Functional Microcircuit Recruited during Retrieval of Object Association Memory in Monkey Perirhinal Cortex

Neuron

Article Functional Microcircuit Recruited during Retrieval of Object Association Memory in Monkey Perirhinal Cortex Toshiyuki Hirabayashi,1 Daigo Takeuchi,1 Keita Tamura,1 and Yasushi Miyashita1,2,* 1Department

of Physiology, The University of Tokyo School of Medicine, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan JST, Kawaguchi, Saitama 332-0012, Japan *Correspondence: [email protected] http://dx.doi.org/10.1016/j.neuron.2012.10.031 2CREST,

SUMMARY

The primate temporal cortex implements neural mechanisms for memory retrieval from visual longterm storage, and memory neurons have been identified at the single-neuron level whose activities following cue presentation encode the presented object (‘‘cue-holding’’ neurons) or to-be-recalled target (‘‘pair-recall’’ neurons). Although these two types of neurons can potentially interact during the target recall, little is known about information flow among these neurons. We conducted simultaneous recordings of multiple single units in macaque perirhinal cortex while they performed a pair-association memory task. Granger causality analysis revealed the emergence of directed couplings during the delay period predominantly from cue-holding neurons to pair-recall neurons. Moreover, these interactions coincided with unidirectional signal flow from the recipient recall neuron to another recall neuron, implying cascade-like signal propagation among the memory cell assembly. These results suggest that directed interactions among perirhinal memory neurons are dynamically modulated to implement functional microcircuitry for retrieval of object association memory. INTRODUCTION The primate perirhinal cortex is situated within the medial temporal lobe memory system, where it receives information from the ventral visual pathway and participates in the storage and retrieval of visual long-term memory (Miyashita, 2004; Squire et al., 2007; Suzuki, 2009). A pair-association memory task requires subjects to memorize pair-wise relationships between two arbitrarily associated objects (Figure 1A, right), and has been utilized to clinically test human amnesia (Wechsler, 1987) and to reveal that perirhinal lesion impairs stimulusstimulus associative memory in monkeys (Murray et al., 1993; Buckley and Gaffan, 1998). Using this task (Figure 1A, left), previous studies have identified several types of perirhinal neurons that are active during retrieval of visual associative 192 Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc.

memory. One type exhibits sustained activity that continues from the presentation of a cue stimulus (‘‘cue-holding [CH] neurons’’) (Figure 1B, right, blue; Naya et al., 2003b), and another type exhibits gradually increasing activity toward the presentation of the paired associate (‘‘pair-recall [PR] neurons’’) (Figure 1B, right, red; Sakai and Miyashita, 1991; Naya et al., 2001, 2003b). The activity of pair-recall neurons was shown to dynamically appear or disappear on demand for retrieval of a learned paired-associate (Naya et al., 1996). These two types of neurons have been separately identified at the single-neuron level, but the functional microcircuitry and cell-to-cell information flow during retrieval of visual associative memory still remain to be elucidated. An intriguing hypothesis has been proposed that directional couplings from CH cells to PR cells would emerge during the delay period to form a cell assembly representing the sought target (Miyashita, 2004), just as a ‘‘phase sequence,’’ which is thought to underlie various cognitive processes (Hebb, 1949; Harris, 2005; Buzsa´ki, 2010). The present study tested this directional neuronal coupling during memory retrieval. Previous functional microcircuit studies have shown some ‘‘connection rules’’ in primary sensory areas using cross-correlation between spike trains (Alonso and Martinez, 1998; Menz and Freeman, 2003; Atencio and Schreiner, 2010). Several studies have applied coherence analysis to spike trains instead of the cross-correlation analysis to reveal the frequency structure of neuronal communications, and the gamma coherence in particular has been observed for local interactions between neurons (Lee, 2003; Fries et al., 2008; Sirota et al., 2008; Zhou et al., 2008; Hirabayashi et al., 2010; Lima et al., 2010). Granger causality has also been utilized as a tool for investigating the directionality of couplings between continuous signals including local field potentials (LFPs) and electroencephalograms (EEGs) ski and Blinowska, 1991; Brovelli et al., 2004; Gregoriou (Kamin et al., 2009; Verhoef et al., 2011) or between spike trains (Nedungadi et al., 2009; Kim et al., 2011; Quinn et al., 2011). Using the Granger causality analysis, we previously demonstrated that pairs of inferior temporal neurons exhibit directed interactions in the gamma frequency range during stimulus presentation in a visual discrimination task (Hirabayashi et al., 2010). In the present study, we conducted simultaneous recordings of multiple single units in macaque perirhinal cortex, while they performed a pair-association memory task. Granger causality analysis revealed the emergence of directed couplings during the delay period predominantly from CH neurons to PR neurons.

Neuron Microcircuit Dynamics in IT Cortex during Recall

Figure 1. Granger Causality between Cue-Holding and Pair-Recall Neurons (A) Sequence of a pair association memory task (left) and a set of stimulus pairs (right, for monkey 2). Monkeys had to retrieve a learned paired-associate in response to the presented cue stimulus. (B) (Left) Lateral and coronal views of a monkey brain. Scale bar, 10 mm. Multiple single units were simultaneously recorded in area 36 (gray) (see Figure S1). Dashed line, anteroposterior level of the coronal view. rs, rhinal sulcus. amts, anterior middle temporal sulcus. (Right) Schematic drawings of the spiking activity of cue-holding (CH, blue) and pair-recall (PR, red) neurons. Gray background, cue and choice periods. (C–F) Representative data sets. (C) Autocorrelograms and peristimulus time histograms of simultaneously recorded CH neuron (top, unit 1) and PR neuron (bottom, unit 2). Black and gray traces, response to optimal and worst stimuli, respectively. Horizontal black bars, cue period. (D) Spectral dynamics of the Granger causality between units 1 and 2 in response to the optimal cue. Vertical white lines, cue onset and offset. Causal influence was calculated for raw spike trains. Note that the choice period was not included in the plot. (E and F) Same as (C) and (D) but for another CH-PR cell pair.

This directional bias of couplings was also observed in crosscorrelograms. Moreover, these interactions coincided with further signal flow directed to another PR neuron, suggesting cascade-like signal propagation among the memory cell assembly. These results suggest that directionally defined patterns of interactions among different classes of memory neurons implement a functional microcircuit for retrieval of visual associative long-term memory. RESULTS Two monkeys were trained to perform a visual pair-association memory task, in which they had to retrieve a learned paired-

associate in response to the presented cue stimulus (Sakai and Miyashita, 1991; Naya et al., 2003a, 2003b; Takeuchi et al., 2011; Figure 1A). We then inserted a multicontact electrode (Hirabayashi and Miyashita, 2005: Hirabayashi et al., 2010; Ohiorhenuan et al., 2010) into area 36 of the perirhinal cortex to record multiple single-units (Figure 1B, left; see Figure S1 available online). Neuronal activities were classified as ‘‘cue-holding’’ (CH) (Naya et al., 2003b) when a given neuron showed cue activity and subsequent decreasing delay activity selectively in response to a given stimulus (Figure 1B, right, blue), or as ‘‘pair-recall’’ (PR) (Sakai and Miyashita, 1991; Naya et al., 2001, 2003b) when a given neuron exhibited increasing delay activity selectively in response to a given stimulus, as Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc. 193

