Perceiving and Re-Perceiving Objects Anne Treisman University of California, Berkeley ES A number of experiments exploring priming effects and ‘automatization in the perception of novel objects are de- scribed, and a framework for understanding the benefits ‘and costs of re-perceiving previously seen objects is pro- posed. The suggestion is that perceiving an object creates @ temporary representation in an object file that collects, integrates, and updates information about its current characteristics. The contents of an object file may be stored ‘as an object token and retrieved next time the object ap- pears. This facilitates its re-perception when all of the attributes match and may impair it if some are changed. Thus, the world molds our minds to capitalize on earlier experiences but at the same time leaves us able readily to detect and represent any novel or unexpected objects and events. When you look at the object in Figure 1, it has some effects on your brain. First you can see it. Second you try, and some of you fail, to identify it by matching it to traces of relevant experiences stored in memory. So you classify it as an unknown object about which I am trying to make some points. When I take it away, it eaves some traces in memory. Some of you may be able to remember it in schematic form tomorrow or later, just as some (not all) of you can visualize the Taj Mahal. In addition, and perhaps independently of these consciously retrievable ‘memories, the single exposure to this visual object will have effects on how you perceive it again, and perhaps (on how you perceive other stimuli. Some of these effects are short lived; others are quite persistent. Tam interested in understanding the changes that happen when we perceive and re-perceive or remember an object. 'am, of course, not alone in asking these ques- tions. It will be obvious that the ideas owe a lot to other people, particularly to Kirsner, Dunn, and Standen (1987), Hinteman (1976, 1986, 1988), Jacoby and Brooks (1984), Logan (1988), Roediger (1990), Schacter (1987; Schacter, Cooper, & Delaney, 1990), and Tulving (1983; Tulving & Schacter, 1990). I will use a selection of ideas that are similar to theirs in an attempt to link perceiving and re-perceiving across different intervals and in different tasks oF contexts Before I get to my story, I need to introduce a con- ceptual distinction between two different ways we might represent objects, both during perception and in visual memory. Some years ago, Daniel Kabneman and I sug- ested that object perception comprises two different Processes, corresponding roughly to seeing and to iden- tifying (Kahneman & Treisman, 1984). In identifying an object, we match it to a set of stored descriptions that summarize our knowledge of what things look like—what ‘makes something a rabbit, or a teapot, or the Taj Mahal, Many psychologists have attributed perceptual identifi- cation to a recognition network of nodes and connections, whether at the symbolic level or at the subsymbolic level, of connectionist nets. I will call these stored representa- tions object types. Seeing is often equated with activity in the same set ‘of nodes that signal the identity or category of the object. However, Kahneman and I argued that some other form of representation is needed in addition, to mediate per- ception of any particular instance of an object in all its, often arbitrary particularity—this yellow teapot, on this, table, with this cracked lid. I wil cali these representations, object rokens. When we form an object token for a newly perceived object, we also try to match its content to our stored dictionary of object types. We may or may not, succeed-—you presumably failed for the object in Figure 1. But we do see the object anyway, perfectly clearly if visibility is good, because of the object token that specifies its features and structure, even without a label telling us what it is, In this article { will try to fill out, both from exper- imental results and from further speculation, how I believe that object tokens are formed and later retrieved or re- perceived. I will describe results from several tasks that ‘measure visual memory indirectly, as changes in latency with immediate repetition or with long-term practice. ‘These are implicit memory effects, and I will ask whether they are best explained in terms of object tokens or of changes in nodes or connections in a recognition network. ‘The aim is to map out a framework for describing priming, Editor's note. Articles based on APA award addesss that appear inthe “American Pyohologist ae scholarly articles by distinguished contributors {the fied. As such, they are given special consideration inthe American Psychologists editorial selection proces. This article was originally presented as a Distinguished Scientific ‘Contribution award atthe 99th Annual Convention of the American Psychological Association in San Francisco “Auhor’s Note. Tis research vas supported by US. Ar Force Ofce of Scientie Research and Olfice of Naval Research, Grant 90-0370, ‘The