Figure 21.1 FMRI activation patterns during working memory for fractal-like patterns (‘WHAT’), the locations of those patterns (‘WHERE’), or both the patterns and their locations (‘BOTH’). There is a dorsal/ventral spatial/nonspatial topography for the what-only and where-only tasks, but the activation for the BOTH task is in between the levels of activation for the preferred and non-preferred information types, consistent with a more selective representation for an object-in-its-location, involving fewer cells than are active during maintenance of the preferred information, but with those cells more highly active than during maintenance of the non-preferred information. Reprinted with permission from Sala and Courtney (2007). See also color plate 11.
(p.375) those cells more active than during the region’s nonpreferred task. What is the neural mechanism that could simultaneously account for these fMRI results and the dynamic selectivity observed in the Rao et al. (1997) and Rainer et al. (1998) studies?
Biased competition model for the selective maintenance of task-relevant information
A neural mechanism has recently been proposed that applies the principles of biased competition mechanisms for attentional selection to the problem of dynamically and selectively maintaining task-relevant information in WM (Sala and Courtney 2007). This model provides a single simple mechanism to explain both neuroimaging and single-cell physiology data regarding WM maintenance of objects, locations and the conjunction of an object and its location.
The model supposes domain-dependent inputs from parietal and inferior temporal cortices to dorsal and ventral PFC regions, respectively. Each of these PFC regions is proposed to then receive task-dependent input from the other. These excitatory inputs, combined with competitive interactions within each region, would result in the dual selectivity observed in the Rao et al. (1997) and Rainer et al. (1998) studies. In this model, when only spatial locations are task-relevant, the excitatory input is mainly from dorsal PFC to ventral PFC. In this case, all dorsal PFC cells which respond well to the to-be-remembered location are active, independent of their preferred object, because that object selectivity arises only when input is received from ventral PFC. When only object identities are task relevant, the excitatory input is mainly from ventral PFC to dorsal PFC. In this case, all ventral PFC cells which respond well to the to be remembered object are active, independent of their preferred location, because that location selectivity arises only when input is received from dorsal PFC. When the integrated representation of an object and its location is task relevant, then the excitatory inputs go in both directions and dual object and location selectivity is observed in both dorsal and ventral PFC. Cells would respond best only if the task trial required both the cell’s preferred location and its preferred object to be remembered. This type of system would result in a greater number of cells in dorsal (or ventral) PFC being highly active during a task that only requires memory for locations (objects, respectively) than when both an object and its location must be remembered. The fMRI activation, which depends on the average activity of all the cells in a region, would be expected to show the intermediate response for the object-in-location task that was observed by Sala and Courtney (2007).
An important feature of this model is that the strength and direction of the interactions between dorsal and ventral PFC are task dependent. A prediction of the model is that if the task requires that objects and locations be remembered independent of their relationship with one another (i.e., which object was in which location), then both dorsal and ventral PFC regions would be expected to be more active than during an task that required an integrated representation to be remembered. In addition, functional connectivity measures between dorsal and ventral PFC would be expected to be lower when objects and locations must be maintained independently than when an integrated representation is necessary.
The question remains how the strength and direction of these interactions are determined. A promising candidate is the other regions of the PFC whose activation levels do not appear to depend on the type of stimulus-specific information being maintained. As mentioned earlier, the level of activity in these regions appears to depend on the amount and complexity of the context and the rules governing task performance (e.g., Sakai et al. 2002, D’Esposito et al. 1999; Rypma et al. 2002; Bor et al. 2003; Derrfuss et al. 2004). Even in simple tasks, however, these regions could provide biasing signals that control the type of information to be maintained or allowed access to WM (as in Raye et al. 2002). Thus, biased-competition-like interactions among PFC regions could create new representations by combining information from multiple sources according to (p.376) the constraints of the current task rules. In the absence of these biasing signals and interactions both the information in WM and the resulting behavior would default to the most prepotent state, which would have been established through experience and previous behavior (see also Miller and Cohen 2001).
