Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2021 Apr 22.
Published in final edited form as: Neuron. 2020 Feb 20;106(2):256–264.e3. doi: 10.1016/j.neuron.2020.01.032

Combined phase-rate coding by persistently active neurons as a mechanism for maintaining multiple items in working memory in humans

Jan Kamiński 1,5, Aneta Brzezicka 1,2, Adam N Mamelak 1, Ueli Rutishauser 1,3,4,5
PMCID: PMC7217299  NIHMSID: NIHMS1569020  PMID: 32084331

Summary

Maintaining multiple items in working memory (WM) is central to human behavior. Persistently active neurons are thought to be a mechanism to maintain WMs, but it remains unclear how such activity is coordinated when multiple items are kept in memory. We show that memoranda-selective persistently active neurons in the human medial temporal lobe phase-lock to ongoing slow-frequency (1-7Hz) oscillations during WM maintenance. The properties of phase-locking are dependent on memory content and load. During high memory loads, the phase of the oscillatory activity to which neurons phase-lock provides information about memory content not available in the firing rate of the neurons. We provide a computational model that reveals that inhibitory-feedback mediated competition between multiple persistently active neurons reproduces this phenomenon. This work reveals a mechanism for the active maintenance of multiple items in WM that relies on persistently active neurons whose activation is orchestrated by oscillatory activity.

Graphical Abstract

graphic file with name nihms-1569020-f0005.jpg

Introduction

Working Memory (WM) is a fundamental human cognitive capacity that allows us to maintain and manipulate information stored for a short period of time in an active form. WM is essential for many high-level cognitive skills, including inference, decision making, mental calculations, and awareness (Baddeley, 2007). Despite its importance, the neuronal mechanisms of WM in humans are only beginning to be understood. Recent human single-neuron studies have revealed that in the human medial temporal lobe (MTL), stimulus-selective cells remain persistently active during the maintenance period of a WM task if the preferred stimulus of the cell is held in WM (Kamiński et al., 2017; Kornblith et al., 2017). The level of activity of these cells in a given trial is predictive of memoranda (which refers to WM content) and the quality of memory, indicating that they are behaviorally relevant. This work is preceded by decades of work in animal models, which has shown content-specific persistently active cells in a multitude of brain areas and tasks (Chafee and Goldman-Rakic, 1998; Chelazzi et al., 1998; Funahashi et al., 1989; Isomura et al., 2003; Rainer et al., 1998; Romo et al., 1999; Watanabe and Funahashi, 2007, 2014). Together, this body of work supports the theory that persistently active cells are a mechanism for the maintenance of WM (Kamiński and Rutishauser, 2019). This interpretation is also supported by recent work utilizing causal manipulations to disrupt ongoing persistent activity using optogenetics (Inagaki et al., 2019), which reveals that inhibiting stimulus-selective neurons during WM maintenance erases memoranda as demonstrated by behavior.

An essential aspect of WM that remains poorly understood is the mechanism to maintain multiple pieces of information in WM. It is thought that oscillations play a critical role in this process. Indeed, many electrophysiological studies in both humans and animal models show that oscillations are modulated during working memory tasks (Roux and Uhlhaas, 2014). A particular class of oscillations that has received much interest are 3-7Hz theta-band oscillations. Modulation of theta-band power during WM has been observed in several brain areas, including the Medial Frontal (Jensen and Tesche, 2002; Onton et al., 2005), Dorsolateral Prefrontal (Brzezicka et al., 2019), Temporal (Raghavachari et al., 2001), and Occipital cortex (Palva et al., 2011). While prominent, it remains unclear what specific role these power modulations play in WM maintenance. One model that conceptualizes a potential function of these oscillations in maintaining WM is the Lisman and Idiart model (Lisman and Idiart, 1995), which proposes that different neurons carrying information about WM content are sequentially activated in different gamma cycles during every cycle of the theta oscillation (Lisman and Idiart, 1995). This model also proposes that such oscillation-mediated non-overlapping, fast, and sequential re-activation of items held in WM enables the creation of long term associations facilitated by NMDA-receptor mediated synaptic plasticity (Lisman and Idiart, 1995). Other models propose that neurons carrying information about different items held in WM are activated in separate theta cycles during gamma bursts (Lundqvist et al., 2011). In summary, models thus propose that low-frequency oscillations play a critical role in organizing the neural activity that allows us to maintain information in WM in an active state.

Confirming this view, experiments in macaques show that the activity of individual neurons during WM maintenance is coordinated by ongoing low-frequency oscillations in a stimulus-specific manner (Jacob et al., 2018; Lee et al., 2005; Siegel et al., 2009). However, the existence and relevance of such modulation for human WM remains unknown. In particular, it remains unknown whether oscillations play a role in coordinating the activity of persistently active neurons. Here, we used simultaneous recordings of persistently active neurons and the local field potential (LFP) in the human MTL to examine whether the firing of stimulus-selective cells is shaped by ongoing oscillations and whether such modulation is relevant for WM.

Results

Subjects (20 sessions from 19 patients) performed a WM task (the ‘Sternberg task’). In each trial, subjects were asked to memorize (encode) 1–3 images that were presented sequentially for 1 sec each. After a waiting (maintenance) period of at least 2.5 sec and at most 2.8 sec, subjects were asked to decide whether a probe stimulus was identical or not to one of the 1–3 images held in memory (Fig. 1A). We used in total 5 different images, which were customized for each subject based on a screening task (Kamiński et al., 2017). We refer to the number of images held in memory as ‘load’ throughout this paper. Subjects performed well: Across all loads, the accuracy was 89 ± 5.7% (± s.d.) with a median reaction time of 1.13 ± 0.05 s (correct trials only). Reaction times increased as a function of load (Fig. 1B; F2,38 =7; p=0.003; permuted repeated-measured ANOVA, note that F-statistics and degrees of freedom are provided for reference only, p-values throughout the manuscript are estimated by permutation rather than from the parametric test). Also, accuracy decreased for higher loads (F2,38 =4.8; p=0.01; permuted repeated-measured ANOVA; Load 1: 92%, Load 2: 89%, Load 3: 87%). The subject’s behavior thus showed the classical effect of a load-related increase of reaction time expected in the Sternberg paradigm.

Figure 1. Task, recording locations and single-neuron results.

Figure 1.

