An Information-Theoretical Approach to Contextual Processing in the Human Brain: Evidence from Prefrontal Lesions
1 Department of Psychology and Institut Universitari d'Investigació en Ciències de la Salut (IUNICS), Universitat de les Illes Balears, 07122 Palma de Mallorca, Spain, 2 Department of Psychology and Helen Wills Neuroscience Institute, University of California at Berkeley, Berkeley, CA 94720-1650, USA
Address correspondence to Francisco Barcelo, Department of Psychology and Institut Universitari d'Investigació en Ciències de la Salut (IUNICS), Universitat de les Illes Balears, Carretera de Valldemossa km 7.5, 07122 Palma de Mallorca, Spain. Email: f.barcelo{at}uib.es.
 |
Abstract |
|---|
 Top Abstract IntroductionResultsDiscussionSummary of Findings and...Supplementary MaterialFundingReferences |
|---|
Context shapes perception, thought, and action, but little is known about the neural mechanisms supporting these modulations. Here, we addressed the role of lateral prefrontal cortex (PFC) in context updating and maintenance from an information-theoretic perspective. Ten patients with PFC lesions and 10 age-matched controls responded to bilaterally displayed visual targets intermixed with repetitive and novel distracters in 2 different task contexts. In a predictable context, targets were always preceded by a novel event, whereas this temporal contingency was removed in an unpredictable context condition. We applied information theory to the analysis and interpretation of behavioral and electrophysiological data. The results revealed deficits in both the selection and the suppression of familiar versus novel information mainly observed at the visual hemifield contralateral to PFC damage due to disrupted frontocortical and frontosubcortical connectivity. The findings support a deficit in the representation of the temporal contingency between contextually related novel and familiar stimulation subsequent to lateral PFC damage.
Key Words: associative learning • cognitive control • information theory • novelty • working memory
|
Introduction |
|---|
|
TopAbstractIntroductionResultsDiscussionSummary of Findings and...Supplementary MaterialFundingReferences |
|---|
The neural mechanisms for processing contextually related information have attracted considerable interest in recent years. Animal and human research points to a key role of prefrontal and posterior multimodal association cortices for the integration of contextual information (Donchin and Coles 1988
; Cohen et al. 1996; Miller and Cohen 2001). Postrolandic parietal and temporal unimodal and multimodal association cortices are important for setting up the spatial context (Rafal and Posner 1987; Rafal et al. 1990; Rushworth and Taylor 2006), whereas the lateral prefrontal cortex (PFC) seems critical for establishing the temporal contingencies between contextually related events (Fuster et al. 2000; Fuster 2002; Koechlin et al. 2003). For instance, visual orienting toward contextually novel events depends primarily on innate reflexive neural programs (Sokolov 1963; Pierrot-Deseilligny et al. 2004). However, visual reflexes can be modulated within a temporal context of past experiences and future goals with obvious advantages for adaptive and flexible intentional behavior (Miller and Cohen 2001). These contextual modulations have been proposed to depend on an intrinsic interplay between exogenous and endogenous sources of information in distributed neural networks including PFC (Crick 1984; Tononi and Edelman 1997; Edelman and Tononi 2000; Miller and Cohen 2001; Friston 2005).
There are contrasting views as to whether temporal context influences cognitive control at a preperceptual (Näätänen 1990; Rafal et al. 1990), perceptual (Sokolov 1963; Donchin and Coles 1988), decisional (Nieuwenhuis et al. 2005; Daw et al. 2006), or sensorimotor (Hommel et al. 2001; Koechlin et al. 2003) stages of neural processing. This issue can be framed in anatomical terms regarding where contextual representations are held and updated in the brain (e.g., at subcortical, unimodal sensory, or posterior vs. prefrontal multimodal association cortices). Two recent models make opposite predictions about how context modulates brain physiology and behavior: the "context-updating" (Donchin 1981; Donchin and Coles 1988) and the "guided-activation" models (Cohen et al. 1996; Braver et al. 2002).
The context-updating model posits that the modulations of an endogenous "P300" component of the human event-related potential (ERP) index updating of working memory representations triggered by a mismatch between a task event and its perceptual context (Donchin and Coles 1988). The neural mechanisms indexed by these P300 brain potentials have been linked to stimulus change detection (Donchin 1981), perceptual distinctiveness (Polich 2003), and stimulus categorization (Donchin and Coles 1988). The novelty or familiarity of the eliciting event within its immediate temporal context determines the type of P300 activation observed. Contextually novel events elicit transient "novelty P3" activations with maximal amplitudes over frontocentral scalp regions (Polich 2003) that depend on the integrity of a distributed neural network including lateral PFC (Knight 1984), mesial temporal cortices (Knight 1996; Ranganath and Rainier 2003), temporoparietal cortices (Knight and Scabini 1998), as well as subcortical structures (Ranganath and Rainier 2003; Nieuwenhuis et al. 2005). Familiar target events elicit transient "target P3b" activations with maximal intensity over midparietal scalp regions. However, the context-updating model does not fully account for novelty P3 activity to task-irrelevant distracters (Donchin and Coles 1988; Dien et al. 2004). To date, there is no integrative view of the human P300 response that explains both the selection of targets and the suppression of contextually related distracters and how each of these operations tax our capacity for processing information in working memory (Miller 1956). Such an integrative theory of the human P300 response should account for those aspects shared by novelty P3 and target P3b activations (i.e., do they both index context-updating operations in working memory?), as well as for those aspects that are unique (i.e., do they each index the updating of different neural representations?).
