Imagery and reflective thinking

My central argument rests on three claims. First, and probably most controversially, I argue that all reflective thinking is accomplished by mental imagery. I am not restricting the term “mental imagery” to refer only to visual images (i.e., something like “mental pictures”). Rather, mental images can occur in each sensory system: a mental image is the same sort of representation that is produced during the first phases of perception, but is created from stored representations in memory (not from on-line sensory input). For example, in visual perception these representations arise in topographically organized cortex; these patterns of neural activation preserve (roughly) the spatial structure of the image laid out on the retina (for a detailed review, see Kosslyn, Thompson, & Ganis, in press). Image representations give rise to conscious experience (although perhaps only indirectly, by sending outputs to frontal lobe areas farther downstream; cf., Crick & Koch, 1995). According to this view, images are not representations of raw sensory information, but rather are organized to make explicit and accessible perceptible properties of objects.

The notion that mental images are a specific type of internal representation (i.e., a means of storing information) is important because it allows us to study imagery objectively. Image representations have specific properties that affect information processing, and studies of response time have documented properties of imagery. For example, if an object is visualized at a tiny size, its features will be obscured; thus, those properties will require more time to “see” than those of an object visualized at a larger size (where “size” refers to visual angle; Kosslyn, 1980, 1994). Moreover, recent research on the neural bases of imagery provides another method whereby imagery can be studied objectively (e.g., for a review, see Kosslyn, Ganis, & Thompson, 2001). Researchers can observe specific deficits that follow brain damage, and can also observe which areas of the brain are selectively activated during imagery. Indeed, my colleagues and I have shown that the amount of activation in key areas the brain used during imagery nicely predicts performance on imagery tasks (Kosslyn et al., 2004). Thus, although I often use introspection to provide examples of imagery, these are not primary data. Introspection, the sense of “inner perception” (e.g., “seeing with the mind's eye”), is a hallmark that specific types of representations are present; but it is not itself evidence for the use of imagery.

The view that mental images are representations like those produced during the first phases of perception leads me to emphasize that we have many different sorts of mental images, most notably visual (representing shape and properties of shape, such as color and texture), spatial (representing size, orientation, and location), auditory (representing pitch and loudness), kinesthetic (representing movement), and tactile (representing sensations on the skin). Imagery is not restricted to the visual modality. For example, to get a sense of spatial imagery, close your eyes and think about the locations of the windows and doors of the room you are sitting in, relative to where you now sit, or think about the direction of the nearest airport. One can have “spatial” images even in the absence of images of objects.

The distinctions between the different sorts of imagery arise from the operation of distinct neural systems. For example, neuropsychological studies have documented that the visual system is organized into two major neuroanatomic pathways (e.g., Ungerleider & Mishkin, 1982), each of which can support distinct types of imagery. The so-called ventral stream, running from the occipital lobes at the back of the brain to the lobes under the temples, processes representations of shapes and colors. In contrast, the so-called dorsal system, running from the occipital lobes to the posterior portion of the parietal lobes (at the top rear of the head), processes spatial relations.

What is the evidence that these separate “object” versus “spatial” visual systems are also used in imagery? First, patients who have had brain damage can sometimes visualize images of shapes but cannot imagine spatial relations, or vice versa. Specifically, those with damage to the ventral stream can answer questions that require spatial imagery (e.g., “In which hand does the Stature of Liberty hold the torch?”) but not those that require object imagery (e.g., “Did George Washington have a beard?”), and vice versa for patients who had damage to the dorsal stream (Levine, Warach, & Farah, 1985). Second, the two neuroanatomical systems are selectively activated, as assessed in neuroimaging studies, by the different sorts of processing (Thompson & Kosslyn, 2000).

Similarly, we can have auditory images (including of words being spoken), images of kinesthetic feedback of the sort we experience when we move (Jeannerod, 1994, 1995), and so forth (Kosslyn et al., 2001).

To get a sense for my claim that imagery (of one sort or another) is an essential ingredient of reflective thinking, reflect for a moment on how you think about taking a new route when driving or prepare to have a difficult conversation with a friend: do you find yourself visualizing the route, and imagining yourself talking to your friend (perhaps “hearing in your mind's ear” parts of the conversation), running through various scenarios? Mental images are the vehicle of reflective thinking in part because they are a way to find out what we have stored in LTM. For instance, just imagining asking your boss for a raise can tell you a lot about what you believe might happen (and can even make your skin sweat and heart beat faster, if you believe that conversation will become a confrontation; Lang, Bradley, & Cuthbert, 1990; Vrana & Lang, 1990).