Open in a new tab

(A) The task. Top: examples of the screens presented to the participants during a trial (load 2). Bottom: Time of the presentation of each segment in the trial. A trial was composed of 1-3 images to be encoded, presented sequentially. This was followed by the maintenance period, during which participants held the image(s) shown in memory. Next, the probe image was shown, for which subjects judged whether it was presented in this trial or not. (B) Behavioral results. Reaction times increased with load. Thick and light blue lines represent the mean and s.e.m. across all 20 sessions, respectively. ** denote p<0.05; F2,38 =7; p=0.003; permuted repeated-measured ANOVA. (C) Recording locations plotted in Montreal Neurological Institute’s MNI152 coordinates displayed on the California Institute of Technology’s CIT168 T1w atlas of the brain (see Methods). Each dot represents the location of a wire bundle, not an individual neuron. All locations are projected onto x=22.1mm. (D) Percentage of all recorded cells that were stimulus selective during encoding. (E,F,G) Example amygdala cell in different stages of the task: encoding (E), maintenance (F), and probe (G). t=0 is stim onset in (E,G) and maintenance onset in (F). This cell’s firing rate was increased when the cell’s preferred stimulus was held in memory (rate-coding). Top panels show post stimulus time histograms (PSTH; bin size 200 ms, step size 1 ms, shaded areas represent ± s.e.m.). Bottom panel shows raster, trials are reordered for visualization purposes only. Insets show the mean extracellular waveform of all spikes associated with this neuron.

Across all subjects, we isolated 320 putative single neurons in the Amygdala (n=183) and Hippocampus (n=137; Fig. 1C shows the locations of recording sites in MNI [Montreal Neurological Institute] coordinates). Only units which were outside of the lobe that contained the seizure onset zone and which had firing rates >0.1 Hz during the maintenance period were included. Spike sorting quality was assessed quantitatively (Supplementary Fig. 1 and methods). Throughout the manuscript, we use the terms neuron, unit, and cell interchangeably to refer to a putative single neuron.

We first selected neurons which showed a preferential increase in firing for one of the images during encoding of the first image in the trial (“Encoding 1” shown in Fig. 1A, spikes counted in window 0.2 – 1 sec following stimulus onset). We found n=60 selective neurons (18.7%, Fig. 1D; Fig. 1E–G shows an example; selection was done using a permuted one-way ANOVA with X groups, where X=5 was the number of unique images followed by a permutated t-test for the image with the maximal response among all images; see methods).

Next, we assessed whether there was a relationship between the time of spiking and the local field potential for selective neurons during the maintenance period in the theta-frequency range. For this analysis, we only used the subset of selective neurons that fired at least 10 spikes in each of six relevant conditions (loads 1-3, each split into when the preferred image was held in memory or not). n=40/60 neurons satisfied this selection (33 in amygdala, Supplementary Fig. 2A shows the recording locations of this group of neurons). For each selected neuron, we next selected the frequency band (6 possible bands; 1-2, 2-3, 3-4, 4-5, 5-6, 6-7 Hz), wire (7 possible wires in the same wire bundle from which the neuron was recorded), and point of time relative to the spike time (see methods) for which phase-locking strength was maximal (assessed by vector length of circular mean of all phases, see methods). Across the selected wire/frequency band pair of all n=40 neurons, the average spike phase-locking strength during maintenance (0-2.5 sec) was R=0.27. This was significantly higher (p=0.002) than chance as estimated by a null distribution computed using the same approach after jittering spike times relative to the LFP (chance was R=0.237, see methods). We also tested spike-phase locking individually for every neuron using a Rayleigh test at p<0.05, corrected for multiple comparisons using False Discovery Rate (see methods). We found that 39/40 neurons showed significant phase-locking. Neurons preferred different frequencies within the tested 1-7Hz range, with a mean preferred frequency of 3.2 Hz (Fig. 2A).

Figure 2. Spike-Field interactions during WM maintenance.

Figure 2.

Open in a new tab

(A) Distribution of frequencies to which selective neurons showed the strongest spike-phase locking strength. The average frequency is 3.2 Hz. (N=40). (B) Example local field potential (2-3 Hz bandpass filtered – for this cell, this frequency band resulted in the highest spike phase-locking) recorded in the amygdala during maintenance. Dashed lines represent the spike times of a stimulus-selective neuron recorded in the same brain area while the preferred image of the cell was held in memory. Right side inset shows the power spectrum of this channel. (C) Spike-Triggered Average of the neuron shown in (B), computed separately across all spikes in trials in which the preferred image of the neuron was held in memory (red) and not (blue). Shaded areas represent s.e.m. Inset shows the spike waveform of this cell. (D) Average spike phase-locking strength of stimulus-selective neurons (N=40) across all loads. Neurons showed significantly stronger phase locking in the 1-7 Hz band (for each neuron a 1 Hz subband was selected depending on highest spike phase-locking regardless of condition) when the preferred stimulus of a cell was held in WM compared to when it was not. (E) Average spike phase-locking strength shown separately for each load. Phase-locking strength was significantly different between when the preferred image was in memory vs. when it was not only in load 1 (permuted t-test t[39]=3.97, p=0.0005). Phase-locking strength was significantly different between load 1 and load 3 in trials in which the preferred images was not held in memory (permuted t-test t[39]=2.45, p=0.015). For (D,E) blue and red stars represent significant differences relative to the null distribution *p<0.05, **p<0.01, ***p=0.002. The three bars represent, from top to bottom, upper s.e.m., mean, and lower s.e.m.

Next, we tested whether the strength of phase-locking of spikes to the LFP differed between when subjects held the preferred image of a given unit in memory relative to when this image was not held in memory (pooled across loads 1-3). We found that neurons showed stronger phase-locking in trials in which the preferred image was in memory compared to when it was not (Fig. 2B–C shows an example cell; Fig. 2D shows the population summary, number of spikes used were equalized by subsampling, repeated 500 times; permuted t-test t[39]=2.76; p=0.008). The strength of phase-locking in both conditions was significantly higher compered to chance (Fig. 2D, vs. null distribution, p=0.002 and p=0.002 for preferred image held in memory and not held in memory, respectively). As a control, we examined whether oscillatory power on the same wire used for estimating the phase for each neuron differed between when the preferred image was held in memory vs. when it was not. Low frequency power was not significantly different between these two conditions (Supplementary Fig. 2B). Together, this shows that neurons increased the extent to which they fired in synchrony with oscillatory brain activity when their preferred stimulus was held in WM.