We utilized an integrative model of PFC function to examine these 2 questions (Miller and Cohen 2001). The guided-activation model considers a functional hierarchy of representations from unimodal association to posterior and prefrontal multimodal association cortices (Fig. 1; cf., also Fuster 2002; Koechlin et al. 2003), thus providing a benchmark for testing predictions about context-updating operations at 3 different levels of cortical representation (Braver et al. 2002). This model assumes that activation of any perceptual element leads to updating of its associated units in the neural network, including higher ordered memory units and appropriate responses (Miller and Cohen 2001). The frequency of updating of each perceptual element then determines the relative strength of the intervening sensorimotor pathways but also the relative recruitment of PFC, with more frequently updated representations requiring lesser PFC resources (Miller and Cohen 2001). The mean probability of occurrence of a task event offers an approximate measure of the information-processing resources associated with a stimulus and has been routinely adopted in most past ERP research (Donchin 1981; Polich 2003). Indeed, rare targets and novel distracters elicit larger P300 activations and tax working memory more than repetitive standard stimulation. However, such coarse estimations fail to consider the mutual information conveyed through the co-occurrence of contextually related distracter and target events, nor do they consider information transmission for perceptual representations (i.e., visual objects) as different from that conveyed at higher ordered levels of neural representation (i.e., memory chunks or semantic categories; Tononi and Edelman 1997; Koechlin and Summerfield 2007).
|
The model depicted in Figure 1 illustrates the perceptual and motor elements of a simple perceptual judgment task, consisting of 3 contextually related visual stimuli (s1, s2, and sx) and 2 behavioral outcomes (r0 and r1) represented at visual association and premotor cortices, respectively (adapted from Miller and Cohen 2001
). By means of explicit instructions and learning, visual targets (s2) become associated with an overt motor response (r1), whereas standard (s1) and novel (sx) distracters require withholding any overt responses (r0). In the model of Miller and Cohen, these sensorimotor (S–R or stimulus–response) pathways are assumed to be held outside the PFC, for example, at posterior multimodal association cortices (e.g., s1–r0 and s2–r1 pathways in Fig. 1a; Rushworth and Taylor 2006). Finally, superordinate PFC units are assumed to connect stimuli with responses in a context-sensitive way through subsets or "chunks" of subordinate sensorimotor pathways, very much like a "switch operator in a system of railroad tracks" (Miller and Cohen 2001, p. 184). Lateral PFC lesions severely disrupt novelty P3 activity but leave target P3b activity relatively intact (Knight 1984, 1997). Accordingly, we examined the hypothesis that familiar and novel information each have distinct superordinate PFC representations, although they may partly share subordinate sensorimotor representations, as shown in Figure 1a,b, respectively.
We also followed original recommendations by George Miller (1956) for estimating the amount of information transmitted between contextually related stimuli and responses (or "input–output correlations"; Miller 1956) along the 3 layers in the hierarchy of representations in Figure 1 (see also Koechlin et al. 2003). Information theory offers a dimensionless yardstick for exploring the universal properties of human working memory independent of specific sensory or motor demands in target and distracter trials and helps formalize concepts such as "context," "novelty," or "stimulus saliency" (Miller 1956; Koechlin and Summerfield 2007). Information-theoretic analyses based on the joint and conditional probabilities between task stimuli and responses were used to estimate the mutual information conveyed through the temporal co-occurrence of targets and distracters and to clarify whether this contextual information was conveyed through subordinate S–R pathways at postrolandic scalp regions or through superordinate PFC representations (cf., Table 1 and see Appendix in the Supplementary Material).
|
Using this novel approach, we examined predictions from the "guided-activation" and "context-updating" models about the role of PFC in updating and maintenance of contextual information by comparing the behavior and brain responses of 10 patients with unilateral PFC lesions (Fig. 2), with a group of age-matched healthy controls. We addressed 2 important issues: 1) whether distracter-locked novelty P3 activations also reflect context-updating operations and 2) whether the neural representations involved are different from those of target-locked P3b activations (cf., Donchin and Coles 1988
; Dien et al. 2004). In so doing, we manipulated contextual predictability through the temporal contingency between novel distracters and familiar targets in a simple visual attention task and then examined how this influenced neural activity and the ability to perform easy visual discriminations. PFC-lesioned patients and controls responded to inverted triangles (targets) embedded in rapid trains of repetitive upright triangles (standard distracters) and unique "oddball" color pictures (novel distracters) randomly displayed to both visual hemifields (cf., Barcelo et al. 2000). In one condition, infrequent novel events conveyed no contextual information about the next visual target response (unpredictable context). In another condition, novels were always followed by a visual target either at the same or at the opposite visual hemifield (predictable context; Fig. 3). The same instructions were used in both task contexts, and the novel–target contingency was manipulated implicitly by varying the statistical regularities between task events (see Table A2 in the Appendix of Supplementary Material). Our bifield stimulus display limited the patients' ability to use their intact PFC to compensate for deficits and allowed us to compare the patients' behavior and brain activations at the visual hemifields both ipsilateral and contralateral to PFC damage (i.e., this bifield display effectively added 1 bit of information across all other task conditions, see Materials and Methods in the Supplementary Material; cf., Barcelo et al. 2000). The combined manipulation of temporal and spatial context also allowed us to dissociate the modulation of information at multimodal PFC from the transmission of information at retinotopically organized visual association cortices (Miller and Cohen 2001; Friston 2005).