Neuron Microcircuit Dynamics in IT Cortex during Recall

Figure 2. Population Granger Causality between CH and PR Cells (A) Spectral dynamics of population Granger causality in the direction from CH to PR (left) and PR to CH (right). Causal influence was calculated for raw spike trains. The choice period was not included in the plot. Time courses of spike firings for both CHs and PRs were depicted below. (B) Time courses of Granger causality in the gamma (left) and alpha/beta (right) frequency ranges. Trialshifted control was subtracted. ++: p < 0.002, comparison with zero, three-way repeated-measures ANOVA followed by LSD. **: p < 0.001. *: p < 0.006. Error bars, SEM. See also Figures S2–S4. (C) Time courses of coherence in the gamma (left) and alpha/beta (right) frequency ranges. Trialshifted control was subtracted. +: p < 0.02, comparison with zero, two-way repeated-measures ANOVA followed by LSD. *: p < 0.02. Error bars, SEM. (D) Proportions of cell pairs showing a significantly directional causal influence. **: p < 0.002, c2 test. (E) Pair-recall indices for putative pre and postunits of significantly directional CH-PR pairs. Blue and red lines, individual cell pairs showing higher PRI values for pre and postunits, respectively. Black line, mean values. *: p < 0.009, paired t test. Error bars, SEM.

well as selective cue activity in response to its paired-associate (Figure 1B, right, red; see Experimental Procedures). Figures 1C–1F show two examples of simultaneously recorded pairs of CH neuron (unit 1) (Figures 1C and 1E, top) and PR neuron (unit 2) (Figures 1C and 1E, bottom). When the optimal stimulus for each cell pair was presented, both cell pairs exhibited prominent gamma Granger causality during the delay period predominantly in the direction from unit 1 to 2, i.e., from CH neuron to PR neuron (Figures 1D and 1F). In total, 180 single units were recorded from two monkeys in 39 sessions. Of these, 55 pairs of simultaneously recorded neurons were composed of CH and PR in terms of the response type (39 and 35 units for CH and PR, respectively) and were further analyzed. We calculated the population dynamics of the Granger causality for all the recorded CH-PR pairs using the optimal stimulus for each CH-PR pair (Figure 2A). During 194 Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc.

the delay period, a population of CH-PR pairs exhibited prominent causality in the gamma frequency range predominantly in the direction from CH to PR. Though the timing of the Granger signal during the delay period varied across cell pairs (Figures 1D and 1F), significant causal influence in the gamma range (30–120 Hz) was, as a population, observed during the delay period only in the direction from CH to PR (from 1.5 to 2.5 s following cue onset; p < 0.002, compared with trial-shifted control, three-way repeated-measures ANOVA followed by least significant difference (LSD) test; Figures 2A and 2B, left), but not in the opposite direction (i.e., from PR to CH). As a result, population causal influence in the gamma range revealed a significant directional bias from CH to PR (p < 0.001; Figures 2A and 2B, left). This directional bias of causal influence was also observed when the time window was extended to the whole delay period (Figure S3A) or when limited to the period during which the firing rates of both CH and PR neurons were significantly above baseline (Figure S3B) but was not observed when calculated using trialshifted spike trains (p > 0.3 for either time window), as expected. In the alpha/beta frequency range (10–25 Hz), a similar trend was observed during the delay period, but the causality from CH to PR was not significant (p > 0.05; Figure 2B, right). We also examined the Granger causality in the theta (4–12 Hz) frequency range during the delay period and found no significant causal influence in either direction (Figure S3C). Population coherence was also

Neuron Microcircuit Dynamics in IT Cortex during Recall

Figure 3. Spike-Jittered Control for Granger Causality Analysis (A) Directionality of gamma Granger causality between CHs and PRs following subtraction of spike-jittered control. ++: p < 0.001, comparison with zero, two-way repeated-measures ANOVA followed by LSD. *: p < 0.009. Error bars, SEM. (B) Effects of variable jitter ranges on the directionality of gamma causal influence between CH-PR pairs during the delay period. ++: p < 0.002, comparison with zero, paired t test. Error bars, SEM.

significant during the delay period in the gamma range (p < 0.02; Figure 2C, left; for further details about the coherence value, see Supplemental Text), but not in the alpha/beta range (p > 0.07; Figure 2C, right). Power spectra of spike trains were also examined for CH and PR neurons, and these spike trains exhibited significant gamma-band power (Figures S2A and S2B) in a stimulus-dependent manner (Figure S2C). Next, we examined individual cell pairs, each of which exhibited a significant causal influence at the single cell-pair level. During the delay period, 29 of 55 CH-PR pairs (53%) exhibited significantly directional causal influence in the gamma frequency range. Of these, a significant majority of CH-PR pairs exhibited significantly stronger causal influence in the direction from CH to PR (23 of 29 pairs, 79%; p < 0.002, c2 test; Figure 2D). For each of these significantly directional CH-PR pairs (n = 29), putative pre and postunits were defined on the basis of the causality value during the delay period. For each of these units, we calculated the pair-recall index (PRI) (Naya et al., 2003b), which evaluates the strength of recall activity on the basis of neuronal activities for all the learned pair associations, and compared the index values between the pre and postunits. In a significant majority of cell pairs, the PRI value of the postunit was larger than that of the corresponding preunit (21 of 29 pairs, 72%; p < 0.02, c2 test; Figure 2E). As a population, activities of postunits exhibited significantly higher PRI values compared to the preunits (p < 0.009, paired t test; Figure 2E), consistent with the finding that the population gamma causality was significantly biased in the direction from CH to PR during the delay period (Figure 2B, left). In the above analysis, trial-shifted control (Lee, 2003; Brovelli et al., 2004; Zhou et al., 2008; Gregoriou et al., 2009; Hirabayashi et al., 2010) was utilized to evaluate the statistical significance of directed interactions. However, within-trial cofluctuation of firing rates for a given cell pair might affect the calculation of neuronal interactions, and this effect might be underestimated in the trialshifted control (Brody, 1998; Lee, 2003). We therefore further calculated spike-jittered control (Fujisawa et al., 2008; Amarasingham et al., 2012), in which spike timings were randomly jittered within a range of ±10 ms before calculating the Granger causality. The resultant directionality was statistically comparable irrespective of whether trial-shifted control or spike-jittered control was utilized to evaluate the causal influence in the

gamma range (Figure 3A). To examine the precision of spike timings in the directed couplings from CH to PR, we further calculated the Granger causality with variable jitter parameters (2, 5, and 10 ms). We then evaluated the effects of the jittering by calculating the directionality of the Granger causality (Figure 3B). The directionality of the Granger causality gradually decayed as the jitter parameter became larger. With a jitter parameter of 5 ms, the directionality of the Granger causality declined to insignificant level (p > 0.09, paired t test), and the directionality was almost completely abolished when using a jitter parameter of 10 ms. These results indicate that the observed Granger causality depended on at least a 5 ms precision of spike timings on average, consistent with the observed interactions in the gamma frequency range: for a given unit, ±5 ms jittering would randomize the spike timings in the range of 10 ms, which corresponds to half the cycle of 50 Hz oscillation. In the population dynamics of Granger causality from CH to PR (Figure 2A, left), gamma-frequency components were predominantly observed in two distinct bands. To examine whether the individual pairs indeed tended to show the causal influence either in one or the other of these bands, we calculated the causality in two distinct frequency bands separately: 30–70 Hz (low-gamma) (Vianney-Rodrigues et al., 2011) and 80–120 Hz (high-gamma) (Griffiths et al., 2010). The distribution of relative amplitudes of causal influence in the high- and low-gamma ranges showed that causal influence of individual pairs were significantly biased to either of the low- or high-gamma component (Figure S4A; see Supplemental Text for details). The highand low-gamma components of the directed interactions tended to show different dynamics during the delay period, but the difference did not reach significance (Figure S4B, see Supplemental Text for details). We also examined whether the directional bias of interactions from CH to PR could be observed in the cross-correlograms between spike trains during the whole delay period (see Figure S4C for an example data) In total, 17 of 55 CH-PR pairs (31%) showed a significant peak on the shift-predictor-subtracted cross-correlogram (SSCC) (Perkel et al., 1967; Steinmetz et al., 2000; Usrey et al., 2000; Kohn and Smith, 2005; Hirabayashi and Miyashita, 2005; Hirabayashi et al., 2010; see Supplemental Experimental Procedures). The z value of the SSCC peak was on average 4.3 ± 0.7 (mean ± SEM). For these SSCCs, Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc. 195