The idea that we use imagery to make explicit what we have stored in LTM is based on the notion that we constantly encode and store perceptual information incidentally, without effort (e.g., Valiquette, McNamara, & Smith, 2003). In contrast, we must make an effort to encode and store a description. Thus, in many situations we will not have described information that we did encode and store perceptually. Hence, to recall and use that information, we need to rely on these perceptual representations and use imagery (as in my Mickey Mouse example).

In addition, we sometimes will use imagery even if we have stored descriptions, and automatic processing in principle could perform a task. For example, please decide whether a mouse has a tail. Even though you know this information, you may still use an image. Why? In earlier work (Kosslyn et al., 1977; summarized and extended in Kosslyn, 1980), my colleagues and I showed that searching descriptive knowledge and using imagery occur in parallel, and constitute a kind of race: whichever process is faster “wins” and produces the response. Thus, if automatic processing is relatively slow, you may not end up actually using it.

In summary, reflective thinking occurs when automatic processing is either not sufficient or fast enough to perform a task; reflective thinking involves working memory; and working memory relies on representations that underlie conscious awareness: mental images. These ideas imply that reflective thinking is important to the extent that the world is not predictable, that new and unexpected events occur. During reflective thinking we pause to examine the consequences of various actions or events, and such processing helps us make novel decisions. This claim is, in fact, a corollary of the definition of reflective thinking: if reflective thinking occurs when one cannot respond automatically, on the basis of information stored in LTM, then it follows that when automatic responses do not work, we engage in reflective thinking.

Stored knowledge

My second claim is that LTM is a vast network of stored units of information, and that searching through such memories is best conceptualized in terms of the process of constraint satisfaction. Let me begin to explain this idea with an analogy: think about how you guess the “target” when you play the game charades: the presenter might get down on all fours, indicating that she has in mind something that has four legs; she might crawl, indicating that it was an animal (not a table or chair or other four-legged object); might hold up fingers at the top sides of her head to indicate pointed ears; might mimic lapping up milk, and so on. Each hint would activate the corresponding stored representation of a characteristic, and when enough representations of characteristics are activated only a single concept is consistent with them all.

In my view processing of information stored in LTM is exceedingly unintelligent: such processing requires only that activating one representation in turn activates or inhibits (as specified by the nature of the association) other representations to different degrees. For example, the representation of the animal we call a cat is associated in LTM not only with a name but also with the categories to which it belongs, common contexts in which it is found, a stored image of the appearance of a prototypical cat (and possibly images of specific individual cats), a description of its shape and other characteristics, and so on. When you hear the word “cat” the representation of this word is activated, which in turn, activates the representation of the animal itself; and this representation in turn activates the representations of associated concepts, such as “milk,” and at the same time inhibits irrelevant representations, such as of the concept “fork.”

The central idea behind this notion of constraint satisfaction is that the input activates specific representations, and if there are enough cues, only a single stored representation of a concept may be consistent with all of the activated representations at once. One last analogy: say you have a watermelon seed and are squeezing it between your thumb and forefinger. If you keep squeezing harder and harder, at some point the seed will fly out. This is another example of constraint satisfaction: the shape of the seed, the distribution of moisture over it, the precise position of the seed within your grip, and the forces on it are all factors that act as constraints. No single constraint dictates what will happen when the pressure is strong enough, but all of them taken together force the seed on a specific trajectory. There will be only one way that all of the constraints can be combined simultaneously. The shape of the seed, moisture, and other constraints are like the information activated in your memory by cues.

This process of constraint satisfaction operates both to access specific descriptive memories and to trigger the processes that generate an image. Generating an image requires locating stored modality-specific representations and activating them, which allows the perceptual qualities of the object or event to be reconstructed in working memory. The simultaneous “race” between looking up stored descriptions and generating an image is a straightforward byproduct of this sort of constraint satisfaction processing.

In short, this sort of constraint satisfaction occurs automatically and unconsciously, and can simply activate a descriptive representation that is explicitly stored or can produce an image. The two processes occur at the same time, and typically we stop doing one once the other has allowed us to access the sought information; thus, we do not need to use imagery if we can quickly locate a stored descriptive representation.