|
|
The context-updating model does not make any explicit predictions about distracter-locked novelty P3 activations or about the role of PFC versus posterior association cortices in updating contextual representations, although task-relevant P300 activations have been linked to temporal–parietal cortical regions (Donchin 1981
; Knight and Scabini 1998). The context-updating model predicts larger parietal P3 activity to unexpected target events in unpredictable—compared with predictable—stimulus contexts. Alternatively, we hypothesized that the more informative predictive novels would elicit the larger novelty P3 activations (Barcelo et al. 2006). In turn, target P3b amplitudes should not be affected by contextual predictability because this did not alter the mutual information conveyed by target stimuli for response selection in the present task (see Table A1 in the Appendix of Supplementary Material). These predictions are contrary to the context-updating model but are consistent with our estimations of mutual information (see the Appendix in the Supplementary Material, Koechlin and Summerfield 2007) and with a series of task-switching studies showing that novelty P3 and target P3b activations index context-updating operations at 2 hierarchically distinct levels of neural representation: target P3b indexed updating of familiar sensorimotor representations at posterior association cortices (Fig. 1a), whereas cue-locked novelty P3 activations indexed updating of novel superordinate sensorimotor representations (Fig. 1b, cf., Barcelo et al. 2002, 2006).
The guided-activation model explicitly implicates PFC in processing contextually related target and distracter information (Fig. 1a,b, Miller and Cohen 2001; Braver et al. 2002). Although, this model has not been previously tested using human lesion P300 ERP data, 4 of its general predictions about cognitive control are relevant to the present study. First, the model predicts a role of PFC in setting up and holding online the task context, rather than any isolated stimulus or perceptual features. Second, lateral PFC enables cognitive control in response to conflict signals from subcortical or frontomedial structures (Miller and Cohen 2001; O'Reilly et al. 2002). Third, a paradoxical amelioration of context-induced errors should follow context-processing deficits in PFC patients (Braver et al. 2002, p. 440). Finally, this model predicts that a unitary PFC-dependent representation of context can explain the selection of target information, the inhibition of distracter information, and working memory operations (Miller and Cohen 2001; Braver et al. 2002).
We explored specific predictions about the role of PFC versus posterior association cortices in the elicitation of the human P300 response. However, one should not dismiss the role of subcortical structures in the cognitive control of visual orienting to novel events (Zink et al. 2003; Pierrot-Deseilligny et al. 2004; O'Reilly 2006), in line with age-old ideas about the hierarchical architecture of control in the nervous system (Jackson 1884). At least 2 sources of extra-PFC influences are worth considering: First, the extracortical control of vision involving well-defined corticotectal connectivity (Goldman and Nauta 1976; Rafal et al. 1990; Gaymard et al. 2003; Pierrot-Deseilligny et al. 2004; Johnston and Everling 2006). Second, subcortical and/or posterior cortical structures have been proposed to support successful delayed target discriminations in the absence of distracters in monkeys and humans with extensive PFC lesions (Malmo 1942; Knight 1984), in accord with the distinct neural substrates for the exploitative processing of familiar information as opposed to the exploration of novel information (Daw et al. 2006; O'Reilly 2006).
|
Results |
|---|
|
TopAbstractIntroductionResultsDiscussionSummary of Findings and...Supplementary MaterialFundingReferences |
|---|
Behavior
Missing Errors and False Alarms
Controls correctly responded to 93.6% of targets compared with an overall 81.0% hit rate in patients (F1,18 = 7.3, P < 0.02). Patients missed more targets at their contralesion—rather than ipsilesion—visual hemifield (F1,18 = 8.5, P < 0.01; 11.6% contralaterally vs. 7.4% ipsilaterally missed targets; cf., Barcelo et al. 2000). Error rates were lower at the ipsilesion visual field for predictable targets (2.6% ipsi- vs. 4.2% contralesion, P < 0.01), with relatively larger miss rates for ipsilesionally displayed unpredictable targets (6.9% ipsi- vs. 4.8% contralesion, P < 0.05). Contextual predictability did not influence miss rates of targets displayed contralaterally to lesion.
Patients committed more false alarms than controls (F1,18 = 7.7, P < 0.02; mean false alarm rates of 1.0% vs. 2.5% for controls and patients, respectively). False alarm rates were lower in the predictable context (F1,18 = 7.8, P < 0.02).
Reaction Times
The pattern of results for response accuracy was mirrored by reaction times (RTs) (Fig. 4). RTs were slower in patients than controls (F1,18 = 6.3, P < 0.03; mean ± standard error of the mean [SEM] for patients: 583 ± 10 ms; controls: 516 ± 11 ms). PFC patients had a contralesion target detection deficit (interaction group by visual field of target display: F1,18 = 9.1, P < 0.008), with RTs 590 versus 577 ms for contralesion versus ipsilesion targets, respectively. Across patients and controls, responses to predictable novel–target pairings were faster (mean ± SEM: 531 ± 15 ms) than those following unpredictive novels (555 ± 14 ms) or standards (556 ± 14 ms) (interaction of predictability by distracter type: F1,18 = 10.3, P < 0.005). A main effect of distracter type (F1,18 = 4.8, P < 0.04), and its interaction with group (F1,18 = 4.7, P < 0.05), revealed that controls benefited more than patients from contextual predictability for speeding up their target responses. Across both visual hemifields, controls made faster target responses following a predictive novel (mean ± SEM: 504 ± 11 ms) than following a standard (529 ± 10 ms), whereas the patients' responses did not differ between these 2 conditions (583 ± 20 ms vs. 583 ± 11 ms for targets following standard and novel events, respectively).