Did the relationship shown above between oscillations and cell firing vary as a function of load? For load 1, the strength of spike phase-locking differed significantly between when the preferred stimulus of a cell was held in memory and when it was not (Fig. 2E, permuted t-test t[39]=3.82, p=0.0025) and phase-locking during load 1 trials during which the preferred stimulus was not in memory was not larger than expected by chance (p=0.18). For loads 2 and 3, however, there was no significant difference in phase-locking strength between whether the preferred image was held in memory or not (load 2: t[39]=1.14, p=0.27; load 3: t[39]=0.38, p=0.67). Instead, in contrast to load 1, phase-locking was significantly stronger than expected by chance even when the non-preferred image of a cell was held in memory (load 2: p=0.01; load 3: p=0.002; Fig. 2E stars at bottom). Overall, phase-locking strength increased moderately as a function of load for trials during which the preferred image was not in the memory (linear mixed-effect regression R2 =0.23, t[118]=2.01, p=0.045; blue in Fig. 2E) and phase-locking was significantly stronger comparing load 1 and 3 (Fig. 2E, t-test t[39]=2.45, p=0.015). Together, these results show that the extent of phase-locking of selective neurons was load and memory content dependent, with increased phase locking for trials when the preferred image was not held in memory in higher loads.

While the maximal response of all cells examined is for a given image (“preferred”), some cells might nevertheless also respond to some degree to a subset of the “non-preferred” images. Was the observed increase in phase-locking during non-preferred trials in loads 2-3 driven by such more broadly tuned selective neurons? If so, this effect would be most pronounced during load 3, where it would be most likely that one of the non-preferred images of a cell held in memory also drive the response of this cell to a lesser degree. To test if this was the case, we tested if the firing rates increased with load for trials during which the preferred image of a cell was not in held in memory. We found no significant difference in firing rates (Supplementary Fig. 2C, permutated ANOVA F2,78=2.34; p=0.1), indicating that there was no systematic selectivity for non-preferred images. As an additional control, we also tested whether the extent of the increase in the strength of spike phase-locking between load 1 and 3 for the not-preferred conditions was correlated with the selectivity of a neuron as assessed by the Depth of Selective (DOS) index (Minxha et al., 2017; Rainer et al., 1998). DOS varies between 0 and 1, with 1 indicating that a neuron responds exclusively for only one image. If the increase in phase-locking is driven by broad selectivity, we would expect an inverse relation between phase locking strength increases and DOS. This was not the case: We did not find a significant relation (Supplementary Fig. 2D, r=0.05; p=0.72). This indicates that increases in phase-locking were not driven by differences in selectivity of neurons.

Why did we not observe differences in spike phase-locking strength between trials during which preferred vs. non-preferred stimuli of a cell were held in memory for load 2 and 3 trials? Note that spike-phase locking strength is a second-order metric that is insensitive to differences in phase. Therefore, we next examined the phases of the LFP to which neurons phase-locked. We found that during higher loads, neurons fired spikes at different phases depending on whether the preferred image was held in memory or not (Fig. 3A shows an example neuron). At the population level, we assessed this effect by comparing the difference in phase between the two conditions to a null distribution for each neuron (computed by scrambling condition labels for each spike, n=500). This revealed that, during loads 2 and 3, there was a significant difference in the phase to which spikes phase-locked (Fig. 3B, load 2: p=0.002; load 3: p=0.002). There was no significant difference during load 1 (Fig. 3B, load 1: p=0.24; Supplementary Fig. 2E–G shows examples of STA during all loads and load 3 conditions). We next used a linear regression to test whether there is a relationship between load and the magnitude of phase differences. A linear regression with load as predictor and phase difference as one of the dependent values revealed a significant positive slope (Fig. 3B, linear mixed-effect regression R2=0.1, t[118]=4.4, p=0.003). Together, this indicates that neurons switched from a phase-locking regime to a phase-coding mode in higher memory loads.

Figure 3. Neurons use a phase code to represent information in WM.

Figure 3.

Open in a new tab

(A) Example Spike-Triggered Average (2-3 Hz bandpass filtered – this band resulted in the highest spike phase-locking for this cell regardless of condition) computed for spikes in trials during which the preferred image of the neuron was in memory (red) or not (blue). Only load 3 is shown. Note the phase difference at t=0 between conditions. t=0 is the time of the spike. Shaded areas represent ± s.e.m. (B) Population summary. Absolute circular difference of LFP median phase (1 – 7 Hz band, for each neuron a 1 Hz subband was selected depending on highest spike phase-locking regardless of condition) between spikes in trials in which the preferred stimulus was in WM vs. when it was not. Stars indicating significant difference (p=0.002) vs. resampled data created by assigning spike phases randomly to each condition. Neurons fired spikes at significantly different phases as a function of whether the preferred stimulus is in memory or not for load 2 and 3, but not load 1. Phase differences were significantly larger in load 3 relative to load 1 (t[39]=3.23; p=0.0015). The three bars mark represents, from top to bottom, upper s.e.m., mean, and lower s.e.m. (C) Decoding accuracy for decoding whether the preferred stimulus of a cell is presently held in WM or not based on only spike phase, firing rate or both. Only during load 3 was it possible to decode information from spike phase alone (p=0.002, compared to null distribution, N=500). During load 3, a decoder with access to both spike phase and firing rate performed better than decoders with access only to firing rate or spike phase alone. Dashed line denotes chance level. Stars denote significant differences *p<0.05, **p<0.01, ***p=0.002. See also Supplementary Figure 2.

We next assessed whether spike phase alone was informative by testing whether a decoder (see methods) with only access to spike phase could discriminate between whether the preferred stimulus of a cell was held in memory or not (during the maintenance period). This revealed that, during load 3, but not load 1 and 2, decoding accuracy was higher than chance (Fig. 3C, blue bars; load 1: 0.48, p=0.76; load 2: 0.46, p=0.99; load 3: 0.60, p=0.002; chance was an accuracy of 0.5; p values estimated by comparing to scrambled null distribution, see methods). Was the information provided by phase redundant to that provided by the rate or did it provide additional information? To test this question, we compared the performance of decoders with access to only firing rate or phase with those with access to both. We found that for load 3, the decoder with access to both firing rate and phase performed best compared to decoders with access to only firing rate or phase (decoding accuracy 0.76, 0.65, and 0.60, respectively; Fig. 3C). Together, this data shows that during load 3, the phase to which neurons phase-locked provided additional information on a single-trial level about the stimulus held in memory that was not available in the firing rate alone.

The above data suggests that phase-coding plays a select role only in higher loads. To further examine this observation, we created a computational model to investigate the underlying mechanisms. The model we constructed consists of a network of several excitatory and one inhibitory leaky integrate and fire neuron (Danziger, 2015). The inhibitory neuron received excitatory input from and provided inhibitory input to all four excitatory neurons, thereby providing inhibitory feedback that was proportional to the overall average activity in the network. Each excitatory cell codes for a different possible item in WM (Fig. 4A). We modeled the fact that an item was held in memory by adding an oscillatory input to only the excitatory neuron(s) that represented the items currently held in memory (similar to (Lisman and Idiart, 1995)). We manipulated the load by adding the same oscillatory input to either 1 or 3 neurons (Fig. 4A).

Figure 4. Model to examine experimental results.