manuscript is submitted for publication with the understanding that the U.S. Government is authorized to reproduce and distribute reprints for government purposes, notwithstanding any copyright no tation thereon. Preparation ofthe article was also supported by the Russell Sage Foundation. Tam grateful to Daniel Kabneman, Nency Kanvisher, eona ‘Khurana, Morris Moscovitch, and Gail Misen for thie helpful com ‘ments, and to Epiram Cohen for programming the displays (Correspondence concerning this article shoul be addrese to Anne ‘Treisman, Department of Psychology, Universit of California, Berke cA 94726, 862 July 1992 + American PsychologistNN el Figure 1 Nonsense Shape With No Prestored Perceptual Representation ———_— effects of visual memory at different delays and through, different re-perception tasks. We may then be able to make inferences about the memory representations responsible for the observed effects. Let me anticipate the conclusions, in order to guide you through a varied sample of experiments. (Table | also gives an overview.) As Jacoby, Logan, and others we done, J will try to account for many of the implicit memory effects we have observed in terms of a large set, of object tokens, one for each experience of a perceived object. Tokens are set up in a single trial and can affect, performance after substantial periods of time (at least a week in one of our studies). They may be retrieved the next time the same object appears and may affect its re- perception, either speeding it up or interfering if the object, ‘was previously irrelevant or ifsome features conflict with those of the present stimulus. If the object was attended, the token specifies its incidental properties—as seen on. ‘a particular occasion—as well as the properties that were relevant to the task. This article describes a sample of preliminary experiments that suggested these conclusions. Some of the conclusions are quite tentative, but we can, use them as straw men to shoot at when further evidence is collected. Object Files and Visual Immediate Memory ‘The first experiments I will describe explored the initial process of setting up object tokens in a perceptual task that also illustrates why we felt the need to propose their existence. The aim was to account for the perception of dynamic scenes containing moving, changing objects. Although in our daily lives we seldom watch pumpkins turn into carriages or frogs into princes, we do face in a milder form the problems of maintaining the perceived unity and continuity ofa single object as its shape changes through motion, perhaps also revealing new parts or properties. To account for the integration over time of information about any particular object, Kahneman and introduced the metaphor of object les by analogy with the file that the police might use to collect information about a particular individual or incident, An object file isa temporary structure within which information about a particular, presently visible object is collected, inter- related, and updated, as the object moves or changes its appearance. It is the collecting box in which an object token is formed. Note that the functional instantiation of an object fle need not be activity ina particular group of cells; it could also be a temporary synchronization of activity in groups of cells located in different areas of the brain (Eckhor et al., 1988; Gray, Konig, Engel, & Singer, 1989). It must, however, have the potential of leaving a ‘more permanent trace, recording the object token beyond the time for which the object is perceptually available, We designed a series of experiments to demonstrate the existence of these object-specific representations (Kahneman, Treisman, & Gibbs, 1992). Most of our ex- periments used a letter-naming task in which two suc- cessive displays of letters are linked by the motion of frames presented around the letters. In a typical experi- ‘ment, two empty boxes appear. Two letters are briefly flashed in the boxes (see Figure 2), which then move, empty, to two new locations. Finally a third letter appears in one of the two, to be named as quickly as possible, We find that naming is faster when the same letter had pre- viously appeared in the same frame than when a new letter is presented. Surprisingly, however, naming speed is completely unaffected (relative to naming a new letter) if the same letter had previously appeared in the other frame. Thus, conventional priming is absent. Under the particular conditions of this experiment, retrieval and ‘matching seem to be determined entirely by the object file, addressed by its spatio-temporal coordinates and not by its content. We suggest that when a new stimulus ap- pears, the visual system tries to associate it with its recent history. It is matched to the most likely past object, con- strained by the spatio-temporal continuity of the object file. If the present stimulus matches the past content of the object file, perception is faster than if a change has occurred. Our experiments demonstrate a brief visual memory