|
PFC patients benefited from target predictability only at their ipsilesion visual hemifield (interaction of group by predictability by target field: F1,18 = 6.8, P < 0.02), and this effect was specific for novel–target pairings (F1,18 = 8.5, P < 0.01; Fig. 4). In controls, the response benefit for predictable targets was equivalent across both hemifields. In the patient group, the costs and benefits, respectively, conveyed by unpredictive and predictive novels were observed only when the following target appeared at the ipsilesion visual hemifield (F1,9 = 11.8, P < 0.003). For contralesion targets, RTs were not influenced by the contextual information conveyed by the preceding novel stimulus (Fig. 4, frontals). The visual hemifield of novel display did not influence this visual target effect (interaction of group by predictability by target field by novel field: F1,18 < 1). In line with missing error rates, contextual predictability improved the speed of visual discriminations only when targets were displayed at the ipsilesion visual field of patients.
Standard–target pairings were not influenced by contextual predictability. Therefore, predictability effects could be specifically attributed to the mutual information conveyed by the novel events about the next target trial and were not related to unspecific factors such as reduced effort or overall difficulty in trial blocks with fully predictable visual target discriminations.
Electrophysiology
Standard, target, and novel events elicited well-known sensory ERPs recorded as positive (P1) and negative (N1) voltage deflections with maximal intensity over temporooccipital regions contralateral to the hemifield of visual display. As reported previously, loss of top-down PFC-dependent input reduced these visual P1 and N1 responses over the temporooccipital cortex ipsilateral to the lesion (Barcelo et al. 2000; Yago et al. 2004). These sensory brain potentials were not influenced by manipulations of contextual predictability and will not be discussed further (see Table 2).
|
Brain Responses to Novel Events
Contextually novel events elicited a series of stereotypical brain potentials in controls (labeled P2, N2, novelty P3, and N4 in Fig. 5; see Table 2 for a summary of ERP results), all with a frontal or frontocentral maximal voltage distribution. All these brain potentials were altered in the patients, but some were also modulated by contextual predictability (Fig. 5). Lateral PFC lesions reduced P2 amplitudes (peak latency 250 ms at Fpz; Fig. 5, Controls) over frontopolar and frontal regions (F1,18 > 4.5, Ps < 0.05 at Fpz and Fz; Fig. 5, Frontals). This P2 reduction was larger at recording sites contralateral to the visual field of stimulation (interaction of visual field by electrode: F1,18 = 9.0, P < 0.01; not shown). In both patients and controls, novel events also elicited a transient negative field potential with a frontocentral scalp distribution (N2; peak latency 340 ms at Fz; Fig. 5, Controls). Patients exhibited increased N2 amplitudes (F1,18 = 8.8, P < 0.01 at Fpz; Fig. 5), as well as a rostral displacement of their scalp topography. For both controls and patients, larger N2 amplitudes were recorded over frontopolar regions contralateral to the visual field of novel display (interaction of visual field by electrode: F1,18 = 40.9, P < 0.0001) and such an effect was larger in the patients (F1,18 = 9.1, P < 0.007; not shown). Mean N2 amplitudes in response to novels were not influenced by contextual predictability in controls or patients (Fig. 5).
|
In controls, mean novelty P3 amplitudes were larger in response to predictive compared with unpredictive novels over frontopolar (Fpz; F1,18 = 5.8, P < 0.03) but not at more posterior regions (Fig. 5, Controls). Mean novelty P3 amplitudes were reduced in patients in all conditions (F1,18 = 5.2, P < 0.04 at Fz), and this reduction was larger for predictive than unpredictive novels displayed ipsilaterally to lesion (interaction of group by predictability by visual field: F1,18 = 5.4, P < 0.03). In turn, context did not influence novelty P3 amplitudes to novels displayed in the hemifield contralateral to PFC damage (Fig. 5a,b, Frontals). These deficits could not be attributed to group differences in the peak latency of novelty P3 (F1,18 = 1.2, not significant; 443 ms for patients, 458 ms for controls at Fz). Likewise, peak-to-peak N2–P3 amplitudes did not differ between groups (3.4 vs. 4.5 µV at Fz for patients and controls, respectively; P = 0.3). Predictive novels elicited the largest novelty P3 responses over frontopolar regions in controls, whereas the largest abnormalities in novelty P3 activity were recorded over midfrontal regions in patients (interaction of group by predictability by electrode: F1,18 = 4.6, P < 0.04). Importantly, predictive novels flashed ipsilaterally to lesion also elicited abnormal sustained negative activity during 50–200 and 400–600 ms poststimulus onset (interaction group by predictability by visual field by electrode: F1,18 = 6.3, P < 0.02). These early and late anomalous sustained negativities showed similar scalp topographies (F19,171 = 1.2, P = 0.4). In turn, the scalp distribution of these abnormal early and late negativities differed from that of novelty P3 activity in the patients (F19,171 = 6.3, P < 0.003, Greenhouse-Geisser = 1.5). These results suggest impaired neural processing of all novel information in PFC patients, consisting of a bilateral disruption in a sequence of stereotypical transient brain responses, together with an anomalous sustained negativity associated with predictive novels displayed to the ipsilesion visual hemifield of patients.
