli subtended 5 72 x 5 72 deg of visual angle on the screen On average observers required an average of 46 and 126 bubbles to reach the 75 performance criterion in upright and inverted conditions respectively The number of bubbles between 197 and 30 bubbles depending on observers and condition and average performance between 86 and 75 did vary across the six experimental sessions to stabilize in the last session In this session observers in upright and inverted respectively required an average of 30 and 65 bubbles for performance levels 75 and 76 The comparatively higher number of bubbles in the inverted condition suggests a higher requirement of visual information suggesting a more difficult inverted condition diagnosing a FIE We can now turn to a comparison of the required information in each condition to attain the same level of performance To this end we first perform a linear multiple regression Practically for each spatial scale we computed two independent sums we added together all the information samples leading to correct responses in one sum and all the information samples leading to incorrect responses in another sum At each spatial scale we then subtracted these two sums to construct an image that discriminates the information
16. f Neuroscience 11 1239 1255 Yarmey A D 1971 Recognition memory for familiar public faces effects of orientation and delay Psychonomic Science 24 6 286 288 Yin R K 1969 Looking at upside down faces Journal of experimental Psychology 81 41 145 Figure captions Figure 1 Pictures in b decomposes a in five scales c illustrate the bubbles applied to each scale d are the revealed information of b by the bubbles of c Note that on this particular trial there is no revealed information at the fifth scale By integrating the pictures in d we obtain e a sample stimulus Gosselin amp Schyns 2001 Schyns Bonnar amp Gosselin 2002 Picture f is a sampled in the image plane Gosselin amp Schyns 2001 Gibson Gosselin Wasserman amp Schyns 2002 Jentzsch Gosselin Schweinberger amp Schyns 2002 O Donnell Schyns amp Gosselin 2002 Picture g is a 3D shape model of a dog sampled in phase space Schyns amp Gosselin 2002 Finally picture h is the ambiguous area of a Dali painting sampled in the same generation space as e Bonnar Gosselin amp Schyns 2002 Figure 2 The first row of this figure applies to the upright condition It gives the five classification images at each scale from finest to coarsest The red areas revealing a face are significantly above chance p lt 05 The rightmost picture is the effective face The second and third rows are the same the first
17. ies have examined more closely the encoding differences that occur when experiencing an upright or an inverted face However the specification of these differences has so far remained largely unknown Rossion amp Gauthier in press To address the FIE with Bubbles we need to make six basic decisions 1 what is the stimulus set 2 in which space will stimuli be generated 3 what is the bubble 4 what is the observer s task 5 what are the observer s possible responses 6 is the analysis per observer or per group of observers In resolving all of these we will set up a search space and vary the parameters of the independent variables upright and inverted face information that determine the measurable dependent variable the observer s response The Bubbles solution should specify the difference between the information subspaces driving the processing of upright and inverted faces 1 Stimulus set In a Bubbles experiment the stimulus set is crucial because it critically bounds what will be tested Here we used a total of 10 greyscale faces 5 males 5 females each one of which displaying 3 different expressions neutral angry and happy Hairstyle was normalized and so were global orientation and the location of the light source Stimuli could be upright or inverted but when inverted we flipped the image so as to keep the light source to the right of the face Generally speaking the larger the stimuli set the be
18. in the second row of Figure 2 As the number of bubbles was unequal in upright and inverted we computed the discrimination image on the normalized Z score images Here the first three scales contain diagnostic information The eyes the nose and the right part of the mouth are the most important local features that explain the difference between inverted and upright information use in face processing Discussion We started this problem with a generic methodological question Given a dependent measurable response behavioral electrophysiological or other of an organism how can we determine the optimal subset of parameters from the independent variables that determine the response With simple stimuli e g Gabor functions or sinewaves this is not much of a challenge because they can only vary along a few degrees of freedom limiting the complexity of the task With the stimuli that are typical of realistic face object and scene categorizations the task had proven so far intractable Bubbles is a technique that can resolve the credit assignment problem of attributing the determinants of a response to the parametric subspace of a carefully specified information search space Using the Face Inversion Effect as an example we reviewed the six basic decisions to be made to set up a Bubbles experiment In order deciding the stimulus set the dimensions within which to search for information the geometry of the unit to sample information the task