Neuron Microcircuit Dynamics in IT Cortex during Recall

Figure 4. Temporal Relationships between the Causal Influence and Recall Activity (A) Dynamics of recall activity around the time of the causality increase. Red and black traces, population data of the normalized dynamics of gamma causality from CH to PR (red) and recall activity (black) around the time of halfmaximum causality. Maximum values of the ordinate, 0.68 and 0.36 for causality and recall activity, respectively. Left traces, corresponding dynamics during the baseline period. Red and black arrows indicate the time when the causality and recall activity significantly exceeded 0. Thin traces, mean ± SEM. (B) Scatter plot of the latencies of recall activity (ordinate) and causality from CH to PR (abscissa) for each cell pair. Cyan and magenta dots, cell pairs shown in Figure 1 (cyan, C and D; magenta, E and F). (C) Population average latencies of Granger causality and recall activity. **: p < 0.002, paired t test. Error bars, SEM.

peak positions were significantly biased to the direction that inferred the signal flow from CH to PR (14 of 17 pairs, p < 0.01, c2 test), consistent with the results obtained by the Granger causality analysis. Moreover, the distribution of asymmetry index (Alonso and Martinez, 1998; Menz and Freeman, 2003; Hirabayashi et al., 2010; Takeuchi et al., 2011) of SSCCs, which evaluates the bias of the directionality, was significantly shifted to the direction from CH to PR (0.52 ± 0.19, p < 0.02, paired t test). These results further support the observed directional bias of causal influence for CH-PR pairs. To assess the temporal relationships between the causal influence and the activities of recipient neurons, we next calculated the latencies of the causality from CHs to PRs and the recall activity of PRs. In total, 30 of 55 CH-PR pairs (55%) showed both significant Granger causality from CH to PR and significant recall activity during the whole delay period, and were included in this analysis. Because both the latencies of the Granger causality and recall activity varied across cell pairs, we first examined the dynamics of the recall activity around the time of the causality increase for each cell pair. Figure 4A shows the population dynamics of causal influence from CHs to PRs (red) and the recall activity of the target PRs (black) that were sorted according to the time when the Granger signal started to increase. Significant increase in the recall activity was observed following elevation of the causal influence (p < 0.001, 0 s versus 1.0 s from the half-maximum causality; Figure 4A). Population recall activity significantly exceeded the baseline level (activity during 800 ms before cue onset) at 260 ms after the time of half-maximum causality (Figure 4A, black arrow), while the causality itself became significant just at the time of half maximum (red arrow). As a population, the half-maximum latencies of causal influence from CHs to PRs were significantly shorter than those of the recall activities (p < 0.002, paired t test; Figures 4B and 4C). Indeed, a significant majority of CH-PR pairs exhibited the directed influence from CH to PR more rapidly than the recall activity of the recipient PR (22 of 30 pairs, p < 0.02, c2 test; Figure 4B). To address whether the signal flow from CH to PR has an impact on subsequent interactions between PR cells, we simultaneously recorded from triplets of delay-selective cells, i.e., a pair of PR cells and a CH cell. With these CH-PR-PR triplets, we examined the interactions between PR cells that develop in concert with the directed influence from CH to PR (Figures 196 Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc.

S5A–5C). Figure 5A shows the population data for the Granger causality between PR cells (PR1 and PR2) sorted based on the time at which the directed influence from CH to PR1 reached half maximum. When the causal influence from CH to PR1 increased, causal influence between PR cells also exhibited a transient increase in the gamma frequency range, predominantly in the direction from PR1 to PR2 (two-way repeatedmeasures ANOVA followed by LSD, p < 0.001; n = 19; Figures 5A, left, and 5B, left). This directionally-biased interaction from PR1 to PR2 was not observed when trials were shifted between spike trains (p > 0.3; Figures 5A, middle, and 5B, left), or when the same calculation was conducted following triplet shuffling (p > 0.1; Figures 5A, right, and 5B, left; see Experimental Procedures). Triplet shuffling demonstrated that the calculated ‘‘GCCH/PR1-triggered GCPR1/PR2’’ for the population did not simply reflect the temporal proximity of GCCH/PR1 and GCPR1/PR2 among the population of triplets. In contrast to the gamma range, directional bias of the interaction between PRs was not observed in the alpha/beta frequency range (p > 0.07; Figure 5B, right). We also investigated whether directionally biased signal flow between PR1 and PR2 occurred when the causal influence between CH and PR1 increased in the opposite direction (i.e., from PR1 to CH), but the results showed that the signal flow between PR1 and PR2 was not significant in either direction (p > 0.4). We next tested whether the observed directional bias in the causal influence between PRs was affected by the signal mediated by CHs (i.e., PR1 to CH to PR2). To test this possibility, we recalculated the Granger causality between PRs after canceling out the effects of CH cells (Geweke, 1984; Cadotte et al., 2008; see Experimental Procedures). The results showed that the directional bias in the causal influence between PRs (i.e., PR1 to PR2) in the gamma frequency range remained significant even after canceling out the effects of CH cells (Figure S5D). Increased coherence between PR cells was also observed only in the gamma range (p < 0.01) in concert with the directed influence from CH to PR. Together, the above results suggest that the neuronal signal from CH to PR was accompanied by further interactions between PR cells predominantly in the direction that causes a cascade-like signal flow from CH to PR to another PR. Finally, we recorded from CH-CH-PR triplets, and examined whether CH-CH interactions occurred around the time at which

Neuron Microcircuit Dynamics in IT Cortex during Recall

Figure 5. CH-PR-PR Triplet Interactions To see whether the signal flow from CH to PR indeed has an impact on the next interaction between PR cells, we examined, for CH-PR-PR triplet data, GCCH/PR -triggered interaction between PR cells. (A) Population causality between PR1 and PR2, triggered by GCCH/PR1. Raw (left), trial-shifted (middle), and triplet-shuffled (right) spike trains were used to calculate the causal influence between PR1 and PR2. Time 0 represents the time at which the GCCH/PR1 value exceeded halfmaximum (see Figures S5A–S5C). (B) Directionality of the triggered causality between PR1 and PR2 in the gamma (left) and alpha/beta (right) ranges. ++: p < 0.001, comparison with zero, two-way repeated-measures ANOVA followed by LSD. **: p < 0.003. *: p < 0.02. Error bars, SEM. See also Figure S5D.

