|Home | About | Journals | Submit | Contact Us | Français|
Does the sustained, elevated neural activity observed during working memory (WM) tasks reflect the short-term retention of information? Functional magnetic resonance imaging (fMRI) data of delayed-recognition of visual motion in human participants were analyzed with two methods, a general linear model (GLM) and multi-voxel pattern analysis (MVPA). Although the GLM identified sustained, elevated delay-period activity in superior and lateral frontal cortex and in intraparietal sulcus (IPS), pattern classifiers were unable to recover trial-specific stimulus information from these delay-active regions. The converse—no sustained, elevated delay-period activity, but successful classification of trial-specific stimulus information—was true of posterior visual regions, including MT+ and calcarine and pericalcarine cortex. In contrast to stimulus information, pattern classifiers were able to extract trial-specific task instruction-related information from frontal and parietal areas showing elevated delay-period activity. Thus, the elevated delay-period activity that is measured with fMRI may reflect processes other than the storage, per se, of trial-specific stimulus information. It may be that the short-term storage of stimulus information is represented in patterns of (statistically) “subthreshold” activity distributed across regions of low-level sensory cortex that univariate methods cannot detect.
For decades, a prevailing view has been that working memory (WM) storage is accomplished via sustained, elevated neural activity. Such activity, first identified with extracellular recordings in the non-human primate (Fuster and Alexander, 1971; Niki, 1974; Funahashi et al., 1989), has been observed in numerous areas of the human brain with functional magnetic resonance imaging (fMRI; Curtis and D’Esposito, 2003). The idea of a role for delay-period activity in storage is reinforced by its sensitivity to manipulation of memory-influencing factors, such as persistence across varying delay lengths, and variation of signal magnitude with memory load (Vogel and Machizawa, 2004; Xu and Chun, 2006; Postle, 2006).
There are, however, complications with the mnemonic interpretation of delay-period activity. One is cases of activity that appear mnemonic, but can be shown to support other functions, such as attention or response preparation (Fuster, 2002; Lebedev et al., 2004). Relatedly, lesion-induced deficits originally interpreted as mnemonic (Jacobson, 1936; Funahashi et al., 1993) have subsequently been reinterpreted as reflecting factors other than memory per se (Malmo, 1942; Tsujimoto and Postle, 2012).
A second complication is that delay-period activity can fail to show properties thought necessary for a mnemonic signal. In one such case, although fMRI activity at several sites was elevated throughout a long delay period (24 seconds), none showed load sensitivity, leaving uncertain these regions contribution to storage (Jha and McCarthy, 2000). In another, monkeys showed excellent short-term memory (STM) for direction of motion despite the absence of directionally tuned neurons in either the middle temporal area (MT) or the prefrontal cortex (PFC) that sustained elevated activity across the delay (Bisley et al., 2004; Zaksas and Pasternak, 2006; Hussar and Pasternak, 2012).
A third complication relates to assumptions of homogeneity of function, often only implied or tacit, but nonetheless inherent, in massively univariate analyses of neuroimaging data. When activity is identified in a large volume of “activated” tissue, the extraction of a spatially averaged signal from contiguous voxels necessarily assumes that all are “doing the same thing”. Further, its interpretation entails assuming that this locally homogeneous activity can be construed as supporting a mental function independent of other brain areas. These assumptions, however, are difficult to reconcile with the increasingly common recognition that neural representations are high-dimensional and supported by anatomically distributed, dynamic computations (Norman et al., 2006; Kriegeskorte et al., 2006; Bullmore and Sporns, 2009; Cohen, 2011).
Given these complications, we sought to test core assumptions about elevated delay-period fMRI activity using an information-based analysis. Multivariate pattern classification was used to test two hypotheses about STM for visual motion information: 1) that elevated delay-period activity carries trial-specific stimulus information, and 2) that trial-specific stimulus information can be encoded in subthreshold patterns of activity. The first tests an assumption that has underlain most neuroimaging research on WM. Confirmation of the second would extend findings for visual STM for oriented gratings (Serences et al., 2009; Harrison and Tong, 2009) and offer insight about the physiological basis of STM for visual motion (Hussar and Pasternak, 2012).