However, reflective thinking can itself provide input to the process of constraint satisfaction, and thus not all of the overall process of constraint satisfaction need rely solely on unconscious processes. In this case, reflective thinking is not simply a product of unconscious processing, but provides input to such processing. This observation leads to my final claim.

Reflective thinking changes automatic processing

My third claim is that reflective thinking is useful not simply as a way to recall information or make novel decisions, but also as a way to reprogram LTM, so we will not need to “think about” a situation in subsequent encounters. During reflective thinking, the inferences or decisions one has drawn on the basis of imagery are then “reflected back” to LTM, changing it. For example, if you are asked to figure out a good route home from a new location, the first time you do so you must use visual and spatial imagery to reflect. If you are asked to do this repeatedly, eventually you will not need to use imagery; you will have stored that information descriptively in LTM. Moreover, after you have used that information a few times, it will be so easy to access (because the associations to the retrieval cues are strengthened) that you will access the description before an image is completed. Thus, you will know how to make your way along that route without needing to reflect before setting out.

Here is another example. Please indulge me and answer the question again. What shape are Mickey Mouse's ears? Please bear with me and answer one more time. What shape are Mickey Mouse's ears? By the last time, you may incidentally have had an image, but if you are like most people you did not actually need to use this image to answer the question. Rather, you simply stored the description of the shape (“round” or “circular”) in LTM; and, because you had practiced this response, you could access this descriptive information more quickly than you could generate an image. Thus, you could answer automatically on the basis of this descriptive information without needing to reflect.

We are not conscious of information in LTM, and the idea that mental imagery has a special relation to unconscious knowledge and emotion is not new. Sigmund Freud thought of images as akin to internal smoke signals, spiraling up from fires smoldering beneath the surface. However, I am saying something very different: imagery is not a spin-off of the unconscious, acting as a symbol that signals deeper mysteries, as Freud believed. Rather, imagery is a two-way street between the conscious and unconscious: it not only taps into unconscious information stored in LTM, and thereby allows us to know more than we initially were aware of knowing, but it also allows us to modify LTM itself. This idea, that imagery feeds back to affect what is stored explicitly in LTM, will prove especially important when we turn to PTSD.

In summary, I have made three major claims: (a) all reflective thinking involves imagery (of one sort or another), (b) searching through LTM is best thought of as a process of constraint satisfaction (which can activate a descriptive representation of the sought knowledge, such as the concept “round,” or can trigger processes that produce and interpret an image), and (c) the results of reflective thinking in turn enter new information into LTM and thereby can alter the circumstances when reflective thinking will be used in the future. The results of such reflective thinking are more profound that that, however; by entering new information into LTM, reflective thinking provides new grist for the constraint satisfaction mill: it provides additional cues that can be used to retrieve specific memories.

In the next section I explore how this framework may inform us regarding PTSD (occurring either in childhood or thereafter). We then end with a more general view of how this framework may provide insights into cognitive development, both normal and abnormal.

PTSD and the brain

Researchers have reported that imagining trauma engages the brain differently in PTSD patients than in control participants. In a typical study, the patient is asked to listen to a script that leads him or her to imagine either the traumatic event or a neutral event. The researchers then examine which parts of the brain were more activated during the traumatic imagery, compared to neutral imagery, and contrast this difference between the PTSD patients and control participants who did not have PTSD and were not abused in childhood. Using this technique, Rauch, van der Kolk, Fisler, and Alpert (1996) found that when PTSD patients imagined trauma-related stimuli, compared to neutral stimuli, limbic, and related structures (specifically on the right side) were activated. Lanius, Williamson, Boksman, Densmore, Gupta, Neufeld, Gati, and Menon (2002) used this technique to study PTSD patients who had been sexually abused, and found increased activation of prefrontal and limbic areas, compared to control participants, as well as increased activity in parts of visual cortex. The frontal areas appear to have included areas that previously had been shown to be activated during working memory (see Smith & Jonides, 1998, 1999). Similarly, Shin et al. (2004), using this paradigm, found decreased activation in the medial frontal gyrus during trauma-based imagery compared to neutral imagery, and further found that this activation was negatively correlated with the amount of blood flow in the amygdala. Furthermore, Shin et al. found that the more severe the PTSD symptoms, the more blood flow in the right amygala and the less blood flow in the medial frontal gyrus during trauma-related imagery.