Figure 4.

Open in a new tab

(A) The leaky integrate and fire network model. Excitatory neurons (blue and red) are connected with an inhibitory neuron (purple). The inhibitory neuron provides feedback inhibition (white empty circles) to excitatory neurons. Excitatory neurons that represent information currently held in WM receive external oscillatory input and are shown in red. (B) Example activity of excitatory neurons. Upper and lower panel shows results for the case of load 1 and 3, respectively. (C) Spike-Triggered Average (STA) of snippets of excitatory neuron activity around each spike (± 300 ms). Note the strong phase-locking of the neurons activity in load 3 even if the neurons preferred stimulus is not held in WM. (D,E) Average firing rate of neurons relative to the peak of the subthreshold oscillatory input (± 150 ms) during load 1 (D) and load 3 (E). In load 3, there is a decrease of tonic activity of neurons whose preferred stimuli are not presently held in WM (blue). This effect is the result of feedback inhibition. (F) Firing rate of neurons during different conditions. During simulations with low noise levels, both rate-and phase encode information about WM content. For large noise levels, in contrast, only phase but not rate coding was present. (G) Same as (C) but for the simulations with large noise levels. For (G,C) shaded areas represent ± s.e.m. See also Supplementary Figure 3 and 4.

We simulated the above model for different loads and quantified the spike-phase locking strength of the simulated neurons to compare the behavior of the network to the experimental data. First, we found that with increasing load, neurons representing stimuli not held in memory started to increase their phase locking (Fig. 4B–C, see Supplementary Fig. 3A for load 2). This is because of feedback inhibition, which increased as a function of load. As a result, the spontaneous activity of neurons which represent items not currently held in memory becomes phase-locked due to the shorter window of time available for them to fire action potentials (Fig. 4D–E). Moreover, because this phase-locking was caused by feedback inhibition rather than direct oscillatory input, the phase at which the two types of neurons (those coding and not coding for information currently in memory) fired spikes was different (Fig. 4D–E). This is similar to what we observed in the experimental data (Fig. 2 and 3). The model produces this phenomenon robustly across a range of values for the frequencies of the oscillatory input (2–5 Hz, Supplementary Fig. 3B–D). Note that the oscillatory frequency during which the network operates is primarily defined by the neuronal time constant we used (we used 75 ms). Changing this time constant will allow the network to operate in different frequency ranges (see Supplementary Fig. 3C–D for network performance expressed as a ratio relative to the neuronal time constant).

We also varied the amplitude of the noise that the network received as input. We calibrated the amount of noise such that there was no longer a difference in firing rate between when the preferred stimulus of a neuron was held in memory and when it was not (Fig. 4F, right). This revealed that, in this situation, the strength of phase-locking and the phase to which spikes phase-lock still differed between conditions even when firing rate did not (Supplementary Fig. 3B, Fig. 4F–G). Thus, in higher loads, the model network could perform in a regime in which only spike phase, but not firing rate, of a neuron would be informative of whether a stimulus is currently held in memory. We next verified this model prediction in the experimental data. To do so, we selected the subset of analyzed neurons (n=25/40, 21 in Amygdala) that did not showed a significant increase in firing rate to the preferred stimulus during the maintenance period (0-2.5 sec). As predicted by the model, this group of neurons had significantly higher spike-phase locking strength in their pre-defined frequency band if the preferred image (defined during encoding) was held in memory compared to when it was not. Also, as predicted, this effect was present in load 1 but not loads 2-3 (Supplementary Fig. 4A). Moreover, we observed that phase differences between trials when the preferred image was held in WM and not were significantly higher than expected by chance during load 2 (p=0.005) and load 3 (p=0.002), but not load 1 (p= 0.6356; Supplementary Fig. 4B). This experimental result thus demonstrates that the timing of spikes of neurons whose firing rate was not modulated by memory content was nevertheless modulated by ongoing low-frequency oscillation such that spike phase was informative about memory content. This model prediction was thus borne out in the experimental data.

Discussion

Our results show that the time at which cells carrying information during WM maintenance fire spikes is not random but rather is shaped by the oscillatory activity of the LFP. Similar spike-field relationships have previously been reported for other tasks. For instance, single neurons in human prefrontal and various parts of the MTL encode goals during virtual navigation using the phase of low-frequency oscillations (3 Hz) at the time a neuron spikes (Watrous et al., 2018). Furthermore, in a study of macaque prefrontal cortex neurons, phase-coding during WM maintenance was also observed at the single-neuron level for both low-and high frequency oscillations (3 Hz and 32 Hz, Siegel et al., 2009). Also, neurons representing task relevant-and irrelevant information during WM fire at different phases of theta oscillations (Jacob et al., 2018). Here, we show that human MTL neurons (located predominantly in the Amygdala) phase-lock their activity to low-frequency oscillations during the maintenance of multiple items in WM. The role of theta oscillation in the Amygdala has been investigated extensively in the context of processing of emotional stimuli (Paré and Collins, 2000; Taub et al., 2018; Zheng et al., 2017, 2019). For instance, the recognition of emotional stimuli is associated with higher phase-amplitude coupling between theta and gamma oscillation (Zheng et al., 2019). In our analysis, we observe that spike-phase locking to theta oscillation also plays a role in WM maintenance. Extending our previous observation of persistent activity in the Amygdala (Kamiński et al., 2017), the result shown here add additional evidence to the hypothesis that the Amygdala plays a role in supporting WM.

The frequency of the LFP to which neurons in the MTL phase-lock in our study and previous studies (including those discussed above) is relatively low. This raises the question of whether the oscillations that these neurons phase-lock to can be considered as theta. In the rat hippocampus, theta is in the 4-10 Hz band (Buzsáki, 2005). However, in humans, oscillations thought to be theta have been observed with frequencies as low as 2 Hz (Lega et al., 2012). Also, in humans, the frequency of theta oscillations depends on the recording location along the anterior-posterior axis (Goyal et al., 2018), with the posterior parts of hippocampus exhibiting oscillations at 8 Hz and anterior parts exhibiting oscillations at lower frequencies of about 3 Hz. A second distinction along the same axis is that posterior theta correlates with movement speed, whereas anterior theta does not (Goyal et al., 2018). This observation suggests that there are two functionally distinct theta oscillation in the human hippocampus, one involved in navigation (posterior) and one in cognition (anterior). Our recordings were performed in the anterior hippocampus and exhibited low-frequency oscillations. We therefore hypothesize that what our neurons phase-locked to where low-frequency theta oscillations.