Brain Responses to Target Events
Target events evoked 2 well-known ERP signatures with a temporoparietal and midparietal scalp topography, respectively (labeled N2 and P3b in Fig. 6), none of which were affected by contextual predictability in controls. In the patients, predictable targets elicited normal target P3b activity over midparietal and temporoparietal regions, whereas unpredictable targets displayed contralateral to PFC lesions elicited reduced target P3b activity over midparietal (F1,18 = 5.8, P < 0.03 at Pz; Fig. 6a, Frontals: contra targets), as well as over ipsi- and contralesional temporoparietal regions (P < 0.004 and P < 0.04, respectively; Fig. 6b).
|
Lesion-Related Asymmetries in Brain Activation
We examined the hemispheric asymmetries in the brain responses to novel distracters and familiar targets between the lesioned and the intact PFC regions in the unpredictable context condition (Fig. 7). Asymmetries were observed in the brain responses to novel distracters, with enhanced P2 activity but reduced novelty P3 activity over the lesioned, as compared with the intact, lateral PFC region (F1,9 > 5.6, Ps < 0.04). In contrast, the brain responses to familiar targets showed enhanced target P3b activity and reduced N2 activity over the lesioned, as compared with the intact, lateral PFC region (F1,9 > 7.3, Ps < 0.02). The abnormally enhanced N2 amplitudes to all novel distracters did not differ between the lesioned and the intact hemispheres.
|
|
Discussion |
|---|
|
TopAbstractIntroductionResultsDiscussionSummary of Findings and...Supplementary MaterialFundingReferences |
|---|
Lateral PFC-lesioned patients showed behavioral and electrophysiological deficits during the updating and maintenance of contextually related familiar and novel information. At least 3 distinct deficits were identified: 1) a general deficit in processing all novel information, 2) a deficit in acquiring the temporal association between predictive novels and targets, and 3) a deficit in the selection of unpredictable targets. The first 2 deficits corresponded with abnormal brain responses to novel events displayed either ipsilateral or contralateral to PFC damage. Both behavioral responses and target ERPs were abnormal to unpredictable targets flashed contralesionally. These results support the hypothesis that damage to a unitary PFC representation of context results in several apparently distinct cognitive deficits (Miller and Cohen 2001
; Braver et al. 2002). The data also support a crucial role of lateral PFC in updating and online maintenance of both familiar and novel information (Malmo 1942; O'Reilly 2006). Impaired Processing of All Novel Information in PFC Patients
The behavioral deficits of PFC patients were matched with 2 distinct abnormalities in the brain responses to novel—but not familiar—information: 1) a bilateral disruption in the stereotypical ERPs to all novel events that was influenced neither by the hemifield of novel display nor by contextual predictability and 2) an anomalous and sustained negative potential related to predictive novels displayed ipsilesionally. The frontal topography of these 2 abnormalities supports a disrupted superordinate sensorimotor representation of novel information (i.e., memory chunk II in Fig. 1b; Barcelo et al. 2006), rather than a purely perceptual representation of the stimulus context at posterior association cortices.
Bilaterally Disrupted Updating of Novel Information
A well-known sequence of stereotypical brain responses to novel stimulation (e.g., P2, N2, novelty P3, and N4 in Fig. 5) was altered in the patients regardless of the hemifield of novel display or contextual predictability. These ERP abnormalities contrasted markedly with the relatively normal ERP activations recorded to familiar task–relevant targets over the same prerolandic regions and with the normal ERP activations to the same novel stimuli over postrolandic multimodal association regions (Table 2). These results confirm previous findings and lend support to the notion that there are distinct PFC representations for novel and familiar information (Knight 1984). These ERP results also agree with our estimations of information transmission within the hierarchy of representations in Figure 1 (see Tables A1 and A3 in the Appendix of Supplementary Material), suggesting that in spite of not requiring any overt responses, unpredictive novels exceeded the theoretical limit of the human capacity for processing information in working memory (cf., Miller 1956) and did so by conveying information at a superordinate level of representation (Koechlin et al. 2003). Targets also conveyed information for response selection but did so through subordinate sensorimotor representations at posterior association cortices (Tables A1 and A3 in the Appendix of Supplementary Material). Accordingly, the updating of target information demanded mainly temporoparietal—rather than PFC—activations (Figs 1a and 6), whereas the updating to a novel memory chunk recruited activity at both PFC and temporoparietal cortices (Figs 1b and 5; cf., Miller and Cohen 2001).
The bilaterally disrupted brain responses in patients with unilateral PFC lesions suggest that updating of novel information involves intercallosal cross talk (Knight 1996) that is known to be driven by arousal systems in the brainstem (Sokolov 1963; Crick 1984; Zink et al. 2003; Nieuwenhuis et al. 2005). The prerolandic scalp distribution of disrupted brain responses to all novel information questions their interpretation in terms of forward transmission of information through sensory-specific geniculostriate pathways. In turn, these findings concur with modulatory interactions between exogenous and endogenous sources of information through modality nonspecific bidirectional corticosubcortical pathways (Sokolov 1963; Edelman and Tononi 2000; Friston 2005). One plausible source of exogenous information is arousal systems in the midbrain (Zink et al. 2003; Nieuwenhuis et al. 2005) that communicate with lateral PFC through well-defined prefrontotectal connectivity (Goldman and Nauta 1976; Gaymard et al. 2003; Johnston and Everling 2006).