To test our hypotheses we scanned subjects (fMRI) while they performed a delayed-recognition task for visual motion (Fig. 1). We then trained pattern classifiers to discriminate the direction of motion from individual time points in the trial and tested to see how these classifiers labeled the data from all the other time points during the trial (creating a decoding time series). A similar approach has been shown to be sensitive to the dynamics of memory content, capturing changes in the memory trace on a continuous basis throughout the trial (Polyn et al., 2005; Lewis-Peacock and Postle, 2008, 2012).
In contrast to analysis methods that assess the magnitude of activity (whether the firing rate of a neuron or the strength of the blood-oxygen level-dependent (BOLD) signal of one or a group of voxels), information-based analyses, such as pattern classification, focus on obtaining quantitative measures of the information content within a given area (Kriegeskorte et al., 2006; Kriegeskorte, 2011). This approach, often termed multi-voxel pattern analysis (MVPA), uses machine learning methods to identify patterns of activity that are reliably associated with different stimuli or categories of stimuli. The extent to which novel patterns of activity can then be correctly categorized provides a measure of the information available in the underlying voxels (Norman et al., 2006; Haynes, 2011; Jimura and Poldrack, 2012).
Ten volunteers (5 females) between 21 and 28 years of age (mean 23.8 years) were recruited from the undergraduate and graduate student community of the University of Wisconsin-Madison and paid for their participation. All subjects had normal or corrected-to-normal vision, no reported history of neurological disease, and no other contraindications for MRI. All subjects gave written informed consent according to the procedures approved by the Health Sciences Institutional Review Board at the University of Wisconsin-Madison. Three subjects (1 female) were excluded from the Results due to failure to perform the task to criterion level (see Results for more details).
Participants were scanned while performing seven runs of a delayed-recognition task for visual motion. A schematic representation of the experiment is illustrated in Fig. 1. Each trial began with a fixation cross changing color to white, indicating to subjects that they needed to fixate the cross and prepare for the start of the trial. After 1.5 s a patch of coherently moving dots was presented (sample, 1 s). Participants were instructed to remember both the direction and the speed of this motion while maintaining fixation for the duration of the trial. The sample was followed by a 15 s delay-period. Seven seconds into the delay period the fixation cross changed color to indicate the dimension on which subjects would be making a match/non-match comparison judgement between the remembered motion and a new set of moving dots; blue indicated they should make the judgement based on the direction of the moving dots, ignoring speed, while magenta indicated that they should make the judgement based on speed, ignoring direction. Following the delay, a second patch of coherently moving dots (probe, 1 s) was presented and subjects were required to indicate as quickly as possible with one of two buttons if the probe motion matched or did not match the sample on the cued dimension. After a 1 s response period the fixation cross changed color to green if the subjects were correct or red if they were incorrect (feedback, 2 s). A 10 s intertrial interval (ITI) followed, during which the fixation cross changed color to gray and subjects were instructed they could break fixation and relax their eyes.
Sample and probe stimuli consisted of circular patches (15° diameter) of coherently moving dots. Sample motion could be in one of four directions (42°, 132°, 222°, 312°) and at one of three speeds (4°/s, 8°/s, 12°/s). Directions were chosen to be off the cardinal axes to reduce potential verbalizations. Probe stimuli on match trials had the same value in the cued dimension, and a different value from the non-cued dimension (e.g., for a match trial where direction was cued the sample might have moved toward 132° at 8°/s, while the matching probe might have moved toward 132° at 4°/s). On direction-cued non-match trials the probe stimulus was rotated (randomly clockwise or counter-clockwise) by a degree threshold value from the direction of sample and the speed was randomly drawn from one of the two values not used in the sample. On speed-cued non-match trials the probe stimulus speed was changed (randomly increased or decreased) by a proportional threshold value from the speed of the sample and the direction was randomly drawn from one of the three values not used in the sample. The threshold values for non-match direction and speed were updated using separate adaptive staircases (Levitt, 1971) to keep performance around 75% correct. The logic of this approach was that by holding task difficulty constant throughout an experimental session, we could assume comparable fidelity of representation on trials for which the response ended up being correct versus incorrect. This would maximize sensitivity by allowing inclusion of all trials in the analyses. Before scanning, subjects practiced a block of 24 trials to familiarize themselves with the experimental procedure and to determine starting threshold values for non-match trials.