the directed influence from CH to PR elevated. For CH-CH-PR triplets in which the causal influence from CH to PR was larger than that in the opposite direction, the gamma coherence between CHs elevated prior to the time at which the causal influence from CH to PR reached half maximum (raw versus trial-shifted spike trains; p < 0.002 for CH-CH; p > 0.1 for CHPR, two-way repeated-measures ANOVA followed by LSD; n = 39 triplets; Figure 6; for further details about the coherence value, see Supplemental Text). These results indicate that interactions between CHs elevated prior to the increase in the causal influence from CH to PR, implying that these CH-CH couplings might facilitate subsequent directed interactions from CH to PR. We also examined the Granger causality between CHs in the CH-CH-PR triplets (Figure S6). The results showed that when the causality from CH1 to PR was elevated, the causal influence between CH1 and another CH (CH2) in the gamma range was significantly biased in the direction from CH1 to CH2, suggesting a divergent activity propagation from a given CH (CH1) to a PR, and from the CH1 to another CH (CH2). DISCUSSION In the present study, we simultaneously recorded from pairs or triplets of perirhinal neurons in monkeys while they performed

a pair-association memory task. Granger causality analysis between memory neurons revealed that directed interactions occurred predominantly from cueholding neurons to pair-recall neurons during retrieval of visual associative memory. Moreover, triplet recordings showed that these interactions coincided with further directed influence predominantly from the recipient recall neuron to another recall neuron, implying cascade-like signal propagation among the memory neurons. These results suggest that directed interactions among perirhinal memory cell assembly are dynamically modulated to implement functional microcircuitry for retrieval of visual associative long-term memory (Figure 7). Granger Causality as a Measure of Directed Interactions between Single Neurons Granger causality estimates whether the present value of a given time series X is better predicted by incorporating the past knowledge of another time series Y than by using only the past knowledge of X (Dhamala et al., 2008; Nedungadi et al., 2009; Hirabayashi et al., 2010; Kim et al., 2011; Verhoef et al., 2011; Quinn et al., 2011). Granger causality, as well as directed coherence or directed transfer function, has been widely utilized to reveal directional interactions between distant brain areas via continski uous signals, such as LFP, EEG, and BOLD signal (Kamin and Blinowska, 1991; Roebroeck et al., 2005; Gregoriou et al., 2009; Verhoef et al., 2011). Recently, Granger causality analysis has also been applied for spike trains (Nedungadi et al., 2009; Hirabayashi et al., 2010; Kim et al., 2011; Quinn et al., 2011). Previous studies have proven that the application of Granger causality to spike trains provided directional information that was consistent with anatomical data (Nedungadi et al., 2009) or with the results of a conventional cross-correlation analysis (Nedungadi et al., 2009; Hirabayashi et al., 2010). In the present Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc. 197

Neuron Microcircuit Dynamics in IT Cortex during Recall

Figure 6. Triplet Interactions for CH-CH-PR (A) Spectral dynamics of coherence for CH-CH (top) and CH-PR (bottom) pairs around the time of directed influence from CH to PR. (B) Gamma coherence for CH-PR (white) and CHCH (black) pairs before (left) and after (right) half maximum of the directed influence from CH to PR. ++: p < 0.002, comparison with zero, two-way repeated-measures ANOVA followed by LSD. **: p < 0.001. Error bars, SEM. See also Figure S6.

comparable results on the directionality of causal influence during the delay period, supporting the present results of the Granger causality analysis.

study, Granger causality was observed predominantly in the direction from cue-holding neurons to pair-recall neurons, and these results were also supported by cross-correlograms, a standard measure for directed interactions between neurons (Alonso and Martinez, 1998; Menz and Freeman, 2003; Atencio and Schreiner, 2010; Hirabayashi et al., 2010; Takeuchi et al., 2011). It has been shown that spectral estimates of spike trains provide an advantageous approach for the detection of neuronal interactions because they reflect not only the primary peak of the cross-correlations, but also their flank structures (Jarvis and Mitra, 2001; Fries et al., 2008). In addition, we used a multitaper method to determine spectral estimates of spike trains, which have also been shown to be effective for the detection of correlation structures by reducing the noises and biases (Jarvis and Mitra, 2001; Fries et al., 2008). These factors might explain why the Granger causality and cross-correlation analyses in the present study revealed a different sensitivity for the detection of neuronal interactions. In calculating the Granger causality, stationarity of the analyzed signal is an important prerequisite (Ding et al., 2000; Young and Eggermont, 2009). In the present study, we utilized a sliding window technique to reduce the effects of data nonstationarity as in previous studies (DeCoteau et al., 2007; Pesaran et al., 2008; Zhou et al., 2008; Gregoriou et al., 2009; Hirabayashi et al., 2010; Verhoef et al., 2011). In addition, we focused on the causal influence during the delay period, for which non-stationarity of the spiking activity is expected to be lower than that for the visual stimulation period. Furthermore, to assess the effects of firing rate dynamics, we calculated two different controls of the Granger causality: a trial-shifted control (Lee, 2003; Brovelli et al., 2004; Zhou et al., 2008; Gregoriou et al., 2009; Hirabayashi et al., 2010), in which the spike trains of a given cell pair were obtained in different trials, and a spike-jittered control (Fujisawa et al., 2008; Amarasingham et al., 2012), in which exact spike timings in each train were randomly jittered (< ±10 ms) within a trial. Both controls provided statistically 198 Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc.

Neuronal Interactions in the Gamma Frequency Range In the present study, directed interactions were predominantly observed in the gamma frequency range. In contrast, none of the results regarding alpha/beta range reached the statistical significance. Therefore, the results in the alpha/beta range will emphasize that the neuronal interactions found in this study were specific to the gamma frequency range, consistent with previous studies describing local neuronal interactions (Lee, 2003; Fries et al., 2008; Sirota et al., 2008; Zhou et al., 2008; Hirabayashi et al., 2010; Lima et al., 2010). However, our results do not exclude the possibility that interactions with different brain areas in lower frequency ranges might be multiplicatively involved (Colgin et al., 2009) in the microcircuit operations in the perirhinal cortex. Temporal Relationships between the Granger Signal and Recall Activity In the present study, we examined the temporal relationships between the Granger signal from CHs to PRs and the recall activity of the recipient PRs, and found that the latency of the causal influence after cue onset was significantly shorter than that of the recall signal (Figure 4). In interpreting this temporal relationship, however, some cautions would be needed as follows. First, considering that a given cortical neuron receives inputs from many other neurons, Granger signal from just one neuron would only partly explain the generation of spikes in the target neuron. As a result, the latencies of the causal influence from different source neurons would distribute widely over the response latency of the target neuron. Second, spike-spike interactions in the gamma frequency range have been previously shown to emerge several hundred milliseconds after stimulus onset in the primary visual cortex, probably due to underlying network dynamics (Zhou et al., 2008). However, it is not known whether similar network dynamics also underlies the memory retrieval in the perirhinal cortex. Furthermore, top-down signals from other cortical areas might also be involved in the recall activity (Tomita et al., 1999; Miyashita, 2004), in addition to the local interactions described above. Given these considerations, it would be an important future challenge to examine

Neuron Microcircuit Dynamics in IT Cortex during Recall

Figure 7. Schematic Diagram of the Inferred Functional Microcircuitry in the Perirhinal Cortex in Retrieval of Object Association Memory Cyan and magenta neurons, CH and PR neurons. Blue and red arrows between neurons depict directed interactions identified in the present study. Lines between neurons represent functional couplings (see also Figure S7). The present results suggest that during memory retrieval, cue information is transmitted from CH cell assembly to PR cell assembly to convert the representation in the microcircuit from the cue to the sought target.