Interference-resistant maintenance versus updating
Another realm of WM where biased-competition-like mechanisms may be at work is in the updating of information in WM. Successful WM maintenance requires that interfering information that is not task-relevant be prevented from overwriting the current contents of the WM buffer. Indeed, many accounts of the role of PFC in WM maintenance emphasize the protection from interference aspect, rather than the storage of the information itself (e.g., Engle et al. 1999; Sakai et al. 2002). Some have suggested that WM storage capacity is fundamentally a function of attentional control over interfering information (Kane and Engle 2002). Simultaneously, however, the system must also allow the currently maintained information to be overwritten with updated information quickly whenever the new information becomes more important to the current task than the previously maintained information. This concept of relative priority is again reminiscent of biased competition for attentional selection, and data from recent neuroimaging studies support the idea that a similar mechanism is at work.
Roth and her colleagues (Roth, Serences and Courtney 2006) used fMRI to investigate the relationship between interference-resistant WM maintenance and updating the contents of WM with a new sample stimulus. They used a modified delayed-recognition task in which participants viewed a continuous stream of either faces or houses. The first object in each task block was the first sample to be maintained in WM. With the presentation of each subsequent stimulus, participants indicated with a button press whether the current stimulus matched the sample stimulus. Randomly, every 4–10 seconds, participants saw one of two well-memorized cue faces or houses. One of these cues instructed the participants that the old sample was now irrelevant and that they were to maintain in WM the next face or house that they saw and to make future match/nonmatch decisions in reference to this new sample stimulus. The other cue served as a control event. That cue instructed participants to continue to maintain the current sample stimulus. For both cue events, participants pressed a button to indicate that they recognized the cue stimulus. There were also control task blocks in which participants viewed the same stream of faces or houses, but did not need to maintain any one of them in WM. Instead, participants made a perceptual categorization decision (male/female for faces, garage/no garage for houses) for each stimulus as it was presented. Thus, the experimental design included both sustained activation components for memory blocks versus control blocks, related to WM maintenance, and transient activation components related to the update events. Contrasting the update events to the control cued maintenance events enabled identification of activity related to replacing the current contents of WM with a current perceptual stimulus.
Independent of the type of object to be remembered (face or house), transient, update-related activations were observed primarily in middle and superior frontal regions and parietal cortex. These regions partially overlapped with those regions showing sustained, maintenance-related activity. Notably, the middle frontal region activated by both maintenance and updating appears to be the same region as has been previously implicated in WM maintenance under conditions of interference (Sakai et al. 2002). A similar region of activation has been observed in other studies in which the task required actively maintained abstract information, such as rules, to influence cognition or behavior (Bunge et al. 2003; Derrfuss et al. 2004; see Figure 21.2.). In the Roth et al. study, participants needed to continue to maintain the sample stimulus while making match/nonmatch (p.377)
Figure 21.2 Activation of middle frontal cortex (possibly Brodmann’s Area 46) in several studies involving the prioritization of one type of information over another: (a) Updating (in green), WM maintenance (blue), and their overlap (yellow) (reprinted with permission from Roth et al.2006), (b) Common activation for task-switching, Stroop task, and n-back working memory task (reprinted with permission from Derrfuss et al. 2004), (c) Refreshing of just-seen information (reprinted with permission from Johnson et al. 2003), (d) Interference-resistant WM maintenance (reprinted with permission from Sakai et al. 2002). See also color plate 12.
Transient changes in activation have also been observed in a highly similar middle frontal region when research participants refreshed the mental representation of a recently seen (but not actively maintained) stimulus (Raye et al. 2002; Johnson et al. 2003; see Figure 21.2.). Both this refreshing task and the update events in the Roth et al. study involve a change in the relative priority of current versus previously seen information. The other regions that showed both sustained, maintenance-related activity and transient, update-related activity in the Roth et al. study were in (p.378) parietal and superior frontal cortex in regions highly similar to those that have been shown to have both sustained and transient activity related to maintaining and shifting attention, respectively (Yantis et al. 2002; Serences et al. 2004).
