tional MRI (fMRI)] appear to be sensitive to physiologic events that correlate with acute changes in neuronal activity on the order of seconds (5–8). This means that we are capable of asking a range of questions: Which brain areas are active during tasks known to promote long term memory storage (e.g., deep encoding tasks)? Which brain areas are active when information is intentionally retrieved from memory (e.g., explicit recall and recognition)? How do the brain areas that are initially activated by a task change after repeated performance of the task or prior exposure to information involved in the task (e.g., skill learning and priming)? However, it seems unlikely that we can observe the temporally distributed processes related to storage and consolidation directly.
A consequence of this notion is that, when we are observing active memory processes with neuroimaging alone, we are, to a greater or lesser degree, simply observing what might be thought of as “specialized” instances of typical information processing—tasks that are known to create, or benefit from, storage- and consolidation-related processes that bridge gaps over time. Often these memory tasks are the same tasks that are elsewhere described as language tasks or attention tasks (9, 10). The fundamental basis for this interdependency might be that many processes subsumed under the concept of memory are a by-product of normal information processing (11–13). Where information processing ends and “memory encoding” begins, for example, is a blurry distinction at best. After all, in typical everyday life how often do we actively try to memorize something? Yet, at the end of the day, we can recall a wide range and number of details regarding what happened to us. The exception to this comes in relation to explicit (or episodic) memory retrieval. During explicit memory retrieval, the active task demand is to intentionally retrieve information acquired at another time. In our discussion of neuroimaging studies of memory, we will examine both findings related to memory encoding and memory retrieval.
Functional neuroimaging techniques are also unlikely to resolve directly the flow of information processing that occurs over very brief time scales, such as on the order of 10s of milliseconds (14). However, these processes can be observed with electrophysiological recording techniques, and perhaps the integration of functional neuroimaging and electrical recording techniques will provide a comprehensive description of the spatio-temporal orchestration of human neural processing (8, 15–20).
Tasks that Promote Long Term Memory Encoding. Func tional neuroimaging studies of memory encoding were first conducted serendipitously (see discussions in refs. 9 and 10). Early studies of language function required subjects to generate and/or elaborate on the meanings of words (21, 22). Although not specifically intended as such, these tasks, as well as similar tasks that followed (23–25), were excellent long term memory encoding tasks. Tulving and colleagues (26, 77) directly demonstrated this by testing subjects for recognition following performance of a task used by Petersen et al. (21) to explore language function. The task was word (verb) generation, in which individuals are presented with nouns (e.g., “dog”) and are asked to generate associated verbs (e.g., “bark”). Tulving found that subjects showed high levels of recognition performance for words encountered during this word generation task. The PET imaging studies showed that word generation activated a pathway of brain regions including left prefrontal areas, the anterior cingulate, and the right-lateral cerebellum. Thus, the first insight from functional neuroimaging related to memory was arrived at: Active encoding of verbal information is tied to activation of a brain pathway including the left prefrontal cortex and functionally related structures. A series of more directed studies followed.
Kapur et al. (26), by using PET, sought to identify brain areas activated by deep encoding. Subjects were instructed to decide whether visually presented words represented entities that were either living or nonliving. This meaning-based word processing task led to 75% correct recognition of the words. Imaging data contrasting this deep encoding task and a shallow encoding task (decide whether the word contains the letter “a,” 57% correct recognition) demonstrated robust left prefrontal activation overlapping with the regions previously activated by the word generation tasks. This basic pattern of findings also has been demonstrated with fMRI. For example, in a series of fMRI studies, Gabrieli and colleagues have explored the functional-anatomic correlates of another deep encoding task, in which participants view words and then decide whether they fall into the category of abstract (e.g., hope) or concrete (e.g., tree) words. They found significantly greater left prefrontal activation during this deep encoding task than during a shallow encoding task in which subjects simply decided whether words were presented in uppercase or lowercase letters (23, 27). Our laboratory, by using fMRI, has followed up on some of these findings, and representative data are shown in Fig. 1.
Fletcher et al. (28) approached the issue in a different manner and compared later recall performance when participants first engaged in word generation concurrently with an easy distractor task (yielding high levels of recall, 83%) to a dual task situation in which word generation was paired with a difficult distractor task (yielding moderate levels of recall, 69%). Word generation paired with easy distraction, which presumably allows for more elaborate encoding, showed significantly greater left prefrontal activation than was observed during word generation in conjunction with the difficult distractor task. The conclusion across all of these studies is that the left prefrontal cortex, at or near Brodmann areas 44 and/or 45 and sometimes extending anteriorally and dorsally, is activated when subjects are engaging in tasks that lead to long term storage as assessed by later explicit retrieval tasks.
As can be seen across a wide range of manipulations, the data suggest that brain areas actively used to elaborate on word meaning or to access new words (word generation) are the same areas that lead to encoding of these events. Verbal working memory tasks also activate these areas (ref. 29; for reviews, see refs. 30 and 31). In a particularly elegant demonstration, prefrontal activation was shown to increase parametrically in relation to working memory load (32). The prefrontal activations tracking memory load were located in several areas of dorsolateral prefrontal cortex including portions of anterior frontal-operculum and the inferior frontal gyrus that directly overlap with areas activated by the “deep encoding” task shown in Fig. 1. What might this mean?
Buckner and Tulving (10) proposed that these functional neuroimaging studies demonstrate how multiple kinds of information processing might interact to promote encoding of long term memory. Effortful word generation tasks, verbal working memory tasks, and long term memory encoding tasks all activate similar brain pathways including left prefrontal regions and related structures. One possible account of this regularity is that the imaging studies conducted to date have not yet teased apart the various memory processes or separated distinct subregions within prefrontal cortex. This possibility currently is being explored further by several different laboratories. Alternatively, brain regions in left prefrontal cortex might be part of the neural substrate that maintains representations on-line (in working memory) while the representations are manipulated and used to guide further functions such as word access and generation. The basis of these representations may involve information coding in terms of phonology, lexical representations, more abstract semantic representations, or even coding of response alternatives. Further work is needed to explore the nature of the information that is being operated on in these left prefrontal regions.
The key point here is that these representations may themselves be the “encoding” that leads to storage and consolida-