Participants completed 168 trials over the course of 7 runs while in the scanner. Sample stimuli included all possible pairwise combinations of directions and speeds (12 different combinations, each seen 14 times). Direction was cued on 96 trials, with the remaining 72 cued for speed. This disparity balanced the overall number of individual examples at each dimension value across the cued dimensions. Match and non-match trials were equally likely. The experimental stimuli were controlled by E-Prime 2.0 (Psychology Software Tools, Inc., PA) and viewed through fiber-optic goggles mounted on the head coil (Avotec Visible Eye, Stuart, FL). Participants responded via two buttons on a fiber-optic button box (Psychology Software Tools, Inc., PA).
Subjects also performed one block of an MT+ localalizer task, similar to that used by Huk et al. (2002). In summary, participants viewed alternating 18 second blocks of stationary and moving dot patterns (alternating from expanding and contracting every second) within a circular aperture (15°) while maintaining fixation, for a total of 8 blocks of each.
Whole brain images were acquired with the 3T scanner (Discovery MR750; GE Healthcare) at the Lane Neuroimaging Laboratory at the University of Wisconsin-Madison. For all subjects a high-resolution T1-weighted image was acquired with an FSPGR sequence (8.132 ms TR, 3.18 ms TE, 12° flip angle,156 axial slices, 256 × 256 in-plane, 1.0 mm isotropic). A gradient-echo, echo-planar sequence (2 s TR, 25 ms TE) was used to acquire data sensitive to the BOLD signal within a 64 × 64 matrix (39 sagittal slices, 3.5 mm isotropic). Seven runs of the delayed-recognition task were obtained for each subject, each lasting 12 min, 8 sec (364 volumes). All task runs were preceded by 10 sec of dummy pulses to achieve a steady state of tissue magnetization. One run of the MT+ localizer was obtained for each subject, lasting 4.8 minutes (144 volumes).
The functional data were preprocessed using the Analysis of Functional NeuroImages (AFNI) software package (Cox, 1996). All volumes were spatially aligned to the final volume of the final run using a rigid-body realignment and corrected for slice time acquisition. Linear, quadratic and cubic trends were removed from each run to reduce the influence of scanner drift. For univariate analyses, data were spatially smoothed with a 6 mm FWHM Gaussian kernel and transformed into Talairach space (Talairach and Tournoux, 1988). For classification analyses, data were z-scored separately within run for each voxel. Data were not smoothed and were left in their native space.
Each within-trial event of the delayed-recognition task (i.e., sample, pre-cue delay, cue, post-cue delay, probe, response) was modeled separately for direction- and speed- cued trials. Sample and probe were modeled as 1 s boxcars, each delay as a boxcar of appropriate duration, and the cue and response as impulses. All were convolved with a canonical hemodynamic response function (HRF). Each of these independent regressors was entered into a modified general linear model (GLM) for analysis using AFNI. For the present purposes, a more generous boxcar-shaped covariate was used to model the delay-periods (rather than a more conservative mid-delay delta function (Zarahn et al., 1999; Postle et al., 2000)), to ensure that we would not miss any delay-active voxels, although at the expense of likely also being sensitive to some variance that is attributable to the sample presentation. In this way, we implemented a generous feature selection step that included as many “delay active” voxels as possible for MVPA, thereby being careful not to exclude any such voxels that may potentially carry stimulus-specific information. The localizer was modeled with boxcars for both stationary and moving dot patterns. This localizer was used to insure that the regions of interest used for the MVPA analysis included MT+.
Classification was performed using the Princeton Multi-Voxel Pattern Analysis (www.pni.princeton.edu/mvpa) toolbox and custom routines in MATLAB (The MathWorks, Natick, MA, USA). Preprocessed fMRI data from individual trial time points were used to train separate classifiers to classify the direction of motion (4 possible directions) or the speed of motion (3 possible speeds) in the sample and by inference, the direction/speed of motion in memory (stimulus-specific classification), or to classify whether the subject had been cued that direction or speed was the relevant dimension on a given trial (trial-dimension classification).