spanning the interval in which the objects move, empty, to their new locations (Kahneman et al., 1992). It is in- teresting that we found no effect of the temporal interval or of the duration of the initial letters over the range we tested, The matching benefit was about equal for initial presentations of 30 milliseconds and 1 second and for motion that took 700 milliseconds or less than 100 mil- liseconds, making it unlikely that the effects depend on iconic memory. ‘When we used words rather than letters in a similar experiment, we did find some priming that was not object specific, indicating retrieval by content. However, this was in addition to object-specific priming of about the same ‘magnitude as is obtained with letters. Words are much ‘more distinctive than letters, and unlike the letters, they July 1992 + American Psychologist 863EEE Table 1 Evidence for Object Tokens in Re-Perception “Tne of test Ts Observations Inerancee During original Name letter or conjoin features in Object-specifi reviewing Object files, addressed by spatio- perceptual final display of sequence with and feature ‘temporal coordinates, are experience. ‘moving, changing objects (with integration. Created to collect and integrate Kahneman and Gibbs) Immediate repetition Match one of two superimposed of fist shapes. Look for negative ‘experience. priming when unattended shape becomes attended one (with DeSeneppe Delayed repetition of Draw briefly presented masked first experience. pattern. Compare with explicit ‘memory in recognition tasks. (with Musen) Transfer to other tasks atter Search for line patterns. Measure transfer to other perceptual tasks with same targets (with Vietra), ‘Search for conjunction or for feature targets; look at trial-to- trial repetition effects (with Hayes) ‘Short-term repetition riming, Search for conjunction or feature targets presented with location or irrelevant feature ‘consistencies (with Hayes). Later sessions of extended practice. Information about each attended object. (Object representations are ‘established in a single tral, even Equal or greater ‘Negative priming from novel shape {or the irrelevant object in a ‘compared with familar superimposed pair. Greater accuracy for pattern previously seen in a single study tral. Little effect of repeated study or delay. Dissociation between implicit and ‘explicit memory. Large effects of practice within the search task but litle or no transfer to other tasks. ‘Object representations established in one tral can last for a Considerable time (up to one week) and are only slightly strengthened by four more exposures. Object tokens also contain information about the response: that was made. Learing in extended practice includes selection of appropriate responses. Object tokens depend on focused attention. They contain irrelevant as well as relevant features. Retrieval from previous trial is cued primarily by location. tn conjunction search benefit of match to previous trial when target, location, and irelevant features are repeated. in feature search, much less specicty Costs and benefits of location and feature consistencies in Conjunction search: much reduced in search for feature targets. Object tokens for attended stimu accumulate across trials; they are retrieved by content, but also contain correlated Information about focation and irrelevant features. ES ‘were not repeated across trials in our experiment, Dif. ferences in the activation of types in a semantic network would therefore be much greater between primed and ‘unprimed words than between primed and unprimed let- ters. We suggest that type activation and token matching make independent contributions to the preview benefit, that we observed with words. Object files are involved in feature integration as well as in speeded identification of a stimulus. In a similar paradigm, we presented four frames, flashed up a hori- zontal line in two of them, and a vertical line in the other ‘two (Figure 3), moved the empty boxes to new locations, and finally, flashed another horizontal or vertical line in, ‘each box in its new location. (The result is summarized in Treisman, 1988.) Subjects were asked to detect a plus sign that on one half of the trials replaced a line in one of the four boxes in the final display. They were slightly slower to say there was 0 plus when the two successive lines within each frame differed in orientation, and so would have made a plus if they had been temporally and, spatially superimposed, compared with when the two successive lines within any given object were both hori- zontal or both vertical. The two successive displays of lines were apparently integrated within the frames that, linked them, making it harder to distinguish illusory pluses from the real target plus. Mordkoff (1991) recently. obtained similar results in a task that required integration of color and shape. Subjects were more likely to sce il- lusory recombinations when the features were seen as ‘belonging to the same moving object in different positions 864 July 1992 + American Psychologist