Our model proposes a neuronal mechanism to explain the observation that, for higher loads, neurons transition from phase-locking for preferred stimuli only to phase-coding. We propose that this can be explained as a result of shared inhibitory feedback (Douglas and Martin, 2004; Rutishauser et al., 2015) . For example, inhibitory neurons in V1 receive excitatory input from local pyramidal cells with different receptive fields (Bock et al., 2011). This effect is consistent with the finding that inhibitory neurons are less stimulus-selective than pyramidal neurons (Ison et al., 2011; Sohya et al., 2007). Inhibitory neurons, in turn, create local dense reciprocal connections with pyramidal neurons (Fino and Yuste, 2011), thereby closing the loop. In our model, such shared inhibition results in phase-locking of spiking because the oscillatory inhibitory input to the excitatory neurons restricts the periods of time during which the neurons can fire. The phase of the oscillatory inhibitory input is different from that of the external oscillatory input provided to the neurons that represent items held in memory. As a result, neurons that encode information currently in memory become active at a phase different from that of neurons that encode an item not currently held in memory. This feature of the network makes it possible to decode memory content from neurons even if they do not exhibit a firing rate difference as a function of which item is held in memory. Of note, persistently active neurons decrease their firing rate for the preferred stimulus as a function of load (Kamiński et al., 2017). Here, we show that phase-coding is a potential mechanism to counteract this reduction in information.

More broadly, the phase-coding we observed could be used as a mechanism for selective routing of information (Fries, 2015). The oscillatory phase alignment with downstream neuronal targets increases the probability of successfully evoking an action potential in the downstream targets (Gupta et al., 2016; Womelsdorf et al., 2007). Thus, only information carried by spikes fired at a particular phase would be passed on to downstream targets in a network. Our model indicates that this mechanism of temporal organization of spiking activity could also be used for routing information represented by neurons who do not exhibit rate-coding. In this way, we hypothesize that activity-silent (Stokes, 2015) neurons might code for information held in WM by being part of a specific neuronal assembly (Womelsdorf et al., 2007).

STAR Methods

Lead Contact and Materials Availability

Further information and requests for resources should be directed to the Lead Contact, Ueli Rutishauser (urut@caltech.edu).

Experimental Model and Subject Details

19 subjects participated in the study (Supplementary Table 1). All of them were implanted with depth electrodes for the surgical treatment of epilepsy. All subjects volunteered for the study and gave informed consent. This study was approved by the Institutional Review Boards of Cedars-Sinai Medical Center and Huntington Memorial Hospital.

Method Details

Task

We used the same modified Sternberg task as described in (Kamiński et al., 2017). Briefly, every trial started with a fixation cross (0.9-1 sec, Fig. 1A). Next, the to be memorized images were presented sequentially (1 sec followed by blank screen 16-200 ms) for one up to three images. Then, there was a maintenance period (2.5-2.8 sec). Lastly, at the end of the maintenance period, the probe image was shown. Subjects were asked to judge if the probe image was present in the trials or not. They were asked to respond as fast as possible using one of two buttons on the response pad.

A set of 5 images was chosen for each subject based on a screening task run 2-3 hours before the WM task (Quiroga et al., 2005). The screening task was composed of 54-64 images shown 6 times for 1 second in randomized order. Images which evoked the most selective response where used in the later WM task. For details, see (Kamiński et al., 2017).

Electrophysiology

Each macroelectrode contained eight 40 μm diameter microwires (Minxha et al., 2018; Rutishauser et al., 2010). We recorded broadband (0.1-9000Hz filter) signal from a total of 32 channels sampled at 32 kHz using a Neuralynx Atlas system. Signals were locally referenced to one of the eight microwires in a given brain area.

Localization of electrodes

Electrodes were localized based on pre-and post-operative T1 structural MRIs. We extracted the brains from the pre-and post-operative T1 scans (Ségonne et al., 2004) and aligned the post-operative to the pre-operative scan with Freesurfer’s mri_robust_register (Reuter et al., 2010). We then computed a forward mapping of the pre-operative scan to the CIT168 template brain (Tyszka and Pauli, 2016) using a concatenation of an affine transformation followed by a symmetric image normalization (SyN) diffeomorphic transform computed by the ANTs suite of programs (Avants et al., 2008). This resulted in a post-operative scan overlayed on the MNI152-registered version of the CIT168 template brain (Tyszka and Pauli, 2016). Note that this procedure does not require ability to align the post-operative scan to the template brain, because this is done based on the pre-operative scan (see Minxha et al., 2017 for details). We then used the Freesurfer’s Freeview program to mark the electrodes as point sets to determine where the tips of the microwires were located. For visualization only, electrode positions were projected onto the 2D sagittal plane.

Quantification and Statistical Analysis

Spike sorting and quality metrics of single units

For sorting, the recorded signal was filtered with a zero-phase lag filter in the 300-3000Hz band and spikes were detected and sorted using a semi-automated template-matching algorithm (Rutishauser et al., 2006). We computed several spike sorting quality metrics for all identified putative single-neurons to assess the quality of identified units (Supplementary Fig. 1): 1) the percentage of interspike intervals (ISIs) below 3 ms was 0.42% ± 0.79%, 2) the ratio between the standard deviation of the noise and the peak amplitude of the mean waveform of each cluster was 6.96 ± 4.88 (peak SNR), 3) the pairwise projection distance in clustering space between all neurons isolated on the same wire was 13.28 ± 7.66 (projection test; in units of s.d. (Pouzat et al., 2002) of the signal), 4) the modified coefficient of variation of variability in the ISI (CV2) was 0.96 ± 0.14, 5) the median isolation distance (Harris et al., 2000) was 31. We calculated the isolation distance in a ten-dimensional feature space (energy, peak amplitude, total area under the waveform and first five principal components of the energy normalizes waveforms (Harris et al., 2000).

Statistical approach

Statistical comparisons were conducted using permutation tests based on a null distribution estimated from B=2000 runs on data with scrambled labels using the EEGLAB toolbox (Delorme and Makeig, 2004). For Supplementary Fig. 2D we used the Depth of selectivity (DOS) index as defined in (Minxha et al., 2017; Rainer et al., 1998) to measure how selective the response of a given neuron is across all used stimuli. For regresions we used linear mixed-effect regression with fixed effect of load and random effect of cluster id.

To test whether an individual cell exhibited significant spike-phase locking we used the Rayleigh test at p<0.05. We corrected for multiple comparisons using False Discovery rate (taking into account the 7 wires, 6 frequency bands and 250 samples around a spike (± 250 ms) which were used to find the highest spike-phase locking).

Selection of neurons

For each recorded neuron, we ran the following two statistical tests to determine whether a cells response is selective for one of the five images shown during the experiment. First, we ran a permuted one-way ANOVA with 5 groups (the number of unique images) during encoding of the first image in the trial (0.2 – 1 sec). Second, we tested whether the activity for the image with the maximal response was significantly different relative to all other images (permutation t-test). A cell was considered selective only if both tests were satisfied at p<0.05.