the cellular/network machineries that link the observed Granger causality with the spiking activity of the recipient neurons. Triplet Interactions among Memory Cell Assembly Higher-order neuronal interactions among multiple (more than three) neurons have been reported previously (Paz et al., 2006; Luczak et al., 2007; Ohiorhenuan et al., 2010). In particular, higher-order interactions in the primary visual cortex have been predominantly observed for local processing compared to more distant neuronal couplings (Ohiorhenuan et al., 2010). These previous studies imply the existence of triplet interactions in perirhinal microcircuits for memory recall (see Supplemental Text for details about the network sizes in the present study). In the present study, we recorded from triplets of memory neurons, CH-PR-PR, and demonstrated that causal influence between PRs increased in concert with the directed interactions from CH to PR (Figure 5). Moreover, this causal influence between PRs was directionally biased from PR1 to PR2, where PR1 was the target of the directed interactions from the CH. In calculating triplet interactions, effects of the third neuron on interactions between the other two neurons should be considered. We confirmed that the signal flow from PR1 to PR2 was not attenuated when the influence of the CH on PR2 was taken into account (Figure S5D), indicating that the observed causal influence from PR1 to PR2 was not due to an indirect interaction mediated by the CH. These results suggest that directed interactions from CH to PR are accompanied with further signal flow from the recipient PR to another PR, i.e., cascade-like signal propagation to the cells representing the sought target (Figure 7). ‘‘Phase Sequence’’ within a Perirhinal Microcircuit during Memory Recall According to a classical theory raised by D.O. Hebb, cognitive processes might be reflected in sequential interactions among multiple populations of neurons, each encoding relevant information as a cell assembly (Hebb, 1949; Harris, 2005; Buzsa´ki, 2010). This neuronal process has been termed ‘‘phase sequence,’’ and some previous studies in rodents have revealed neural correlates of the phase sequence (Harvey et al., 2012; Xu et al., 2012). In particular, a previous study showed that, in the context of spatial alternation memory task, the recorded ensemble of hippocampal neurons exhibited, during the delay period without visual inputs, a defined pattern of sequential firing

on demand of the retrieval of a specific spatial memory (Pastalkova et al., 2008). In the present study, we demonstrated directional signal flow between two different well-characterized neuron types, i.e., directed influence from CH to PR during the delay period, which was accompanied by further influence on another PR (see Supplemental Text for details about the relationships between neuronal interactions and firing patterns). This PR to PR causality might be important for propagation of the activity in the cell assembly of PRs that represent the sought target. Activity and its functional roles of CHs have been described in many previous literatures (see Naya et al., 2003b, and references therein). Based on these findings, we propose a model of circuit dynamics during associative memory retrieval that incorporates these cellular interactions (Figure 7). Together with earlier results, therefore, ‘‘phase sequence’’ might be a fundamental neuronal process for cognitive demands, including retrieval of visual associative long-term memory. Previous studies have raised possible mechanisms underlying the development of memory cells via neuronal interactions (Durstewitz et al., 2000; Brunel, 2003; Deco et al., 2010; Woloszyn and Sheinberg, 2012). Regarding the pair-association memory, one possible mechanism for associative coding lies in the interaction between neurons through learning (Messinger et al., 2001), each coding for the presented cue stimulus and its paired associate, respectively (Erickson and Desimone, 1999; Brunel, 2003; Albright, 2012). Cue-holding activity continuing to the presentation of the paired-associate would provide a time window for near-synchronous firing of these two types of neurons, which might lead to synaptic plasticity strengthening the connectivity between them (Brunel, 2003; Li and DiCarlo, 2010). Spike-timing-dependent plasticity would further provide a directional bias of these couplings on the basis of the firing order of these neurons (Dan and Poo, 2004), i.e., from the cueholding neuron to the other neuron that encodes the paired associate. Understanding the contribution of such synaptic plasticity to the dynamic operations of microcircuits would be an important issue for future study.

EXPERIMENTAL PROCEDURES Subjects and Behavioral Task All animal procedures complied with the U.S. National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by

Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc. 199

Neuron Microcircuit Dynamics in IT Cortex during Recall

the Institutional Review Committee of the University of Tokyo School of Medicine. The subjects were two adult macaque monkeys (see Supplemental Experimental Procedures for details). The procedure for a pair-association memory task (Figure 1A) was previously described in detail (Naya et al., 2003a, 2003b; Yoshida et al., 2003; Takeuchi et al., 2011; see Supplemental Experimental Procedures for details). In brief, monkeys were trained to perform a pair-association memory task using a set of six pairs of visual stimuli (12 monochrome Fourier descriptors, different sets were used for two monkeys). In each trial, a fixation point was presented for 0.8 s, after which a cue stimulus (one of the 12 visual stimuli) was presented for 1 s. Following a 2.2 s delay period, two stimuli were presented: the paired associate of the cue stimulus and a distractor. The monkey obtained a reward for correctly choosing the paired associate within 1.5 s. In this task paradigm, monkeys were required to choose not the presented cue itself, but instead the memorized paired associate, and thus, the task required cued recall, not recognition (Basile and Hampton, 2011). Throughout the recording sessions after training, monkeys’ performance was >90% correct. If the eye position deviated more than 1.5
–2.0
from the fixation point before the end of the delay period, the trial was automatically terminated.

matrix S (Dhamala et al., 2008; Nedungadi et al., 2009). According to Geweke’s formulation (Geweke, 1982), the autospectrum of a given signal can be decomposed into an intrinsic component and a component that is predicted from another signal. The Granger causality can thus be described as the ratio of the predicted component to the total of the autospectrum. Here, the Granger causality from spike train Nj(t) to Ni(t) at frequency f can be represented as INj/Ni ðfÞ = ln

Sii ðfÞ 


P jj

Sii ðfÞ 2 P 2 P  Hij ðfÞ ij = ii



(Equation 1)

The Granger causality in the opposite direction can be estimated similarly. Note that ‘‘causality’’ as estimated by Granger’s measure does not exclude the possibility that a third ‘‘hidden’’ variable could be the prior ‘‘cause’’ that influences both neurons in a given pair.

Spectral Analysis of Spike Trains Spectral estimates of spike trains (auto- or cross-spectrum and coherence) were calculated using a multitaper method with Chronux, an open-source MATLAB software package (DeCoteau et al., 2007; Pesaran et al., 2008; Zhou et al., 2008; Hirabayashi et al., 2010). With this method, spectral estimates of spike trains were determined as the average of the results computed after multiplying several different taper functions by the original signal. These processes yield less biased and less noisy spectral estimates than conventional methods (Jarvis and Mitra, 2001; Fries et al., 2008). Seven or nine orthogonal Slepian tapers (three tapers only for the analysis in the theta frequency range) were applied to calculate the spectral estimates of spike trains that were downsampled at a rate of 1 kHz.

Coherence and Granger Causality Analyses for Stimulus-Selective Neurons To calculate the spectral estimates between recorded neurons, the spiking response of each single-unit was examined as follows. A given unit was defined as cue-selective if the cue activity (mean firing rate during a 900 ms period beginning 100 ms after cue onset) was significantly stimulus-selective (one-way ANOVA, p < 0.05). Similarly, a given unit was defined as delayselective if the delay activity (mean firing rate during a 1,700 ms period beginning 500 ms after cue offset) showed significant stimulus selectivity (one-way ANOVA, p < 0.05). In total, 107 neurons (65 and 42 for monkeys 1 and 2, respectively) exhibited both cue- and delay-selective responses (CD-selective cells) and were further analyzed. For a given CD-selective cell, neuronal activity elicited by a given stimulus was classified as cue holding (CH) when (1) the neuron showed significant cue activity in response to a given stimulus, detected as a significant increase in firing rate compared to that during the preceding baseline period (800 ms prior to cue onset; paired t test, p < 0.05), and (2) the delay-activity elicited by the stimulus was less than the preceding cue activity (Figure 1B, right, blue). Likewise, for a given CD-selective cell, neuronal activity elicited by a given stimulus (c-stimulus) was classified as pair-recall (PR) when (1) the response to the paired-associate of the c-stimulus was significantly larger than the baseline activity and significantly larger than the response to the c-stimulus (paired t test, p < 0.05) and (2) the delay activity elicited by the c-stimulus was greater than the preceding cue activity (Figure 1B, right, red). For a given cell pair, coherence and/or Granger causality were calculated for a given stimulus if the response of each constituent cell for that stimulus was classified as CH and PR, respectively: CH and PR activities were required for the same stimulus. If more than two stimuli were classified as candidates for calculating spectral estimates of coupling as a CH-PR pair, the product of the delay responses of both constituent cells was calculated for these stimuli to determine the optimal stimulus that elicited the largest responses from both cells. The optimal stimulus was then used to conduct the coherence and/or Granger causality analysis. To correctly estimate the connectivity utilizing the current method, it would be better to use as large a number of spikes as possible. From a practical perspective, we used the optimal stimulus for each cell pair, because it was an effective way to satisfy this requirement with a reasonable number of trial repetitions in a single recording session. Spike trains were obtained for 85 ± 30 (mean ± SD) trials for each optimal stimulus in these analyses. Worst stimuli were also determined for each cell as worst three stimuli for eliciting responses. The strength of coherence or Granger causality in each frequency range (gamma or alpha/beta) was defined as the average value in the corresponding frequency range (gamma, 30–120 Hz; alpha/beta, 10–25 Hz).