Classification was accomplished using L2-regularized logistic regression (LR), a linear classification approach used widely in the machine learning community and well suited for application to fMRI data as it tends to generalize well after learning in high-dimensional feature spaces with limited training examples (Pereira et al., 2009). The lambda penalty term, which reduces the contribution of less informative voxels to classification and thus improves generalization, was determined (λ = 25) by repeating the wholebrain testing procedure described below for penalty terms at powers of 10 from −4 to 4, then at a finer grained resolution within the best interval. The penalty term was chosen to maximize the mean decoding performance across all subjects. During decoding, a trial was considered correctly classified if the correct direction/speed had the highest likelihood estimate (winner-take-all classification). Overall classification accuracy was determined using leave-one-trial-out cross-validation, in which the classifier was repeatedly trained on data from all but one trial, and then tested on the left out trial, rotating through all the trials as the left out testing trial.
Stimulus-specific classifiers were always trained with data from direction-cued trials or speed-cued trials, never both. Testing of trained classifiers was done on trials of both type. This allowed us to compare how representations changed when subjects were cued that they would be judging speed, and thus direction was no longer relevant, and vice versa. Trial-dimension classifiers were trained and tested on all data.
In order to examine the dynamics of the memory trace, each classifier was trained using data from only a single time point in the trial (e.g., the first volume acquisition following the target) and then tested on all time points in the left-out trial (i.e., including both time points preceding and following the training time point). The result of this procedure is a time course of decoding accuracy for the entire trial. By doing the initial training of the classifier using different time points in the trial (e.g., a time point just after the sample, a time point in the later part of the delay, etc.), it was possible to estimate the stability of a given representation throughout the duration of the trial.
Classification was initially performed on whole brain data that had undergone a basic feature-selection step whereby only those voxels that showed a main effect for task (t > 2) in the univariate GLM were included. This step was included to reduce the chances of overfitting during training. Subsequent Region of Interest (ROI)-based analyses used only those voxels within individual ROIs, created from the intersection of anatomically defined ROIs and voxels that showed either significant sustained delay-period activity or no delay-period activity, depending on the specific hypothesis. Four anatomically defined ROIs were hand drawn for each subject by tracing gray matter on the high-resolution anatomical scans: frontal, parietal, lateral occipital and temporal, and medial occipital. The frontal region included the entire precentral sulcus (PCS) and the posterior portion of the inferior frontal sulcus (IFS). For finer-grained analyses, this was subdivided into three frontal ROIs that showed robust delay-period activity: the superior rostral bank of the PCS bounded superiorly by the intersection of the superior frontal sulcus (SFS); a more inferior portion of the rostral bank of the PCS bounded ventrally by the intersection with the IFS; and the caudal third of the IFS. The parietal region included the entire intraparietal sulcus (IPS) and the superior parietal lobule (SPL). For finer-grained analyses, it was also subdivided into three ROIs: medial-caudal IPS comprising the descending segment and the caudal half of the horizontal segment, dorsolateral IPS comprising the rostral half of the horizontal segment and the ascending segment, and SPL. The lateral occipital and temporal region included all of the lateral occipital gyrus, the fusiform gyrus, the posterior portion of the middle and inferior temporal gyri and the posterior portion of the inferior temporal sulcus. The medial occipital region covered the medial portion of the occipital lobe from the lingual sulcus to the occipitoparietal sulcus, including all of the calcarine sulcus.
The significance of classifier performance was determined using a random permutation test (Golland and Fischl, 2003) to determine the likelihood of observing a specific accuracy under the null hypothesis that there is no relationship between the data and the specific class labels used to train the classifier (directions/speeds of motion). A null distribution was generated by rerunning the entire classification cross-validation procedure 1000 times, randomly shuffling the class labels each time. A p-value was then computed by determining the proportion of permuted accuracies that were higher than the observed accuracy. This procedure was repeated for all classification results.