In short, my reading of the neuroimaging literature leads me to two general conclusions. First, some parts of frontal cortex (particularly those involved in working memory) are overly active during trauma-related imagery, and other parts (in particular, those that play a role in motor programming) are underactivated. Second, various parts of the limbic system, including the amygdala, are being driven unusually strongly by the imagery. These results may indicate that trauma-related imagery is (a) easily triggered in these patients and (b) once triggered is relatively persistent. If so, then trauma-related imagery would engage working-memory areas more than neutal imagery, and to the extent that the patient is absorbed with the imagery they would have a greater emotional reaction (and, hence, the limbic activation) and be less prepared to act (and hence the reduced activation in the medial frontal gyrus).

A prediction of this interpretation of the results is that trauma-related imagery should be especially compelling for these patients; it should be vivid and lifelike. In fact, there is some suggest evidence that supports this view. For example, Stutman and Bliss (1985) found that PTSD patients rated their images (of neural scenes) as more vivid than did people with other disorders, and Laor, Wolmer, Wiener, Weizman, Toren, and Ron (1999) found that PTSD patients scored higher on a measure of imagery control.

Imagery in PTSD

Why does trauma-related imagery have such effects in PTSD patients, and what might be the role of imagery in this disorder? The connection I will suggest hinges on three facts. Let me first outline the facts, and then draw them together.

Alterations in imagery processing

This account of the role of imagery in PTSD leaves open a crucial question: are people with specific types of imagery more susceptible to developing PTSD following trauma, or does PTSD affect the nature of imagery? These alternatives are not mutually exclusive, and both factors may, in fact, be operating. Considering the first possibility, the previous discussion implies that people who already have very lifelike imagery might be more prone toward developing the disorder: by imagining the traumatic incident repeatedly after the fact, they continually reexperience it, which in turn would sensitize fight or flight systems as well as produce wildly varying levels of stress hormones. This hypothesis implies that the same images are experienced repeatedly, which is, in fact, what PTSD patients typically report (Hackmann, Ehlers, Speckens, & Clark, in press). Moreover, this hypothesis leads us to look for evidence that people who develop PTSD had especially vivid and controllable imagery prior to the disorder. Prospective studies would be difficult to do, but not impossible. For example, soldiers could be given imagery tests during their initial training phase, and the results on these measures could later be used to predict which soldiers will suffer PTSD following combat.

Turning to the second possibility, could PTSD actually affect the nature of imagery? Let us now go back to the theory, and consider imagery as a means of reflective thinking. According to the theory, when the appropriate cues are present, LTM is automatically searched (no attention or effort is required) and a “race” ensues: at the same time the constraint satisfaction process searches for stored descriptive encodings, it is also triggering the processes that generate an image. Imagery may “win” the race because one is very good at generating that particular image, so it is very fast, or because one is relatively slow at searching the relevant descriptive information. I want to consider the possibility that both alternatives are possible in PTSD patients.

First, these patients may be particularly adept at generating mental images of trauma-related incidents. This would occur if they dwell on such images, so that those images become practiced and easy to generate. In fact, such patients do report that they repeatedly generate a small number of trauma-related images (Hackmann et al., in press). Moreover, each time one generates the image and reflects on it, one can establish new associations in LTM. Each of those associations then can serve as a retrieval cue for the image. Over the long run, this implies both that more cues will access the representations of the image and that the image will be formed more quickly. It is as if the imagery system has developed a “hair trigger” for such images.

Second, PTSD patients may be relatively slow at searching the relevant descriptive information. These patients are easily aroused, and such arousal could disrupt descriptive search more than imagery, in part because the imagery processes are so overlearned and practiced that they are more robust to the internal “noise” produced by arousal. The arousal itself may increase over time, as the trauma-related images are generated more often, each time resulting in new trauma-related associations being stored. If so, then over time it will be more likely that the trauma-related images will be triggered.

In other words, I am suggesting a feedback loop: the more often the images are generated, the more associations to the image are stored (via the processes of reflective thinking discussed at the outset); and the more associations, the more likely the image is to be triggered in the future. In addition, each time it is triggered, the trauma engages fight or flight mechanisms, which disrupt descriptive thinking more than imagery. This view makes the somewhat counterintuitive prediction that patients with higher baseline arousal (e.g., as assessed by propensity to being startled) should have more lifelike imagery of the traumatic events.