Ipsilesionally Disrupted Maintenance of Novel Information
This deficit in novelty processing was accompanied by a secondary context-sensitive deficit related to our use of predictive novels as anticipatory cues for target selection. In controls, predictive novels elicited enhanced novelty P3 amplitudes over rostrofrontal regions (Fig. 5). This finding supports the hypothesis that novelty P3 and target P3b activations each index different context-updating operations at different levels in the hierarchy of representations in Figure 1. This finding also raises interpretative problems for the context-updating model that predicts lesser P300 activations in the more predictable stimulus contexts (Donchin and Coles 1988). Alternatively, the more informative novels elicited the larger novelty P3 activations, consistent with our use of predictive novels as contextual cues for anticipatory target selection (cf., Barcelo et al. 2002; Barcelo et al. 2006). Information-theoretic analyses indicated that the extra contextual information of predictive novels was not conveyed through posterior S–R pathways ("sensorimotor control," Koechlin and Summerfield 2007) but rather through superordinate PFC representations ("contextual control," Koechlin et al. 2003; cf., Fig. 1 and Appendix, see Supplementary Material).
In PFC patients, contextual predictability modulated brain physiology and behavior only when novel information accessed the intact PFC of patients through their ipsilesion visual field (Figs 4 and 5). The lack of any contextual modulations for predictive novels flashed at the visual hemifield contralateral to damage supports a critical role of PFC in establishing the temporal context for the anticipatory control of vision. The lesioned PFC could assist neither in building up the representation of a novel visual object at posterior association cortices (cf., Barcelo et al. 2000) nor in holding it online to establish an association with the following target response. In contrast, when predictive novels were displayed at the good visual hemifield of patients, they implemented the temporal contingency and could readily anticipate the next target response, thus improving their behavioral performance.
Predictive novels elicited further ERP abnormalities in the form of anomalous sustained early 50–200 ms and late 400–600 ms negativities over the lesioned lateral PFC (Fig. 5). These negativities were observed neither in response to unpredictive novels nor when predictive novels were displayed at the contralesional (bad) visual hemifield of patients. There are at least 2 alternative explanations for these abnormal sustained negativities. One, they could reflect volume-conducted compensatory activity from the intact PFC during online maintenance of the novel–target contingency. However, this is unlikely given their maximal intensity over the lesioned cortex. Two, these negativities could index signals from frontomedial (e.g., anterior cingulate) or else subcortical structures (e.g., striatum, midbrain) that could not be regulated by missing PFC representations. This latter hypothesis concurs with evidence that predictive novels were categorized at extremely short latencies, before visual information could reach PFC through classic flow from extrastriate pathways, and with the existence of well-defined prefrontotectal connectivity originally described in primates by Goldman and Nauta (1976). The hypothesis of a subcortical route for the updating of PFC representations of novel information concurs with the critical importance of modality nonspecific pathways for orienting to visual novelty (Sokolov 1963; Crick 1984; Johnston and Everling 2006), with extracortical influences on visual attention (Rafal and Posner 1987; Rafal et al. 1990), and with the disinhibition of subcortical reflexes in patients with cortical lesions (Jackson 1884). The extremely fast timing of these modulations could not be easily inferred from behavioral or metabolic brain imaging studies with a coarser temporal resolution or from ERP studies in humans without cortical lesions (cf., Fig. 5).
The acquisition of contextual predictability presumably required the building up of new sensorimotor associations between target representations and a novel memory chunk assisted by the intact PFC of patients (Fig. 1a,b), rather than any new linkage between posterior subordinate sensorimotor or perceptual units. The lesioned PFC could not efficiently update to a novel memory chunk, nor could it maintain this novel information online until the onset of the next target. This may explain the absence of contextual effects for novels and targets displayed at the contralesional visual hemifield of patients. In contrast, predictive novels flashed ipsilateral to lesion could be linked to familiar chunk I and led to the anticipation of the next target response, most likely with the support of subcortical structures (Edelman and Tononi 2000; Zink et al. 2003; O'Reilly 2006). In sum, the implicit learning of the temporal contingency between predictive novels and targets required the acquisition and online maintenance of a PFC representation of the novel event and the linkage of this novel memory chunk (Fig. 1b) with the neural representation of the forthcoming target stimulus (Fig. 1a).
Impaired Inhibition of Novel Distracters in PFC Patients
The patients showed problems inhibiting attentional capture by all novel distracters, most apparent in response to ipsilesion targets following an unpredictive novel event. Behavioral distractibility was independent of the visual hemifield of novel display, consistent with bilaterally disrupted brain responses to all novel stimulation described in the previous section. This finding bears 2 corollaries: 1) the effect was not related to the retinotopy of geniculostriate pathways or to any perceptual representation of the novel–target contingency at visual cortices and 2) this distractibility was sensitive to top-down (predictability) rather than bottom-up (visual field) contextual manipulations suggesting a superordinate locus of this effect (cf., Friston 2005) and consistent with a disrupted updating of superordinate representations of novel information (Fig. 1b).