Task difficulty was equated across subjects by feeding real-time, trial-by-trial performance information to a staircasing algorithm that dynamically adjusted the difficulty of non-matching probe stimuli (direction and speed independently) so as to maintain performance at a target level of 75% correct. Overall mean behavioral performance (n = 7) was 80.38% correct for direction trials and 80.32% correct for speed trials, both slightly better than the target performance level because several subjects reached a predefined minimum threshold value, at which point the staircase procedure could no longer reduce the threshold to further increase task difficulty. The average non-match direction threshold was 10.9° (SD 5.33°). The average non-match speed threshold was a 38% change (SD 15.3%). Three subjects were dropped from the study because their inordinately high thresholds indicated that they were responding at random.
To test the first hypothesis, that elevated delay-period activity carries trial-specific stimulus information and can thus be interpreted as a neural correlate of memory storage, we first identified areas showing elevated, sustained delay-period activity by solving a massively univariate GLM for each subject. The resultant individual thresholded statistical maps were then used, on a subject-by-subject basis, to select the voxels used for training the classifiers. Group-level statistical maps for the sample-evoked (Fig. 2A) and delay-period (Fig. 2B) activity illustrate several important characteristics. Activity evoked by the sample was widespread, located bilaterally in posterior visual areas, lateral occipitotemporal regions including the inferior temporal sulcus (including the putative MT+ complex identified with the localizer), IPS, posterior IFS and SFS, and PCS. Delay-period activity was more circumscribed, with clusters of significant activity (p < 0.05, FDR corrected) bilaterally in the IPS, IFS, and PCS. This pattern of elevated activity agrees with many other findings from studies of visuo-spatial WM (Curtis and D’Esposito, 2003).
Before directly testing our hypotheses, it was necessary to demonstrate that motion information could be decoded on a time point-by-time point basis. All analyses were performed on single-subject data, with statistical reliability subsequently assessed across the sample. For each subject’s data, a set of four classifiers was trained to discriminate the direction of the sample motion stimulus (and by inference, the remembered motion direction; 4 possible directions), each using feature-selected whole-brain data from direction-cued trials and restricted to a single time point within the trial. The first classifier (Sample) was trained on data from only the time point 4 seconds after the sample onset, which corresponds to the peak of the sample-evoked response. The second classifier (Late Delay) was trained on data from the time point 16 seconds after sample onset, corresponding to the post-cue portion of the delay period, prior to the onset of the probe stimulus. The third classifier (Probe) was trained on data from the time point 20 seconds after sample onset (4 seconds after probe onset), corresponding to the peak of the probe-evoked response. The final classifier (ITI) was trained on data from the time point 26 seconds after sample onset, corresponding to the middle of the intertrial interval. This classifier was included as a control, as we would not expect there to be any stimulus-specific information retained once the trial had been completed.
Each trained classifier was then used to construct a decoding time course of direction representation on every trial time point from the held-out direction-cued trials. This approach allowed us to detect changes in the neural representation of information across the trial. For example, by training a classifier on data from very early in the trial (i.e., Sample), we would expect to capture early, sensory-based representations. By testing such a classifier on every time point in the trial, we could determine if such a representation remains stable throughout the trial, or if it deteriorates over time. Similarly, training a classifier with data from a time point late in the delay-period (Late Delay) might capture a recoded representation (e.g. verbal or numeric or clock-face), which would be expected to be absent at the beginning of the trial and to strengthen over the course of the delay-period when we look at the time point-by-time point decoding performance of that classifier. By limiting our training data to a single time point during the trial, we hoped to maximize our ability to resolve time point-by-time point changes, at the expense of lower signal-to-noise for the classifier inputs.
Results from this analysis can be seen in Fig. 3A. The mean decoding accuracy for the Sample classifier was significantly above chance (25 %, p < 0.05, permutation test) throughout the entire delay period. Similar results were obtained with the Late Delay and Probe classifiers. These results suggest that the memory representation remains relatively stable and unchanging across the delay period. Decoding accuracies for the ITI classifier were always near chance, which was expected given the absence of any stimulus-related information for the classifier to learn (i.e., signal from 8 sec after the offset of the memory probe would not be expected to carry information about the stimulus from the preceding trial). As with the univariate analyses, all trials were included in the analyses. Follow-up analyses using only correct trials produced qualitatively similar results.