Granger Causality Analysis We conducted a nonparametric Granger causality analysis using MATLAB as described in previous studies (Dhamala et al., 2008; Nedungadi et al., 2009; Hirabayashi et al., 2010). In brief, spectral estimates S(f) of spike trains were first calculated using the multitaper method. To calculate the Granger causality, a transfer function and error covariance matrix were estimated from S(f) using spectral factorization (Wilson, 1972), which decomposes S(f) into a unique corresponding transfer function H(f) and a noise covariance

Dynamics of Causality and Recall Activity Dynamics of coherence and Granger causality (Figures 1D, 1F, and 2A) were examined by calculating spectral dynamics for these measures using a 500 ms sliding window that was slid in 50 ms steps (F-T plot; frequency resolution: ±10 Hz; the start time of a given window was assigned as the time point of the window). For calculating the latency of causal influence from CH to PR (Figure 4), the F-T plot of the causality (trial-shifted control was presubtracted) was calculated, and the maximum value during the delay period was detected

Recording Procedures The activities of multiple single-units were recorded from area 36 of the perirhinal cortex (Figures 1B, left, and S1; see Supplemental Experimental Procedures for details) of the two monkeys using a multicontact electrode (Tetrode, Thomas Recording, Giessen, Germany) (Hirabayashi and Miyashita, 2005; Hirabayashi et al., 2010; Ohiorhenuan et al., 2010). The spacing of the contacts was in the range of several tens of micrometers (see Supplemental Text about the relationships between the channels where CH-PR pairs were recorded and their coherence values). In the multiple single-unit recordings, neuronal signals were amplified, band-pass filtered (250 Hz to 5 kHz), and sorted online into pairs of single units using the standard window discrimination technique to monitor the responses of recorded neurons. Neuronal signals were also stored and then digitized offline at 25 kHz, after which waveform analysis was used to more precisely sort them into multiple single-units (Datawave Technologies, Longmont, CO) (Usrey et al., 2000; Roy and Alloway, 2001; Lee et al., 2005; Hirabayashi and Miyashita, 2005). The presence of a refractory period was confirmed in the autocorrelogram (Usrey et al., 2000; Hirabayashi and Miyashita, 2005; Takeuchi et al., 2011). If the number of spikes with interspike intervals < 2 ms exceeded 1% of the total for a given unit, that unit was discarded or reisolated (Hirabayashi and Miyashita, 2005; Takeuchi et al., 2011). Cross-cluster interspike interval histograms for each pair of units were also confirmed not to show any artificial peak which would imply the false sorting of spikes from one cell into two different clusters. These offline-sorted spike data obtained for correct trials were further analyzed for responses and functional connectivity (for the analyses of error trials, see Supplemental Text).

200 Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc.

Neuron Microcircuit Dynamics in IT Cortex during Recall

in the gamma frequency range to determine the time point of the corresponding half-maximum value. For calculating the normalized dynamics of gamma causality for each cell pair (Figure 4A), mean value of the causality in the gamma range for each time point (trial-shifted control was presubtracted) was normalized using its maximum value during the delay period. Normalized dynamics of recall activity (Figure 4A) was similarly calculated for the same time windows as those for the causal influence, using the maximum value during the delay period and the baseline value during 800 ms period before cue onset. Half-maximum latency of the recall activity was then determined using the normalized dynamics. Note that for both causal influence and recall activity, the time of half maximum value was determined by continuously falling down from the maximum value of each plot.

the coherence of CH-CH and CH-PR (Figure 6B), these coherence values were calculated in 250-ms windows before and after TCH/PR1. Some neurons were used in more than one CH-CH-PR triplet analysis. All statistical tests in the present study were two-sided unless otherwise stated. SUPPLEMENTAL INFORMATION Supplemental Information includes seven figures and Supplemental Experimental Procedures and can be found with this article online at http://dx.doi. org/10.1016/j.neuron.2012.10.031. ACKNOWLEDGMENTS

Triplet Granger Causality Analysis When we could simultaneously record from a triplet of cells that included one CH cell and two PR cells (CH and PR activities were required for the same stimulus), we conducted a triplet Granger causality analysis for the CH-PR-PR as follows (Figures S5A–S5C). First, the F-T plot of the Granger causality (trialshifted control was presubtracted) between a CH cell and a PR cell (PR1) was calculated as described above (with 25 ms steps), and the half maximum value in the gamma frequency range was detected. If the peak value of causal influence was larger in the direction from CH to PR1 than in the opposite direction, the F-T plot for the Granger causality between PR1 and another PR (PR2) was also calculated. Some neurons were used in more than one CH-PR-PR triplet analysis. We defined GCX/Y as INj/Ni in Equation 1, where Nj is the spike train of cell X and Ni is the spike train of cell Y (e.g., CH and PR). For each of above triplets, the ‘‘GCCH/PR1-triggered GCPR1/PR2’’ was defined as GCCH/PR1 -triggered GCPR1/PR2 ðtÞ = GCPR1/PR2 ðTCH/PR1 + tÞ

(Equation 2)

where TCH/PR1 represents the time point when the GCCH/PR1 value reached the half maximum, and t represents the time lag from TCH/PR1. The GCCH/PR1-triggered GCPR2/PR1 was also similarly defined as GCCH/PR1 -triggered GCPR2/PR1 ðtÞ = GCPR2/PR1 ðTCH/PR1 + tÞ

(Equation 3)

For all CH-PR1-PR2 triplets, F-T plots of the Granger causality between PR1 and PR2 were calculated as a function of time t. These F-T plots for all the triplets were then compiled separately for the two directions (i.e., PR1 to PR2, and PR2 to PR1) to determine the predominant direction of triplet interaction among the population (i.e., CH/PR1/PR2 or CH/PR1)PR2). Note that any of the PRs could be the PR1 in a given triplet if the maximum GCCH/PR was larger than that in the opposite direction: if both PRs met this criterion, each PR was used as the PR1, but in a different triplet. Trial-shifted data were also calculated, in which TCH/PR1 was conserved from the original data, but the trials were shifted in computing the interactions between PRs. To quantify the triggered Granger causality/coherence in the gamma and alpha/beta ranges (Figures 5B and S5D), mean amplitudes of these measures within each frequency range were calculated in the epoch of 100 to 400 ms from TCH/PR1. Because the timing of causal influence from CH to PR was relatively clustered in the delay period (Figure 2A), the calculated ‘‘GCCH/PR1-triggered GCPR1/PR2’’ for the population might simply reflect the temporal proximity of GCCH/PR1 and GCPR1/PR2 among the population, instead of individual relationships between GCCH/PR1 and GCPR1/PR2 for each triplet. To test whether the timing of GCCH/PR1 was indeed critical for triggering the corresponding GCPR1/PR2 in each triplet, the same measure was calculated after randomizing TCH/PR1 among all triplets (‘‘triplet-shuffled’’ data). To control for the effect of a CH cell on the interaction between PRs in a given triplet, a conditional Granger causality was calculated for each triplet, where GCCH/PR2 was subtracted from GCPR1/PR2, and GCCH/PR1 was subtracted from GCPR2/PR1 (Geweke, 1984; Cadotte et al., 2008) for each time window (Figure S5D). Causality values between PRs after these subtractions represent those, from which the components of indirect interactions mediating the CH cells were precluded. Population directionality of the resultant conditional Granger causality between PR1 and PR2 was then calculated and statistically evaluated (Figure S5D). Triplet analysis for CH-CH-PR (Figure 6) was similarly conducted as in the above analysis for CH-PR-PR triplets. To compare the dynamics between