The guided-activation model predicts that damage to a PFC representation of context may either impair or improve behavior depending on whether context helps or hurts performance, respectively (Miller and Cohen 2001; Braver et al. 2002). Accordingly, we found relative impairments versus improvements in contralesional versus ipsilesional visual discriminations under predictable versus unpredictable task contexts, respectively. However, our ERP data argued against a unique PFC locus for these contextual effects. The hemispheric asymmetries in frontally distributed transient P2 and novelty P3 activity to novel distracters and familiar targets suggest an interaction between PFC and other—possibly subcortical—structures as a function of the relative novelty or familiarity of task information (Fig. 7; Zink et al. 2003; McHaffie et al. 2005; O‘Reilly 2006). For instance, the absence of any target P2 asymmetry, together with the comparatively larger target P3b responses to familiar targets recorded over the lesioned PFC, supports a subcortical locus for these effects (Fig. 7, Targets). This ERP evidence could reflect an abnormal disinhibition of the ipsilesional striatum during processing of familiar targets in PFC-damaged patients (e.g., unit s2–r1 in Fig. 1a; cf., O’Reilly 2006). The purported role of basal ganglia in holding familiar information online could also account for the preserved visual discrimination ability of monkeys with bilateral PFC resections in the absence of distracters (Malmo 1942).
Impaired Selection of Familiar Targets in PFC Patients
A deficit in the selection of familiar information in patients was observed in response to all contralesion targets (cf., Barcelo et al. 2000; Yago et al. 2004). This target selection deficit matched with reduced P3b activity to targets displayed contralesionally—but not ipsilesionally—, and only in the unpredictable—but not the predictable—task contexts (Figs 4 and 6). In line with the guided-activation model, the onset of a predictive novel would quickly activate the sensorimotor units necessary for selecting the next target response (e.g., s2–r1 in Fig. 1b). When the next predictable target came up on display, its sensorimotor representation could be readily implemented at posterior association cortices resulting in normal target P3b elicitation. Hence, the selection of the target response in the predictable context could be implemented based primarily on extremely fast top-down control and without any fine-grained perceptual analysis at extrastriate cortices. The acceptable discrimination ability of PFC patients in the predictable context, even at their bad visual hemifield, indicates that the updating of sensorimotor target representations relied mainly on subcortical and/or posterior cortical structures (O'Reilly 2006).
This situation changed radically in an unpredictable context, where the absence of advance information about the identity of the next stimulus forced the brain to carry out a detailed analysis of perceptually similar targets and standard distracters. Contextual uncertainty forced the brain to rely primarily on sensory-driven control for the selection of the next action. However, stimulus identification at the ipsilesion visual cortex was impaired due to loss of intrahemispheric PFC–dependent modulatory input, resulting in faulty perceptual categorization of familiar information (e.g., units s1 and s2 in Fig. 1a). A defective perceptual analysis at ipsilesional visual cortices impaired the updating of sensorimotor units at posterior association cortices (Barcelo et al. 2000; Yago et al. 2004). Therefore, contextual uncertainty forced the neural system to implement familiar chunk I based on missing or incomplete perceptual information. The result was elicitation of reduced target P3b activity in a situation with maximal stimulus uncertainty and reduced cognitive control. This visual selective attention deficit could be likened to thalamic selection deficits described by Rafal and Posner (1987), based on the notion of a thalamic link between cortical visual attention and pattern recognition systems during the exploitation of familiar information (Crick 1984; Friston 2005). These results disclose a functional difference between the role of target P3b activations in the selection of familiar target information, in contrast with novelty P3 activations, and contrary to predictions from the context-updating model (cf., Donchin and Coles 1988). The absence of any predictability effects upon target P3b amplitudes in controls agrees with recent P3 results from task-switching studies (Barcelo et al. 2002, 2006), as well as with our information-theoretic analyses showing that targets conveyed the same information for response selection in both our predictable and unpredictable task conditions (see the Appendix in the Supplementary Material).
|
Summary of Findings and Conclusions |
|---|
|
TopAbstractIntroductionResultsDiscussionSummary of Findings and...Supplementary MaterialFundingReferences |
|---|
The present results partly support predictions from the guided-activation model that damage to one single PFC representation of context causes 3 different deficits in 1) the online maintenance of novel information, 2) the inhibition of novel distracters, and 3) the selection of familiar target information (Miller and Cohen 2001
). The present findings are consistent with a functional hierarchy of neural representations involving anterior and posterior multimodal association cortices (Fuster 1998; Koechlin et al. 2003) that implement cognitive control in cooperation with subcortical structures (Jackson 1884; Malmo 1942; Edelman and Tononi 2000). The results also support the notion that PFC neurons do not merely process information about a stimulus or its perceptual context, but rather about the temporal context of our goal-directed behavior (Tononi and Edelman 1997; Miller and Cohen 2001).
Lateral PFC played a crucial role in updating contextually novel information, whereas subcortical and/or posterior cortical structures seemed to have a larger contribution during the online maintenance and updating of familiar information (McHaffie et al. 2005; O'Reilly 2006). The present findings are consistent with ERP evidence from simple perceptual judgment tasks in PFC patients (Knight 1984, 1997), as well as with ERP evidence from task-cueing paradigms in healthy volunteers (Barcelo et al. 2002, 2006), suggesting activation of functionally distinct neural networks during the exploration of novel versus the exploitation of familiar information (Fig. 1a,b, Daw et al. 2006).