The existence of both sustained, maintenance-related activity and transient, update-related activity in brain regions that have previously been implicated in protecting information in WM from interference and regions that have previously been implicated in controlling attentional selection among current perceptual stimuli, suggests that a common mechanism may be involved in all of these functions. We propose that the common requirement in all of these functions is the setting of relative priorities according to current task context and rules. Biased competition could be at work not only in selecting whether object identity or location or both would be attended in the perceptual input, but also which type of information would continue to be maintained during a WM delay. In addition, actively maintained information about task context and rules could also set the relative priorities for information already being maintained in WM versus current perceptual input. Changes in these relative priorities could result from either explicit instruction cues, as in the Roth et al. (2006) study or from unexpected reward feedback (see Rougier et al. 2005). Such resolution among competing sources and representations of information is necessary because of the limited storage capacity of WM.
Summary and future directions
Taken together, the studies reviewed in this chapter suggest a role for abstract contextual or rule information maintained via the mid-dorsolateral PFC to bias competitive interactions within and among other brain regions (including other regions within the PFC) in the service of selecting, creating, maintaining, and updating the optimal representation of the most important information for the current task. This select, more stimulus-specific information then in turn could serve as the biasing signal for selection of particular actions (Cisek and Kalaska 2005; for reviews see Fuster 2001, Bunge 2004), selection of competing conceptual representations in long-term memory (Kan and Thompson-Schill 2004), or for attentional selection, as in the original biased competition model (Desimone and Duncan 1995, Desimone 1998). This framework implies a hierarchical structure with domain-dependent information maintenance and selection via biased competition occurring at every level. The proposed model is shown in Figure 21.3.
In the model there are mutually inhibitory connections within each level of the hierarchy and excitatory inputs across levels. Which representation will ‘win’ that competition at each level depends on how the balance of competition is biased both by the saliency of feedforward perceptual information and by feedback signals from higher levels. The model allows for multiple representations to be active simultaneously, but it is their relative levels of activity that will ultimately determine their influence on other brain regions and, thus, on behavior (see also Bisley and Goldberg 2003).
The most powerful aspect of the model is that it enables many aspects of perception, action and cognition to be explained by a common mechanism: relative levels of sustained activity in one brain area biasing the interactions within mutually inhibitory neural networks in another brain area. While such mechanisms have been proposed before, the unusual aspect of the current proposal is that these interactions could occur within the PFC as well as between the PFC and perceptual or motor regions. Such a unifying framework could simplify discussions of WM, cognitive control and other executive processes. If the model holds true, the difference between the brain regions responsible for various types of executive processes or between the phonological loop and the visuospatial sketch pad would be defined primarily on the type of information represented (as determined by the inputs to that region) and the type of influence that the information (p.379)
Figure 21.3 Hierarchical biased-competition (HBC) model for working memory maintenance and cognitive control. ‘A’ represents the current task schema with all of its associated goals, rules, and sensory information, each of which is proposed to be represented by activity in a different level of the hierarchy. Note that there are two parts to the rule representation (1A and 2A). The greater activation of 1A relative to other possible rules determines whether the current contents of working memory or the current perceptual input is more important for future maintenance in WM. The greater activation of 2A relative to other possible rules determines which objects, locations, or other stimulus-specific information will be maintained.
Related posts:
What do estimates of working memory capacity tell us?
What do working memory span tasks like reading span really measure?
The general-purpose working memory system and functions of the dorsolateral prefrontal cortex
Separating processing from storage in working memory operation span
Accounting for age-related differences in working memory using the feature model
The interpretation of temporal isolation effects
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