We thank G. Rangarajan for kindly providing scripts to conduct the nonparametric Granger causality analysis, H. Kasahara and T. Watanabe for experimental assistance, and M. Takeda and K.W. Koyano for helpful discussions. This work was supported by MEXT/JSPS KAKENHI Grant Numbers 19002010 and 24220008 to Y.M. and 18700378 to T.H. Accepted: October 26, 2012 Published: January 9, 2013 REFERENCES Albright, T.D. (2012). On the perception of probable things: neural substrates of associative memory, imagery, and perception. Neuron 74, 227–245. Alonso, J.-M., and Martinez, L.M. (1998). Functional connectivity between simple cells and complex cells in cat striate cortex. Nat. Neurosci. 1, 395–403. Amarasingham, A., Harrison, M.T., Hatsopoulos, N.G., and Geman, S. (2012). Conditional modeling and the jitter method of spike resampling. J. Neurophysiol. 107, 517–531. Atencio, C.A., and Schreiner, C.E. (2010). Columnar connectivity and laminar processing in cat primary auditory cortex. PLoS ONE 5, e9521. Basile, B.M., and Hampton, R.R. (2011). Monkeys recall and reproduce simple shapes from memory. Curr. Biol. 21, 774–778. Brody, C.D. (1998). Slow covariations in neuronal resting potentials can lead to artefactually fast cross-correlations in their spike trains. J. Neurophysiol. 80, 3345–3351. Brovelli, A., Ding, M., Ledberg, A., Chen, Y., Nakamura, R., and Bressler, S.L. (2004). Beta oscillations in a large-scale sensorimotor cortical network: directional influences revealed by Granger causality. Proc. Natl. Acad. Sci. USA 101, 9849–9854. Brunel, N. (2003). Dynamics and plasticity of stimulus-selective persistent activity in cortical network models. Cereb. Cortex 13, 1151–1161. Buckley, M.J., and Gaffan, D. (1998). Perirhinal cortex ablation impairs configural learning and paired-associate learning equally. Neuropsychologia 36, 535–546. Buzsa´ki, G. (2010). Neural syntax: cell assemblies, synapsembles, and readers. Neuron 68, 362–385. Cadotte, A.J., DeMarse, T.B., He, P., and Ding, M. (2008). Causal measures of structure and plasticity in simulated and living neural networks. PLoS ONE 3, e3355. Colgin, L.L., Denninger, T., Fyhn, M., Hafting, T., Bonnevie, T., Jensen, O., Moser, M.-B., and Moser, E.I. (2009). Frequency of gamma oscillations routes flow of information in the hippocampus. Nature 462, 353–357. Dan, Y., and Poo, M.M. (2004). Spike timing-dependent plasticity of neural circuits. Neuron 44, 23–30. Deco, G., Rolls, E.T., and Romo, R. (2010). Synaptic dynamics and decision making. Proc. Natl. Acad. Sci. USA 107, 7545–7549. DeCoteau, W.E., Thorn, C., Gibson, D.J., Courtemanche, R., Mitra, P., Kubota, Y., and Graybiel, A.M. (2007). Learning-related coordination of striatal and hippocampal theta rhythms during acquisition of a procedural maze task. Proc. Natl. Acad. Sci. USA 104, 5644–5649.

Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc. 201

Neuron Microcircuit Dynamics in IT Cortex during Recall

Dhamala, M., Rangarajan, G., and Ding, M. (2008). Estimating Granger causality from fourier and wavelet transforms of time series data. Phys. Rev. Lett. 100, 018701.1–018701.4. Ding, M., Bressler, S.L., Yang, W., and Liang, H. (2000). Short-window spectral analysis of cortical event-related potentials by adaptive multivariate autoregressive modeling: data preprocessing, model validation, and variability assessment. Biol. Cybern. 83, 35–45. Durstewitz, D., Seamans, J.K., and Sejnowski, T.J. (2000). Neurocomputational models of working memory. Nat. Neurosci. Suppl. 3, 1184–1191. Erickson, C.A., and Desimone, R. (1999). Responses of macaque perirhinal neurons during and after visual stimulus association learning. J. Neurosci. 19, 10404–10416. Fries, P., Womelsdorf, T., Oostenveld, R., and Desimone, R. (2008). The effects of visual stimulation and selective visual attention on rhythmic neuronal synchronization in macaque area V4. J. Neurosci. 28, 4823–4835. Fujisawa, S., Amarasingham, A., Harrison, M.T., and Buzsa´ki, G. (2008). Behavior-dependent short-term assembly dynamics in the medial prefrontal cortex. Nat. Neurosci. 11, 823–833. Geweke, J. (1982). Measurement of linear dependence and feedback between multiple time series. J. Am. Stat. Assoc. 77, 304–313. Geweke, J. (1984). Measures of conditional linear dependence and feedback between time series. J. Am. Stat. Assoc. 79, 907–915. Gregoriou, G.G., Gotts, S.J., Zhou, H., and Desimone, R. (2009). Highfrequency, long-range coupling between prefrontal and visual cortex during attention. Science 324, 1207–1210. Griffiths, T.D., Kumar, S., Sedley, W., Nourski, K.V., Kawasaki, H., Oya, H., Patterson, R.D., Brugge, J.F., and Howard, M.A. (2010). Direct recordings of pitch responses from human auditory cortex. Curr. Biol. 20, 1128–1132. Harris, K.D. (2005). Neural signatures of cell assembly organization. Nat. Rev. Neurosci. 6, 399–407. Harvey, C.D., Coen, P., and Tank, D.W. (2012). Choice-specific sequences in parietal cortex during a virtual-navigation decision task. Nature 484, 62–68. Hebb, D.O. (1949). The Organization of Behavior (New York: Whiley). Hirabayashi, T., and Miyashita, Y. (2005). Dynamically modulated spike correlation in monkey inferior temporal cortex depending on the feature configuration within a whole object. J. Neurosci. 25, 10299–10307. Hirabayashi, T., Takeuchi, D., Tamura, K., and Miyashita, Y. (2010). Triphasic dynamics of stimulus-dependent information flow between single neurons in macaque inferior temporal cortex. J. Neurosci. 30, 10407–10421. Jarvis, M.R., and Mitra, P.P. (2001). Sampling properties of the spectrum and coherency of sequences of action potentials. Neural Comput. 13, 717–749. ski, M.J., and Blinowska, K.J. (1991). A new method of the description of Kamin the information flow in the brain structures. Biol. Cybern. 65, 203–210. Kim, S., Putrino, D., Ghosh, S., and Brown, E.N. (2011). A Granger causality measure for point process models of ensemble neural spiking activity. PLoS Comput. Biol. 7, e1001110. Kohn, A., and Smith, M.A. (2005). Stimulus dependence of neuronal correlation in primary visual cortex of the macaque. J. Neurosci. 25, 3661–3673. Lee, D. (2003). Coherent oscillations in neuronal activity of the supplementary motor area during a visuomotor task. J. Neurosci. 23, 6798–6809. Lee, H., Simpson, G.V., Logothetis, N.K., and Rainer, G. (2005). Phase locking of single neuron activity to theta oscillations during working memory in monkey extrastriate visual cortex. Neuron 45, 147–156. Li, N., and DiCarlo, J.J. (2010). Unsupervised natural visual experience rapidly reshapes size-invariant object representation in inferior temporal cortex. Neuron 67, 1062–1075. Lima, B., Singer, W., Chen, N.-H., and Neuenschwander, S. (2010). Synchronization dynamics in response to plaid stimuli in monkey V1. Cereb. Cortex 20, 1556–1573. Luczak, A., Bartho´, P., Marguet, S.L., Buzsa´ki, G., and Harris, K.D. (2007). Sequential structure of neocortical spontaneous activity in vivo. Proc. Natl. Acad. Sci. USA 104, 347–352.