Spike Phase-Locking strength and Spike-Phase differences

Before computing Spike-Field interactions, we subtracted the average spike waveform from each point of time at which a spike was detected (from −1.5 to 2 ms relative to the detected spike) from the raw LFP signal (Zanos et al., 2012). To avoid contamination by evoked potentials following stimulus onset, we only used spikes between the period of time starting 450 ms after onset of and ending 450 ms before the end of the maintenance period. We also rejected spikes if the z-scored value of the LFP associated with them (± 450 ms around spike) was higher than z=3.5 at any point of time. Next we filtered the LFP signal in six 1-Hz wide subbands (1-2, 2-3, 3-4, 4-5, 6-7 Hz) using a Hamming windowed FIR filter for each of the seven wires recorded from in a given brain area (we recorded from 8 wires in each, one of which was the reference). For each neuron, we chose the wire and subband in which spike phase-locking strength was maximal. We quantified spike phase-locking strength of a neuron as the resultant length of the circular mean of the phases across all spikes (collapsed across all loads and preferred/non-preferred conditions, making this selection unbiased). The phase of a spike was extracted by applying the Hilbert transform to the narrow-band filtered LFP. Because the highest concentration of LFP phases can occur at points of time different from when the neuron spikes (Siapas et al., 2005), we computed spike-phase locking using different temporal offsets (up to ± 250 ms) between spikes and LFP and used the offset that resulted in the highest spike-phase locking strength. To compute the difference in phases between conditions we used the absolute circular difference between the median phases of each group.

To compare the strength of spike phase-locking between experimental conditions, we equalized the number of spikes in each condition by taking a random subset of spikes from the condition with more spikes. We repeated this procedure 500 times and compute the average spike phase-locking strength. We only used neurons which had at least 10 spikes in each of the 6 conditions we used: loads 1-3, each split into trials in which the preferred image of the cell was in memory or not. For Fig. 2D–E and Supplementary Fig. 4A, we estimated the null distribution using the same methodology as for the recorded data (choosing highest spike phase-locking across all wires, sub-bands and 250 samples around each spike [± 250 ms]) after first adding a random offset to the point of time of every spike (± 450 ms around spike, uniform; referred to as ‘spike time shuffling’). As an additional control, we also used interspike interval (ISI) shuffling to estimate the null distribution. We found very similar results compared to spike time shuffling: the average of the resampled distribution of phase-locking values after ISI shuffling was R=0.22, which was similar to the time shuffling result (R=0.237). Also, the recorded spike-field locking values were significantly bigger (p=0.002, N=500) than chance as estimated from ISI shuffling. To compute the null distribution used in Fig. 3B and Supplementary Fig. 4B, we randomly assigned phases of spikes to conditions and computed the difference. This procedure was repeated 500 times.

Decoding

For decoding whether the preferred image was held in memory or not we used a pseudopopulation of neurons pooled across all subjects. Only neurons which had at least 10 spikes in each condition were used (N=40). We used a linear support-vector machine decoder as implemented in the ndt toolbox (Meyers, 2013). For decoding of phase, we used the cosine of each angle as variables. We excluded trials with no spikes for this analysis because in this case the spike phase could not be estimated. Leave-one-out cross-validation was used to estimate the performance of each decoder. Out of all possible train-test combinations, we used 500 randomly chosen train-test sets to estimate the cross-validated testing error. To estimate if a decoder accuracy was significantly above chance we compared decoder performance with a null distribution that was estimated using same approach as described but with scrambled labels (repeated 500 times).

Modeling

We used a network composed of leaky integrate-and-fire neurons. We simulated this network using the SimLIFNet function (Danziger, 2015). All parameters have been scaled such that the spiking threshold is 1 and the resting potential is 0 (Lewis and Rinzel, 2003). The derivate of the membrane potential is:

dVidt=−Vi+Iapp+Isyn

Iapp is an external forcing current applied to all cells (representing the sum of the noise and external oscillatory input). Isyn are synaptic inputs from other neurons in the network contributions:

Isyn=∑[a2*e−a(t−Kj)*(t−Kj)]*Wji

a is a synaptic density parameter and Kj is time of the vector of spike times from neuron j. Wji is a coupling strength from the jth neuron to ith neuron. We modeled the fact that given information was held in memory by adding a zero-mean oscillatory input to only the neuron(s) that represented the stimuli currently held in memory. For the model shown in Fig. 4, the frequency of the oscillation was set to 4.1 Hz, so one cycle of an oscillation spanned 3.23x of the neuronal time constant (which here was set to 75 ms (Ehrlich et al., 2012)). The amplitude of oscillatory input for the model in Fig. 4F–G was decreased to 40%. The amplitude of the gaussian noise was 0.25 for the first simulation (Fig. 4B–F) and 0.75 for the second (Fig. 4F–G). Additionally, we added a 0.75 amplitude bias current that was injected continuously into each pyramidal cell. The strength of the connections from pyramidal to inhibitory neurons was set to 1. The strength of the connections from the inhibitory to the pyramidal neurons was −0.9. We tested different values of all the parameters used and found that our results held for a wide range of the critical parameters (Supplementary Fig. 3).

Data and Software Availability

The spike detection and sorting toolbox OSort and the EEGLAB toolbox was used for data processing, both of which are available as open source. Data and custom MATLAB analysis scripts are available upon reasonable request from Ueli Rutishauser (ueli.rutishauser@cshs.org).

Supplementary Material

2

Key Resources Table

REAGENT or RESOURCE SOURCE IDENTIFIER
Software and Algorithms
MATLAB R2016a MathWorks SCR_001622
OSort http://www.rutishauserlab.org/osort SCR_015869
Psychophysics toolbox PTB3 http://psychtoolbox.org SCR_002881
EEGLAB N/A SCR_007292
Other
Neuralynx Neurophysiology System Neuralynx Inc Cat# ATLAS 128
Cedrus Reponse Box Cedrus (https://cedrus.com/) Cat#RB-844
Open in a new tab

Acknowledgments

We thank the staff of the Epilepsy Monitoring Unit at Cedars-Sinai Medical Center for assistance, Jeffrey M. Chung and Chrystal M. Reed for patient care and facilitating experiments, and all patients for their participation. This work was supported by a National Science Foundation CAREER award (1554105 to U.R.), the National Institute of Health (R01MH110831 and U01NS098961 to U.R.) and the McKnight Endowment Fund for Neuroscience (to U.R.).

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Declaration of interests

The authors declare no competing interests.