In closing this section, I must emphasize that just as not every one who experiences trauma develops PTSD, not everyone with very lifelike imagery will be prone to develop the disorder. According to this line of thinking, the imagery feedback loop is only one part of the reason why PTSD develops. In addition, it is possible that one's brain must be particularly susceptible to arousal (which in turn, may be related to the release of specific stress hormones), and one must have difficulty in using cognitive control to dampen down unwanted processing. Furthermore, such patients may have more reactive autonomic or stress-related hormonal systems at the outset. Moreover, some or all of these sensitivities may arise from one's genes (e.g., Segman & Shalev, 2003). Without question, PTSD is a complex disorder, and the present goal is only to present an analysis that leads to relatively modest hypotheses about the possible role of imagery in the disorder.

Evidence of children's use of imagery

Years ago I provided evidence that older children and adults do indeed use imagery less often than younger children to answer questions about properties of objects (Kosslyn, 1976). In this study, I asked three groups of participants(5-year-olds, 14-year-olds, and undergraduates) to decide whether common animals have specific properties. The trick was that I selected the properties for each animal so that the more associated one was smaller than the less associated one. For example, for “mouse” the participants decided whether it has “whiskers” and also decided whether it has a “back” (“whiskers” is smaller but more associated with “mouse” than is “back”). There were two additional properties that were not characteristic of the object being queried (such as wings and fangs for a mouse). The animals and properties were presented in a random order, with an animal's being named and then, 3 s later, a property's being named.

In this study, the participants each took part in two conditions. In one, I explicitly asked them to visualize the named animals, and to use the image to evaluate the properties. In the other condition, I simply asked the participants to evaluate the properties as quickly as possible. There were three important results. First, in the imagery condition, all three groups responded faster for the larger (but less associated) properties than for the smaller (but more associated) properties, as expected if perceptual discriminability was crucial when using imagery. Second, although the participants generally required more time in the imagery condition than in the nonimagery condition, this difference in time between the two conditions was greater for the 12-year-olds than the 5-year-olds, and greater still for the adults. Moreover, although the older participants were faster than the younger ones even in the imagery condition, this advantage paled when compared to the very large differences in the nonimagery condition. Apparently, with age another sort of processing was increasingly available in addition to imagery, and this strategy was faster than using imagery. This is just as we would expect if the older people had encoded increasingly large sets of descriptions, which could be accessed immediately, whereas the youngest children had to rely on imagery.

Third, this interpretation was buttressed by a final result. When not explicitly asked to use imagery the 14-year-olds and adults were faster for the smaller (but more associated properties) than the larger (but less associated properties). This is as expected if they based their decisions on searching a network, where association strength reflects the strength of the connections or distance through the network (cf. Collins & Loftus, 1975). In contrast, the youngest children showed no such difference. In fact, they tended to respond faster for the larger properties. Moreover, when I explicitly asked the children after the study whether they used imagery even when not asked to do so, most of the youngest group reported that they did. Further, when I looked at the data from these participants, I found that they clearly responded faster when evaluating the larger properties, compared to the smaller. This finding did not reflect a difference in how strongly the youngest children associated the properties with the objects: even the youngest children ranked the smaller properties as more strongly associated with the animals, despite the fact that they required more time to evaluate these properties.

These findings are very much in line with the claim that people tend to recode perceptual information as descriptions after they access such information, and thus with age and experience they come to have increasingly larger sets of descriptive information in LTM. These representations allow us to respond automatically, without needing to engage in reflective thinking. This is a good thing, in that having to reflect about each and every decision posed by circumstance would be exhausting. But it is also a bad thing, in that we may often fail to notice something new or to reconsider our habitual responses. In my view, young children do not have as much stored information that allows them to respond automatically, and thus will need to reflect more often. It is tempting to speculate that it is this phenomenon, of reflecting about what we adults consider ordinary, that produces the “childlike freshness” of the young child, who can delight in what we adults find commonplace and can be annoyingly persistent in asking “why?”.

Developmental disorders of imagery

The present perspective not only provides a new way to regard PTSD, but leads us to consider a new way to look at symptoms that cut across the standard taxonomy of disorders, which are often characterized in terms of behavioral tendencies, such as impulsivity, hyperactivity, and problems in self control. Specifically, consider the consequences when specific aspects of processing are impaired.