202 Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc.

Menz, M.D., and Freeman, R.D. (2003). Stereoscopic depth processing in the visual cortex: a coarse-to-fine mechanism. Nat. Neurosci. 6, 59–65. Messinger, A., Squire, L.R., Zola, S.M., and Albright, T.D. (2001). Neuronal representations of stimulus associations develop in the temporal lobe during learning. Proc. Natl. Acad. Sci. USA 98, 12239–12244. Miyashita, Y. (2004). Cognitive memory: cellular and network machineries and their top-down control. Science 306, 435–440. Murray, E.A., Gaffan, D., and Mishkin, M. (1993). Neural substrates of visual stimulus-stimulus association in rhesus monkeys. J. Neurosci. 13, 4549–4561. Naya, Y., Sakai, K., and Miyashita, Y. (1996). Activity of primate inferotemporal neurons related to a sought target in pair-association task. Proc. Natl. Acad. Sci. USA 93, 2664–2669. Naya, Y., Yoshida, M., and Miyashita, Y. (2001). Backward spreading of memory-retrieval signal in the primate temporal cortex. Science 291, 661–664. Naya, Y., Yoshida, M., and Miyashita, Y. (2003a). Forward processing of longterm associative memory in monkey inferotemporal cortex. J. Neurosci. 23, 2861–2871. Naya, Y., Yoshida, M., Takeda, M., Fujimichi, R., and Miyashita, Y. (2003b). Delay-period activities in two subdivisions of monkey inferotemporal cortex during pair association memory task. Eur. J. Neurosci. 18, 2915–2918. Nedungadi, A.G., Rangarajan, G., Jain, N., and Ding, M. (2009). Analyzing multiple spike trains with nonparametric Granger causality. J. Comput. Neurosci. 27, 55–64. Ohiorhenuan, I.E., Mechler, F., Purpura, K.P., Schmid, A.M., Hu, Q., and Victor, J.D. (2010). Sparse coding and high-order correlations in fine-scale cortical networks. Nature 466, 617–621. Pastalkova, E., Itskov, V., Amarasingham, A., and Buzsa´ki, G. (2008). Internally generated cell assembly sequences in the rat hippocampus. Science 321, 1322–1327. Paz, R., Pelletier, J.G., Bauer, E.P., and Pare´, D. (2006). Emotional enhancement of memory via amygdala-driven facilitation of rhinal interactions. Nat. Neurosci. 9, 1321–1329. Perkel, D.H., Gerstein, G.L., and Moore, G.P. (1967). Neuronal spike trains and stochastic point processes. II. Simultaneous spike trains. Biophys. J. 7, 419–440. Pesaran, B., Nelson, M.J., and Andersen, R.A. (2008). Free choice activates a decision circuit between frontal and parietal cortex. Nature 453, 406–409. Quinn, C.J., Coleman, T.P., Kiyavash, N., and Hatsopoulos, N.G. (2011). Estimating the directed information to infer causal relationships in ensemble neural spike train recordings. J. Comput. Neurosci. 30, 17–44. Roebroeck, A., Formisano, E., and Goebel, R. (2005). Mapping directed influence over the brain using Granger causality and fMRI. Neuroimage 25, 230–242. Roy, S.A., and Alloway, K.D. (2001). Coincidence detection or temporal integration? What the neurons in somatosensory cortex are doing. J. Neurosci. 21, 2462–2473. Sakai, K., and Miyashita, Y. (1991). Neural organization for the long-term memory of paired associates. Nature 354, 152–155. Sirota, A., Montgomery, S., Fujisawa, S., Isomura, Y., Zugaro, M., and Buzsa´ki, G. (2008). Entrainment of neocortical neurons and gamma oscillations by the hippocampal theta rhythm. Neuron 60, 683–697. Squire, L.R., Wixted, J.T., and Clark, R.E. (2007). Recognition memory and the medial temporal lobe: a new perspective. Nat. Rev. Neurosci. 8, 872–883. Steinmetz, P.N., Roy, A., Fitzgerald, P.J., Hsiao, S.S., Johnson, K.O., and Niebur, E. (2000). Attention modulates synchronized neuronal firing in primate somatosensory cortex. Nature 404, 187–190. Suzuki, W.A. (2009). Perception and the medial temporal lobe: evaluating the current evidence. Neuron 61, 657–666. Takeuchi, D., Hirabayashi, T., Tamura, K., and Miyashita, Y. (2011). Reversal of interlaminar signal between sensory and memory processing in monkey temporal cortex. Science 331, 1443–1447.

Neuron Microcircuit Dynamics in IT Cortex during Recall

Tomita, H., Ohbayashi, M., Nakahara, K., Hasegawa, I., and Miyashita, Y. (1999). Top-down signal from prefrontal cortex in executive control of memory retrieval. Nature 401, 699–703.

Woloszyn, L., and Sheinberg, D.L. (2012). Effects of long-term visual experience on responses of distinct classes of single units in inferior temporal cortex. Neuron 74, 193–205.

Usrey, W.M., Alonso, J.M., and Reid, R.C. (2000). Synaptic interactions between thalamic inputs to simple cells in cat visual cortex. J. Neurosci. 20, 5461–5467.

Xu, S., Jiang, W., Poo, M., and Dan, Y. (2012). Activity recall in a visual cortical ensemble. Nat. Neurosci. 15, 449–455.

Verhoef, B.E., Vogels, R., and Janssen, P. (2011). Synchronization between the end stages of the dorsal and the ventral visual stream. J. Neurophysiol. 105, 2030–2042. Vianney-Rodrigues, P., Iancu, O.D., and Welsh, J.P. (2011). Gamma oscillations in the auditory cortex of awake rats. Eur. J. Neurosci. 33, 119–129. Wechsler, D. (1987). Wechsler Memory Scale-Revised (San Antonio, TX: The Psychological Corporation. Harcourt Brace Jovanovich). Wilson, G.T. (1972). The factorization of matricial spectral densities. SIAM J. Appl. Math. 23, 420–426.

Yoshida, M., Naya, Y., and Miyashita, Y. (2003). Anatomical organization of forward fiber projections from area TE to perirhinal neurons representing visual long-term memory in monkeys. Proc. Natl. Acad. Sci. USA 100, 4257–4262. Young, C.K., and Eggermont, J.J. (2009). Coupling of mesoscopic brain oscillations: recent advances in analytical and theoretical perspectives. Prog. Neurobiol. 89, 61–78. Zhou, Z., Bernard, M.R., and Bonds, A.B. (2008). Deconstruction of spatial integrity in visual stimulus detected by modulation of synchronized activity in cat visual cortex. J. Neurosci. 28, 3759–3768.

Neuron 77, 192–203, January 9, 2013 ª2013 Elsevier Inc. 203