The early timing and scalp distribution of abnormal frontal negativities supports a subcortical involvement in computing the contextual predictability of novel stimulation (Rafal et al. 1990). Originally described in primates by Goldman-Rakic and Nauta (1976), prefrontotectal pathways could convey sufficient sensory information to lateral PFC for the categorization of a novel, unexpected, or biologically salient visual stimulus (Gaymard et al. 2003; Redgrave and Gurney 2006). Efferent cortical–subcortical connections from brain regions associated with expectation and timing, like PFC and the basal ganglia, offer a potential circuit for the rapid detection of unexpected sensory signals (
50 ms), as enough information can be conveyed through this route to detect a mismatch between visual input and active PFC representations (i.e., a change in luminance and/or spatial location; Johnston and Everling 2006; Redgrave and Gurney 2006). The observed extra-PFC signals in response to predictive novels support a nonrepresentational nature of contextual memories that could be best understood in terms of mutually informed cortical–subcortical dynamics in line with global mapping theories of perceptual categorization (Edelman and Tononi 2000, p. 93–101).
The present ERP findings relied on information-theoretic estimations—rather than on mean stimulus probabilities—of the mutual information between contextually related task events in order to explore the dimensionless properties of human working memory regardless of specific sensory or motor demands in target and distracter trials (Miller 1956; Koechlin and Summerfield 2007). These analyses indicate that the information conveyed by a stimulus for response selection depends on the intrinsic interactions between exogenous (e.g., sensory features) and endogenous (e.g., recent memories and future goals) sources of information along a functional hierarchy of neural representations in the brain. In the present task, the contextual information of a predictive novel stimulus could be best described in terms of large-scale cortical–subcortical dynamics (Tononi and Edelman 1997; Edelman and Tononi 2000; Friston 2005).
Finally, the disrupted brain responses to all novel information in PFC patients lent support to the hypothesis of lateral PFC as a "switch operator" that flexibly connects activation between exploitative and exploratory modes for processing familiar and novel information at the highest level in the hierarchy of cognitive control (Miller and Cohen 2001; Daw et al. 2006). Cognitive deficits in PFC patients can be easily overlooked by neuropsychological assessment methods that use familiar and repetitive material predictably presented at central vision, thus allowing patients with unilateral PFC damage to compensate for their deficits (Knight 1997). Information theory tools could help us gauge stimulus and task uncertainty in order to assess the degree and quality of the information-processing deficits in PFC patients (Miller 1956; Koechlin and Summerfield 2007).
|
Supplementary Material |
|---|
|
TopAbstractIntroductionResultsDiscussionSummary of Findings and...Supplementary MaterialFundingReferences |
|---|
Supplementary material can be found at: http://www.cercor.oxfordjournals.org/.
|
Funding |
|---|
|
TopAbstractIntroductionResultsDiscussionSummary of Findings and...Supplementary MaterialFundingReferences |
|---|
Ministerio de Educación y Ciencia (PR2006-0210), DG d'R+D+I Govern Balear (PRIB-2004-10136) to F.B.; National Institute of Neurological Disorders and Stroke (NS21135, PO40813) to R.T.K.
|
Acknowledgments |
|---|
Special thanks to Clay Clayworth and Donatella Scabini for their assistance in all phases of this work. Conflict of Interest: None declared.
|
References |
|---|
|
TopAbstractIntroductionResultsDiscussionSummary of Findings and...Supplementary MaterialFundingReferences |
|---|
Barcelo F, Escera C, Corral MJ, Perianez JA. Task switching and novelty processing activate a common neural network for cognitive control. J Cogn Neurosci (2006) 18:1734–1748.
Barcelo F, Periañez JA, Knight RT. Think differently: a brain orienting response to task novelty. Neuroreport (2002) 13:1887–1892.[CrossRef][ISI][Medline]
Barcelo F, Suwazono S, Knight RT. Prefrontal modulation of visual processing in humans. Nat Neurosci (2000) 3:399–403.[CrossRef][ISI][Medline]
Braver TS, Cohen J, Barch DM. The role of prefrontal cortex in normal and disordered cognitive control: a cognitive neuroscience perspective. In: Principles of frontal lobe function—Stuss DT, Knight RT, eds. (2002) New York: Oxford University Press. 428–447.
Cohen JD, Braver TS, O'Reilly RC. A computational approach to prefrontal cortex, cognitive control and schizophrenia: recent developments and current challenges. Philos Trans R Soc Lond B Biol Sci (1996) 351:1515–1527.[ISI][Medline]
Crick F. Function of the thalamic reticular complex: the searchlight hypothesis. Proc Natl Acad Sci USA (1984) 81:4586–4590.
Daw ND, O'Doherty JP, Dayan P, Seymour B, Dolan RJ. Cortical substrates for exploratory decisions in humans. Nature (2006) 441:876–879.[CrossRef][Medline]
Dien J, Spencer KM, Donchin E. Parsing the late positive complex: mental chronometry and the ERP components that inhabit the neighborhood of the P300. Psychophysiology (2004) 41:665–678.[CrossRef][ISI][Medline]
Donchin E. Surprise! Surprise? Psychophysiology (1981) 18:493–513.[ISI][Medline]
Donchin E, Coles MGH. Is the P300 component a manifestation of context updating? Behav Brain Sci (1988) 11:343–356.[ISI]
Edelman GM, Tononi G. Consciousness. In: How matter becomes imagination (2000) London: Penguin Books.
Friston K. A theory of cortical responses. Phil Trans R Soc B (2005) 360:815–836.[CrossRef][Medline]
Fuster JM. Linkage at the top. Neuron (1998) 21:1223–1229.[