Bibliography

  1. Avants B, Duda JT, Kim J, Zhang H, Pluta J, Gee JC, and Whyte J (2008). Multivariate Analysis of Structural and Diffusion Imaging in Traumatic Brain Injury. Acad. Radiol 15, 1360–1375. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Baddeley A (2007). Working Memory, Thought, and Action (Oxford University Press; ). [Google Scholar]
  3. Bock DD, Lee WCA, Kerlin AM, Andermann ML, Hood G, Wetzel AW, Yurgenson S, Soucy ER, Kim HS, and Reid RC (2011). Network anatomy and in vivo physiology of visual cortical neurons. Nature 471, 177–184. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Brzezicka A, Kammski J, Reed CM, Chung JM, Mamelak AN, and Rutishauser U (2019). Working Memory Load-related Theta Power Decreases in Dorsolateral Prefrontal Cortex Predict Individual Differences in Performance. J. Cogn. Neurosci 1–18. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Buzsáki G (2005). Theta rhythm of navigation: Link between path integration and landmark navigation, episodic and semantic memory. Hippocampus. [DOI] [PubMed] [Google Scholar]
  6. Chafee MV, and Goldman-Rakic PS (1998). Matching patterns of activity in primate prefrontal area 8a and parietal area 7ip neurons during a spatial working memory task. J. Neurophysiol 79, 2919–2940. [DOI] [PubMed] [Google Scholar]
  7. Chelazzi L, Duncan J, Miller EK, and Desimone R (1998). Responses of neurons in inferior temporal cortex during memory-guided visual search. J Neurophysiol 80, 2918–2940. [DOI] [PubMed] [Google Scholar]
  8. Danziger Z (2015). https://www.mathworks.com/matlabcentral/fileexchange/50339-easily-simulate-a-customizable-network-of-spiking-leaky-integrate-and-fire-neurons.
  9. Delorme A, and Makeig S (2004). EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J Neurosci Methods 134, 9–21. [DOI] [PubMed] [Google Scholar]
  10. Douglas RJ, and Martin KAC (2004). Neuronal Circuits of the Neocortex. Annu. Rev. Neurosci 27, 419–451. [DOI] [PubMed] [Google Scholar]
  11. Ehrlich DE, Ryan SJ, and Rainnie DG (2012). Postnatal development of electrophysiological properties of principal neurons in the rat basolateral amygdala. J. Physiol [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Fino E, and Yuste R (2011). Dense inhibitory connectivity in neocortex. Neuron 69, 1188–1203. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Fries P (2015). Rhythms for Cognition: Communication through Coherence. Neuron 88, 220–235. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Funahashi S, Bruce CJ, and Goldman-Rakic PS (1989). Mnemonic coding of visual space in the monkey’s dorsolateral prefrontal cortex. J. Neurophysiol 61, 331–349. [DOI] [PubMed] [Google Scholar]
  15. Goyal A, Miller J, Qasim S, Watrous AJ, Stein JM, Inman CS, Gross RE, Willie JT, Lega B, Lin J-J, et al. (2018). Functionally distinct high and low theta oscillations in the human hippocampus. BioRxiv 498055. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Gupta N, Singh SS, and Stopfer M (2016). Oscillatory integration windows in neurons. Nat. Commun 7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Harris KD, Henze DA, Csicsvari J, Hirase H, and Buzsaki G (2000). Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular measurements. J. Neurophysiol 84, 401–414. [DOI] [PubMed] [Google Scholar]
  18. Inagaki HK, Fontolan L, Romani S, and Svoboda K (2019). Discrete attractor dynamics underlies persistent activity in the frontal cortex. Nature 566, 212–217. [DOI] [PubMed] [Google Scholar]
  19. Isomura Y, Ito Y, Akazawa T, Nambu A, and Takada M (2003). Neural coding of “attention for action” and “response selection” in primate anterior cingulate cortex. J. Neurosci 23, 8002–8012. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Ison MJ, Mormann F, Cerf M, Koch C, Fried I, and Quiroga RQ (2011). Selectivity of pyramidal cells and interneurons in the human medial temporal lobe. J. Neurophysiol 106, 1713–1721. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Jacob SN, Hähnke D, and Nieder A (2018). Structuring of Abstract Working Memory Content by Fronto-parietal Synchrony in Primate Cortex. Neuron 99, 588–597.e5. [DOI] [PubMed] [Google Scholar]
  22. Jensen O, and Tesche CD (2002). Frontal theta activity in humans increases with memory load in a working memory task. Eur. J. Neurosci 15, 1395–1399. [DOI] [PubMed] [Google Scholar]
  23. Kamiński J, and Rutishauser U (2019). Between persistently active and activity-silent frameworks: novel vistas on the cellular basis of working memory. Ann. N. Y. Acad. Sci nyas.14213. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Kamiński J, Sullivan S, Chung JM, Ross IB, Mamelak AN, and Rutishauser U (2017). Persistently active neurons in human medial frontal and medial temporal lobe support working memory. Nat. Neurosci 20, 590–601. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Kornblith S, Quian Quiroga R, Koch C, Fried I, and Mormann F (2017). Persistent Single-Neuron Activity during Working Memory in the Human Medial Temporal Lobe. Curr. Biol 27, 1026–1032. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Lee H, Simpson GV, Logothetis NK, 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. [DOI] [PubMed] [Google Scholar]
  27. Lega BC, Jacobs J, and Kahana M (2012). Human hippocampal theta oscillations and the formation of episodic memories. Hippocampus 22, 748–761. [DOI] [PubMed] [Google Scholar]
  28. Lewis TJ, and Rinzel J (2003). Dynamics of spiking neurons connected by both inhibitory and electrical coupling. J. Comput. Neurosci 14, 283–309. [DOI] [PubMed] [Google Scholar]
  29. Lisman JE, and Idiart MA (1995). Storage of 7 +/− 2 short-term memories in oscillatory subcycles. Science (80-.). 267, 1512–1515. [DOI] [PubMed] [Google Scholar]
  30. Lundqvist M, Herman P, and Lansner A (2011). Theta and gamma power increases and alpha/beta power decreases with memory load in an attractor network model. J. Cogn. Neurosci 23, 3008–3020. [DOI] [PubMed] [Google Scholar]
  31. Meyers EM (2013). The Neural Decoding Toolbox. Front. Neuroinform 7, 8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Minxha J, Mosher C, Morrow JK, Mamelak AN, Adolphs R, Gothard KM, and Rutishauser U (2017). Fixations Gate Species-Specific Responses to Free Viewing of Faces in the Human and Macaque Amygdala. Cell Rep. 18, 878–891. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Minxha J, Mamelak AN, and Rutishauser U (2018). Surgical and electrophysiological techniques for single-neuron recordings in human epilepsy patients In Neuromethods, (New York: Humana Press; ), pp. 267–293. [Google Scholar]