19. leading to correct and incorrect responses see the first row of Figure 2 for the discrimination images at each scale If all regions of the search space were equally effective at determining the response the image would be a uniform gray To pull out the most effective region we computed Z scores for each discrimination image and indicated in red the regions that are 1 65 std away from their mean corresponding to a p lt 05 These regions circumscribe the subspace driving upright and inverted face classification responses If we project the original face in Figure 1 a into this diagnostic subspace we obtain the effective stimuli displayed the extreme right of the rows in Figure 2 Technically each effective stimulus is obtained by multiplying the face information at each scale in Figure 1b with the corresponding thresholded coefficients in the rows of Figure 2 For upright faces the eyes are the most important local features The only scales with diagnostic information are the second and the third This is consistent with the results of Gosselin and Schyns 2001 and Schyns Bonnar and Gosselin 2002 However these experiments did not include an inverted condition Observers saw two not three expressions and they were less familiar with the faces i e 1 000 rather than 3 900 trials For inverted faces observers do not seem to rely on any specific features to perform the task They all seem equally good or bad This is reflected
20. n temporal cortex respond to complex object properties Kobatabe amp Tanaka 1994 such as orientation in depth Perrett Mistlin amp Chitty 1987 object similarities Vogels 1999 and the information responsible for visual object categorization Sigala amp Logothetis 2002 Vogels 1999 However even though IT cells are just a few synaptic connections away from primary visual cortex their optimal stimuli are hidden in a much more complex search space the physical world With its many faces objects and scenes this space does not comprise just the few degrees of freedom required to represent the little spots of light positioned within the visual field or the moving orientated bars Instead IT cells respond to structured information that varies in 2D retinal position 2D rotation in the image plane 3D rotation in depth illumination articulation and so forth Amongst these multiple degrees of freedom different subspaces of parameters represent the effective stimuli of IT cells The challenge is to understand what these subspaces are And it is still one of the greatest methodological challenges in Cognitive Science when dealing with complex visual stimuli how can a brain event an ERP response or an fMRI measurement or a human behavior e g a categorization response be attributed to a specific object category e g a beach scene a specific object e g a deck chair a specific feature e g the texture of the deck chair
21. n the bins e g subtracting the wrong response from the correct response bin divide the correct response bin by the sum of the correct and incorrect response bins The result of this operation is usually transformed into Z scores and thresholded e g at 1 65 p lt 05 or 2 33 p lt 01 The outcome of this analysis is the product of Bubbles revealing the effective subspace of input information In the visual domain this outcome is a filtering function that can be applied on the original stimulus to reveal the information that drives the task In the FIE example three observers learned to criterion perfect identification of all faces twice in a row the name attached to each of the 10 faces from printed pictures with corresponding name at the bottom During the experiment observers had to determine the identity of each sparse face from 10 possibilities The experiment comprised six sessions of 780 trials i e 13 presentations of the 30 faces upright and inverted but we only used the data of the last five sessions for a total of 3900 trials per subject when observers were really familiar with the faces and experimental procedure In a trial one sparse face computed as described earlier appeared on the screen either upright or inverted To respond observers pressed labeled computer keyboard keys self paced and with correct vs incorrect feedback A chin rest maintained subjects at a 100 cm However it is also useful to k
22. nd the nose are the most useful features to make face decisions It would therefore be advisable that the resolution of the sampling unit in a FIE task be lower i e smaller than the resolution of the important features A very low resolution sampling unit e g the pixel of an image provides a precise sample of the search space but many trials are required to converge on a solution of the search Clearly the resolution of the sampling unit must be chosen with a priori considerations of the expected scale of the solution For the reasons just discussed the bubble of our FIE example has a Gaussian geometry The scale of the bubble was chosen to sample three cycles per face i e stds of 13 27 54 1 08 and 2 15 deg of visual angle from fine to coarse scales see Figure 1 On any given trial information is sampled from the search space by a number of bubbles The sampling is typically performed randomly and is thus non biased Figure 1 a e illustrates the sampling procedure In b the face shown in a is decomposed into five independent scales In c bubbles with a Gaussian geometry sample the information space at random locations overlap is permitted In d the bubbles c are applied to the appropriate scales in b Finally in e the pictures of d are added together to produce a sub sample of the face information in a One important point about bubbles their number It can either be adjusted on line to maintain performance
23. one except that they apply to the inverted condition and the difference between the upright and inverted conditions respectively Figure 1 Figure 2 Finest scale Coarsest scale Effective faces 9109 puta Upright Difference
24. onnar Gosselin amp Schyns 2002 and response latencies Schyns et al 2002 In addition we also used preferential looking Humphreys Gosselin Schyns Kaufman amp Johnson 2002 and N170 amplitudes Schyns et al 2002 Other responses could be the firing rate of single cells fMRI galvanic skin response pleismograph eye movements and so forth To the extent that Bubbles is essentially an empirical tool it is useful to record as many different responses as possible e g correct incorrect latencies and N170 in a face recognition experiment It is difficult to predict before the experiment how responses will correlate with the parameters of the search space Analyses Now that the search has been run the data are collected and the analyses can be performed Remember that the goal of the search is to isolate a subspace of information that determines the measured response s Technically a multiple linear regression on the samples explanatory variable and the responses predictive variable provides this solution This reduces to summing all the bubble masks in different response bins where the number of responses is a function of the nature of the response itself For example two bins are sufficient to tease apart correct and correct responses but more bins are necessary to cover the range of electric activity or cell firing rate of a brain response To reveal the most important information we can perform a linear operation o