To assess the specificity of the classification, the set of classifiers trained to discriminate direction with data from direction-cued trials were also used to decode direction information from speed-cued trials. (Fig. 3B). Classification with the Sample, Late Delay and Probe classifiers was above chance for time points before the cue, when subjects needed to hold both speed and direction information in memory, but fell to chance levels following the cue, suggesting that subjects discarded direction information when it was no longer relevant to the current trial.
Two features in the data confirm that successful Late Delay decoding of direction-cued trials represents the sustained retention of stimulus information and does not reflect an artifact of the slow recovery of the hemodynamic response, nor effects of motion adaptation. First, if the results were purely driven by the residual hemodynamic response to the sample stimulus, or to adaptation, we would expect to see similar decoding performance on both direction-cued and speed-cued trials, because the sample stimuli are identical across these two conditions. Second, such “residual” effects would also result in successful decoding by the ITI classifier, which was trained with data 10 seconds following the probe stimulus (compared to 16 seconds following the sample for the Late Delay), which it clearly does not.
The whole-brain classification procedure was also applied to the stimulus dimension of speed, using data from only speed-cued trials. Unlike direction, however, classifier performance never exceeded chance for any time point during the trial. Because subjects performed these trials at the same level of proficiency as direction-cued trials, we interpret this null result to mean that the representation of speed, at least across the range used in this experiment, may be carried out on too fine a spatial scale neurally to be discriminated with our fMRI procedure. It is also possible that by collapsing across different directions when classifying the speed trials we added too much noise to the signal to classify, given the close relationship between speed and direction (Born and Bradley, 2005). Additionally, the use of only three speeds may have encouraged a coding strategy that varied over time (e.g., verbal labels that changed as the stimuli became more familiar to subjects). The remainder of the Results and Discussion will focus on decoding direction information.
To test our first hypothesis—that elevated delay-period activity carries stimulus-specific information—we repeated the classification procedure that we used with the whole brain data, but applied it to only those voxels in the frontal or parietal cortices that showed elevated delay-period activity as identified with the univariate analysis. As can be seen in the average BOLD time series in Fig. 4A and 4B, each of these areas showed elevated activity that was sustained throughout the delay period. However, decoding performance for motion direction never differed from chance in either area, regardless of the time point used to train the classifier (Fig. 4E and 4F). Therefore, we failed to find evidence that the sustained, delay-period BOLD activity in these regions carried stimulus-related information; the first hypothesis was not supported. To rule out the possibility that using such large ROIs may have obscured the presence of stimulus-representation in smaller regions, we repeated the ROI-classification with the smaller, more specific ROIs described in the Methods section. Results in all smaller ROIs mirrored those of the larger regions: no evidence for stimulus-related information was found in any of the regions.
To test our second hypothesis—that stimulus-specific information can be recovered from subthreshold patterns of activity (i.e., activity whose signal intensity does not surpass a statistical threshold in conventional univariate analysis)—we repeated the classification procedure as above, limiting the training data to only voxels in medial occipital or lateral occipital and temporal cortex that showed no evidence of elevated activity during the delay period in the GLM. As can be seen in the average BOLD time series in Fig. 4C and 4D, these regions showed large sample- and probe-evoked responses, as well as smaller cue-evoked responses, but no sustained, elevated delay-period activity. Decoding performance from these regions, however, was significantly above chance throughout the delay-period (Fig. 4G and 4H).
Although these posterior regions did not show sustained delay-period activity at the group level (Fig. 2B.), there were voxels in the individual-subject data of each subject that did. The results of pattern classification did not change appreciably when these voxels were included or excluded from the analyses. Additionally, when the classifiers were trained only on these posterior delay-active voxels, in no subject was decoding performance sustained at above-chance levels across the delay period. Overall, these findings were consistent with the second hypothesis, that brain regions can carry stimulus-specific information in a sustained manner despite the absence of sustained, above-baseline levels of activity.