  34. Onton J, Delorme A, and Makeig S (2005). Frontal midline EEG dynamics during working memory. Neuroimage 27, 341–356. [DOI] [PubMed] [Google Scholar]
  35. Palva S, Kulashekhar S, Hamalainen M, and Palva JM (2011). Localization of Cortical Phase and Amplitude Dynamics during Visual Working Memory Encoding and Retention. J. Neurosci [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Paré D, and Collins DR (2000). Neuronal correlates of fear in the lateral amygdala: Multiple extracellular recordings in conscious cats. J. Neurosci [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Pouzat C, Mazor O, and Laurent G (2002). Using noise signature to optimize spike-sorting and to assess neuronal classification quality. J. Neurosci. Methods 122, 43–57. [DOI] [PubMed] [Google Scholar]
  38. Quiroga RQ, Reddy L, Kreiman G, Koch C, and Fried I (2005). Invariant visual representation by single neurons in the human brain. Nature 435, 1102–1107. [DOI] [PubMed] [Google Scholar]
  39. Raghavachari S, Kahana MJ, Rizzuto DS, Caplan JB, Kirschen MP, Bourgeois B, Madsen JR, and Lisman JE (2001). Gating of human theta oscillations by a working memory task. J. Neurosci 21, 3175–3183. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Rainer G, Asaad WF, and Miller EK (1998). Selective representation of relevant information by neurons in the primate prefrontal cortex. Nature 393, 577–579. [DOI] [PubMed] [Google Scholar]
  41. Reuter M, Rosas HD, and Fischl B (2010). Highly accurate inverse consistent registration: A robust approach. Neuroimage 53, 1181–1196. [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Romo R, Brody CD, Hernandez A, and Lemus L (1999). Neuronal correlates of parametric working memory in the prefrontal cortex. Nature 399, 470–473. [DOI] [PubMed] [Google Scholar]
  43. Roux F, and Uhlhaas PJ (2014). Working memory and neural oscillations: Alpha-gamma versus theta-gamma codes for distinct WM information? Trends Cogn. Sci 18, 16–25. [DOI] [PubMed] [Google Scholar]
  44. Rutishauser U, Schuman EM, and Mamelak AN (2006). Online detection and sorting of extracellularly recorded action potentials in human medial temporal lobe recordings, in vivo. J. Neurosci. Methods 154, 204–224. [DOI] [PubMed] [Google Scholar]
  45. Rutishauser U, Ross IB, Mamelak AN, and Schuman EM (2010). Human memory strength is predicted by theta-frequency phase-locking of single neurons. Nature 464, 903–907. [DOI] [PubMed] [Google Scholar]
  46. Rutishauser U, Slotine JJ, and Douglas R (2015). Computation in Dynamically Bounded Asymmetric Systems. PLoS Comput. Biol 11. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Ségonne F, Dale AM, Busa E, Glessner M, Salat D, Hahn HK, and Fischl B (2004). A hybrid approach to the skull stripping problem in MRI. Neuroimage 22, 1060–1075. [DOI] [PubMed] [Google Scholar]
  48. Siapas AG, Lubenov EV, and Wilson MA (2005). Prefrontal phase locking to hippocampal theta oscillations. Neuron. [DOI] [PubMed] [Google Scholar]
  49. Siegel M, Warden MR, and Miller EK (2009). Phase-dependent neuronal coding of objects in short-term memory. Proc. Natl. Acad. Sci 106, 21341–21346. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Sohya K, Kameyama K, Yanagawa Y, Obata K, and Tsumoto T (2007). GABAergic neurons are less selective to stimulus orientation than excitatory neurons in layer M/MI of visual cortex, as revealed by in vivo functional Ca2+ imaging in transgenic mice. J. Neurosci 27, 2145–2149. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Stokes MG (2015). ‘Activity-silent’ working memory in prefrontal cortex: a dynamic coding framework. Trends Cogn. Sci 19, 394–405. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Taub AH, Perets R, Kahana E, and Paz R (2018). Oscillations Synchronize Amygdala-to-Prefrontal Primate Circuits during Aversive Learning. Neuron. [DOI] [PubMed] [Google Scholar]
  53. Tyszka JM, and Pauli WM (2016). A high resolution in vivo MRI atlas of the adult human amygdaloid complex. Hum. Brain Mapp 37, 3979–3998. [DOI] [PMC free article] [PubMed] [Google Scholar]
  54. Watanabe K, and Funahashi S (2007). Prefrontal delay-period activity reflects the decision process of a saccade direction during a free-choice ODR task. Cereb. Cortex 17, i88–i100. [DOI] [PubMed] [Google Scholar]
  55. Watanabe K, and Funahashi S (2014). Neural mechanisms of dual-task interference and cognitive capacity limitation in the prefrontal cortex. Nat. Neurosci 17, 601–611. [DOI] [PubMed] [Google Scholar]
  56. Watrous AJ, Miller J, Qasim SE, Fried I, and Jacobs J (2018). Phase-tuned neuronal firing encodes human contextual representations for navigational goals. Elife 7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Womelsdorf T, Schoffelen JM, Oostenveld R, Singer W, Desimone R, Engel AK, and Fries P (2007). Modulation of neuronal interactions through neuronal synchronization. Science (80-. ). 316, 1609–1612. [DOI] [PubMed] [Google Scholar]
  58. Zanos S, Zanos TP, Marmarelis VZ, Ojemann GA, and Fetz EE (2012). Relationships between spike-free local field potentials and spike timing in human temporal cortex. J. Neurophysiol [DOI] [PMC free article] [PubMed] [Google Scholar]
  59. Zheng J, Anderson KL, Leal SL, Shestyuk A, Gulsen G, Mnatsakanyan L, Vadera S, Hsu FPK, Yassa MA, Knight RT, et al. (2017). Amygdala-hippocampal dynamics during salient information processing. Nat. Commun [DOI] [PMC free article] [PubMed] [Google Scholar]
  60. Zheng J, Stevenson RF, Mander BA, Mnatsakanyan L, Hsu FPK, Vadera S, Knight RT, Yassa MA, and Lin JJ (2019). Multiplexing of Theta and Alpha Rhythms in the Amygdala-Hippocampal Circuit Supports Pattern Separation of Emotional Information. Neuron. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

2
NIHMS1569020-supplement-2.pdf (535.7KB, pdf)

Data Availability Statement

The spike detection and sorting toolbox OSort and the EEGLAB toolbox was used for data processing, both of which are available as open source. Data and custom MATLAB analysis scripts are available upon reasonable request from Ueli Rutishauser (ueli.rutishauser@cshs.org).

RESOURCES