25. response than the original natural stimulus For example a white stick with three red annuli moving up and down produces a stronger pecking response than the head of the herring gull mother s At an abstract level the search for the super stimulus can be framed as a generic search problem Given a measurable dependent variable e g the pecking rate response the problem is to find the specific parameters of the independent variable s e g the characteristics of the mother s head that optimize the dependent variable Obviously this approach is not limited to ethology An approach similar in spirit is that of Nobel prize winners Hubel and Wiesel who searched for the stimulus that optimizes the response of cells in the primary visual cortex see Hubel 1988 for a review Much to their surprise they discovered that small spots of light which are so effective in the retina and Lateral Geniculate Nucleus LGN were much less effective in visual cortex Instead simple cells in primary visual cortex responded optimally to inputs with linear properties such as a line with a specific width and orientation in the plane At the next level of cortical integration optimal inputs become more complicated For example complex cells tend to respond better to a stimulus with a critical orientation but also a characteristic speed and direction adding to the width and orientation search space a third dimension Further up the integration ladder cells i
26. tter the Bubbles solution should be A large stimulus set will tend to prevent observers from adopting strategies atypical of natural processing In the FIE example the stimulus set restricts the search space for differences in upright and inverted face encodings to a few males and females with a limited set of expressions in highly restricted conditions of presentation only one light source two poses and static images Although this also applies in most face recognition experiments it is important to point out that the Bubbles solution will be tied to these limitations In our research we have already used faces in other experiments with human participants Gosselin amp Schyns 2001 Schyns Gosselin amp Bonnar 2002 Schyns Jentzsch Schweinberger Johnson amp Gosselin 2002 but also in animals experiments Gibson Wasserman Gosselin amp Schyns 2002 Other stimuli used ranged from 3D models of Gouraud shaded animals Schyns amp Gosselin 2002 to a painting of Dali Bonnar Gosselin amp Schyns 2002 Other researchers have also applied Bubbles to natural scenes Nielsen Rainer Brucklacher amp Logothetis 2002 Although these applications only involved visual stimuli the technique should straightforwardly generalize to auditory and tactile stimulus sets or to cross modal combinations of these 2 Stimulus Generation Space The choice of a proper stimulus generation space is one of the most important decisions when setting
27. up a Bubbles experiment Remember that we are searching for the parameters of the independent variables upright and inverted face information that determine the FIE Each independent variable considered constitutes one independent dimension whose parametric values will be searched To illustrate our face stimuli are 2D pictures The axes of the 2D plane could be searched to find the x y coordinates of face information that determine upright vs inverted performance The stimulus generation space would then be two dimensional and the solution would be a subset of the plane However there is evidence that early vision does analyze the input at multiple spatial scales or spatial frequencies see de Valois amp de Valois 1990 for a review and that mechanisms of face recognition rely on this input see Morrison amp Schyns 2001 for areview Thus a better space to search for the determinants of FIE could include a third dimension of spatial scales Specifically we segmented the third dimension into 5 independent bands of fine to coarse spatial frequencies of one octave each with cutoffs at 90 45 22 5 11 25 5 62 and 2 81 cycles per face The solution subspace becomes an interaction between the 2 dimensions of face feature location and the third dimension of spatial scale In setting up this search space we are making a number of assumptions that are worth pointing out We are assuming that the face pictures are normalized for the
28. xample observers will identify the faces in the upright and inverted conditions 5 Observers Depending on the objectives of the research different types of observers can interact with the Bubbles algorithm For example we have applied the technique to groups of human observers Gosselin amp Schyns 2001 Schyns Bonnar amp Gosselin 2002 Bonnar Gosselin amp Schyns 2002 individual observers to track down effects of expertise acquisition Jentzsch Gosselin Schweinberger and Schyns 2002 Gosselin amp Vinette 2002 O Donnell Gosselin amp Schyns 2002 infants to tackle issues in development Humphreys Gosselin Schyns Kaufman amp Johnson 2002 pigeons Gibson Gosselin Wasserman amp Schyns 2002 and ideal observers which are models providing a benchmark of the information available in a task Gosselin amp Schyns 2001 We have several on going research projects involving brain damaged patients patients suffering from prosopagnosia and hemi neglect 6 Response The response is an interesting parameter of a Bubbles experiment because the technique is in principle sensitive to any measurable dependent variable Here observers pressed labeled keys corresponding to the names of 10 individuals We have used such key press responses to derive correct and incorrect responses Gibson et al 2002 Gosselin amp Schyns 2001 Gosselin amp Vinette 2002 O Donnell Gosselin amp Schyns 2002 Schyns et al 2002 B
Download Pdf Manuals
Related Search