These results form a double dissociation, with frontal and parietal regions showing elevated, sustained delay-period activity but no delay-period stimulus representation, and posterior regions the converse. One possible concern about applying these results to our understanding of WM storage, however, is that doing so requires the acceptance of the null MVPA findings in frontal and parietal cortex. Might it be the case, for example, that these regions are simply less amenable to MVPA (perhaps, e.g., because they represent information at a finer grain of spatial detail than is measurable with our fMRI methods)? To address this possibility, we trained a new set of classifiers on a different discrimination – cue identity (i.e., whether the relevant stimulus dimension was direction or speed) - and repeated the procedure with each of the four ROIs. As shown in Fig. 4I-L, the relevant trial-dimension was decodable from each of the four ROIs. Importantly, for each this was only true for the Late Delay and Probe time-point trained classifiers, and only for time points following cue presentation. These results indicate that frontal and parietal regions are not inherently “undecodeable” with our methods, and, thus, lend more credence to the possibility that they did not represent stimulus-specific information during the delay period of our WM task.
The aim of the present study was to test long-standing views about the relationship between the short-term retention of information and sustained delay-period activity. Using an information-based analysis approach with fMRI data collected during a delayed-recognition task for visual motion, we tested two hypotheses: first, that sustained, elevated delay-period activity carries stimulus-specific information; and second, that stimulus information can be encoded in distributed patterns of subthreshold activity. To test the first hypothesis we trained pattern classifiers with BOLD signal from frontal and parietal areas that showed sustained, elevated delay-period activity. We failed to find evidence that these voxels carried stimulus-specific information during the delay period. To test the second hypothesis, we applied the same procedure to BOLD data from posterior regions that showed robust responses to visual stimuli, but no elevated delay-period activity. The classifiers were successfully able to decode the remembered direction throughout the delay period, providing strong evidence in support of this hypothesis.
The first finding can be seen as a failure to support an enduring assumption in cognitive neuroscience, albeit one that is increasingly being called into question (Curtis and D’Esposito, 2003; Lebedev et al., 2004; Curtis and Lee, 2010; Lewis-Peacock and Postle, 2012). Although on its own it might be qualified as a null result, there are several factors that must influence its interpretation. Most saliently, it is paired with a positive result using the same method, and derived from statistically “subthreshold” voxels located in areas that are active during the perception of the to-be-remembered information. Empirical evidence thus shows that this method is sensitive. Indeed, although there remains some controversy about the physiological and representational factors that underlie the patterns of activity that correspond to different brain states (Freeman et al., 2011; Thompson et al., 2011), we are not familiar with any suggestion that there may exist brain states to which MVPA is less sensitive than traditional analysis of activation levels of individual voxels or groups of voxels. To the contrary, the near-consensus view is that MVPA methods are more sensitive than traditional activation-based analyses (Kriegeskorte et al., 2006; Norman et al., 2006; Haynes, 2011; Lewis-Peacock and Postle, 2012; Jimura and Poldrack, 2012).
Further, although we cannot rule out the possibility that stimulus information might be represented in frontoparietal cortex at either a spatial scale that is too fine to be detected with our fMRI methods, or perhaps via a signal to which BOLD is relatively insensitive (e.g., low-frequency oscillations in local field potentials), we did demonstrate that this is not a limitation for the decoding of trial-specific task instruction-related information. From this perspective our results are consistent with, for example, the finding from monkeys that PFC and posterior parietal cortex (PPC) represent the category to which a stimulus belongs (Freedman and Assad, 2006; Swaminathan and Freedman, 2012). It is also worthy of note that although MVPA has been applied successfully to sensory processing in topographically organized cortex (e.g., the decoding of orientation (Harrison and Tong, 2009; Serences et al., 2009)), it has also been successfully applied to “higher level” processing in polymodal cortex. Thus, for example, MVPA has demonstrated contextual reinstatement during episodic memory retrieval (Polyn et al., 2005), the recognition of individual faces (Kriegeskorte et al., 2007), and neural correlates of free choice (Soon et al., 2008), all entailing the decoding of information from polymodal temporal, parietal, and/or frontal cortex.
Consistent with our preferred interpretation of the null findings in frontal cortex are several factors. First, there are the results from extracellular recording in monkeys performing a similar task with similar stimuli, where no evidence for direction-selective persistent activity was found in the PFC throughout the delay period (Zaksas and Pasternak, 2006; Hussar and Pasternak, 2012). Second, a similar pattern to the MVPA results that we describe here has been reported for STM for four categories of visual objects (Linden et al., 2012) and for complex artificial visual stimuli (Christophel et al., Revision under review). Third, the fact that STM can be intact despite lesions of PFC (D’Esposito and Postle, 1999; Tsujimoto and Postle, 2012) is consistent with the failure to find physiological evidence for STM representations in this region.
The frontoparietal network that has been a focus of this study is known to support the endogenous control of attention (Corbetta and Shulman, 2002; Beck and Kastner, 2009; Noudoost et al., 2010). Interestingly, one account of WM storage is that it is supported by this same top-down mechanism (e.g., Armstrong et al., 2009; Curtis and D’Esposito, 2003; Postle, 2011). From this perspective, the sustained delay-period activity observed in this study may correspond to a control signal that does not vary with stimulus identity. Future work would need to reconcile this possibility with the finding that multivariate patterns of frontoparietal activity do discriminate between directions of motion during a sustained attention task (Liu et al., 2011). In addition to specifically memory-related functions, many other functions might be supported by sustained delay-period activity of frontal and parietal regions. Because the frontal and parietal activity observed in the present study (Fig. 4) resembles activity that has been reported in countless prior neuroimaging studies (Curtis and D’Esposito, 2003), it may well be that it does not correspond to a stimulus-specific or even task-specific function. More general demands that many cognitive tasks (including working memory tasks) have in common include decision making (Curtis and Lee, 2010), prioritizing certain task-relevant representations and/or processes over others (Miller and Cohen, 2001), monitoring the environment to control the processing of potentially interfering exogenous events (Chao and Knight, 1998; Postle, 2005), actively representing a “behavioral set” (Woolgar et al., 2011), and monitoring behavior so as to prevent prepotent responses (Knight and D’Esposito, 2003), including perseverative responses (Milner, 1963; Tsujimoto and Postle, 2012). (Note that although the behavioral set account might be consistent with the successful decoding of cue identity in frontal and parietal regions, this explanation does not generalize to the first portion of the delay period.) This is, of course, an incomplete list.
One important question for future study is the nature of the mental codes with which subjects represent motion information across the delay period. In the monkey, a psychophysical study using backward masking provided evidence that the initial memory trace is perceptually based, retaining a high-fidelity representation of the sample (including such trial-irrelevant information as the local velocity of individual dots in the random-dot motion stimulus). This representation only endured few hundred milliseconds into the delay period, however, perhaps because, in this study, the animals could predict the major features of the impending memory probe (Zaksas et al., 2001). Although the BOLD signal did not afford high temporal resolution in the present study, results with classifiers trained on different portions of the trial suggested that the mnemonic representation is relatively stable. We cannot know with certainty, however, whether this representation was primarily perceptual, motoric, or categorical in nature, or perhaps some combination of these. Our working assumption is that the mnemonic representation of direction was perceptually based, because it is from visual regions that we were able to recover stimulus direction information. Had subjects used, for example, a covert eye-movement strategy, we would have expected to have been able to decode stimulus information from frontal and parietal regions (Ikkai and Curtis, 2011). The same reasoning makes us skeptical that subjects depended on a verbal strategy for remembering either direction or speed. We did not, however, monitor eye movements, nor take steps to discourage covert speech.
The results presented here highlight the differing conclusions that can be drawn from activation- vs. information-based analyses of the same data set. In so doing, they raise questions about the longstanding belief that information retained during working memory is stored via sustained delay-period activity, preferentially in frontal and parietal cortex. Instead, the memory trace may be represented in patterns of “subthreshold” levels of activity distributed across regions of low-level sensory cortex.
This work was supported by the National Institutes of Health (R01-MH064498 to BRP).
The authors declare no conflicts of interest