An Investigation of Localist and Distributed Representation Models using the
Neuronal Coherence of Concept Cell Activity with the Human Medial Temporal
Lobe
An Internship ReportDate: July 2016
Author: Suzanne Martens Student number: 10194045
Supervisor: Conrado Bosman (Daily Supervisor: Tom Sikkens) Co-assessor: Umberto Olcese
2 An Investigation of Localist and Distributed Representation Models using the Neuronal Coherence of
Concept Cell Activity with the Human Medial Temporal Lobe
Abstract Introduction
Since the discovery of concept cells in the human hippocampus, scientists have debated whether their existence supports localist (LM) or distributed representation models (DRM). LM view concept cells as the last step in the hierarchy of visual processing, while DRM think of concept cells as parts of high-level memory networks. Implicitly, LM assume that concept cells primarily communicate with low-level downstream neurons while DRM assume that they primarily communicate with (relatively) nearby members of the network they belong to. Using the ING/PING model of communication within neuronal ensembles, this study investigates whether concept cells behave like they are part of neuronal ensembles (supporting DRM) or not (supporting LM).
Methods
Data was collected from 16 patients suffering from pharmacologically intractable epilepsy. Patients were implanted with chronic depth electrodes with micro-wires at their tips for 10 – 14 days. During this time, patients were asked to perform a computerized task designed for finding concept cells that consisted of viewing pictures of celebrities, famous places and personally relevant people or locations. Results
A total of 13 concept cells was found and analyzed. Local LFP signals of tissue directly surrounding the cells showed a sharp increase of power in the low gamma band and weaker increases in all other frequencies 0.2-0.6s after presenting a preferred stimulus. None of these effects were observed when other pictures were presented. Local LFP signals slightly further removed from the cells showed an increase in coherence in the low gamma band 0.1-0.7s after presenting other pictures as well as a short decrease in high gamma (0-0.2s) and a late onset increase in theta (0.5-0.7s). None of these effects were observed when preferred pictures were presented. Finally, an increase in spike-LFP coherence in the low and high gamma band was observed between spikes of the concept cells and the LFPs when preferred stimuli were presented compared to when other stimuli were presented, as well as a decrease in the alpha band.
Conclusions
The results from the WPLI and spike-LFP coherence analyses are in line with ING/PING models and support the idea that concept cells behave like they are part of neuronal ensembles. Therefore, this study supports DRM over LM. These findings need to be confirmed in a study using larger sample sizes per picture and a larger population of concept cells.
3 §1. Introduction
The field of neuroscience has regularly studied the brain by isolating a mechanism, function or region of interest and investigating its properties without considering the rest of the system (Fodor, 1983; Gazzaniga, Ivry, & Mangun, 2008). On the neuronal level, studies isolating neurons and manipulating the input they receive have revealed how neurons generate signals and how specific neurons can be in the stimuli they respond to (Hubel & Wiesel, 1965; Quiroga et. al., 2005). This research has inspired the so called localist representation model of neuronal representation of the outside world (Gross, 2002). The localist representation model proposes that concepts are
represented by single neurons, and single neuron activity can therefore have meaning by itself (Barlow, 1972; Bowers, 2010). However, it is also clear that the brain is highly interconnected, both locally and distantly (Achard et. al., 2006; Van den Heuvel et. al., 2008). Activity of neurons could therefore always be influenced by neighboring neurons, or even by activity of distant brain regions. Models of distributed representation take this interconnectivity in to account and state that the activity of neurons by themselves is meaningless. Instead, concepts are proposed to be represented by the combined activity in groups of neurons (Plaut & McClelland, 2010). Though distributed
representation models have long been favored, the discussion has recently been revived by advocates of the localist representation models (Bowers, 2010; Roy, 2012; Roy, 2013). The models implicitly assume different strategies of communication between neurons. Localist models assume that
activation of a single neuron is sufficient to represent a concept and therefore would mainly expect to see communication between single neurons representing a concept, distant neurons that process lower level information and distant brain areas that need the information a concept cell represents (e.g. the prefrontal cortex). Distributed representation models on the other hand assume activity of many nearby neurons is needed to represent a concept and therefore expect communication within neuronal ensembles on top of communication with downstream neurons. The current study aims to contribute to this discussion by investigating patterns of communication between cells that show high specificity for certain stimuli (and could therefore at least potentially represent a concept in a way consistent with the localist view) and other nearby neurons.
§1.1 Concept Cells
Certain single neurons in the hippocampus of the medial temporal lobe (MTL) show high specificity for stimuli. These cells were first discovered in the human cortex by Quiroga et. al. (2005). These cells respond selectively to pictures of specific people (e.g. Luke Skywalker or Jennifer Aniston), but not to any of the other presented pictures. Moreover, the cells respond to various different stimuli of the same concept. For example, they invariantly respond to pictures of the person in question as different movie characters they played and to their written or spoken names. The cells therefore do not seem to respond to any low-level stimulus attributes such as contrast or borders, but to high-level cognitive concepts connected to a particular stimulus. Therefore, they have been named concept cells. Concept cells specific for locations and other types of concepts have also been found (Quiroga, 2012). Additionally, concept cells are primarily found in the hippocampus, an area heavily involved in declarative memory (Tulving & Markowitsch, 1998; Gazzaniga, Ivry, & Mangun, 2008).
Concept cells were first proposed to be at the top of the hierarchy of visual processing. They were envisioned to be the result of integrating all visual and memory related information in the system, and every concept was thought to be represented by one concept cell (Bowers, 201 0). This idea is in line with localist theories (Barlow, 1972). This view of concept cells is associated with a number of problems however. Most importantly, the problem of combinatorial explosion (there wouldn’t be enough neurons to represent every concept a person knows) is seen as evidence that localist representations are too inefficient to be biologically plausible (Plaut & McClelland, 2010). Alternatively, distributed representation models do not have this problem because every neuron can be part of multiple distributed networks and can therefore represent many concepts (Gray, 1999). However, some argue that it is actually more efficient to store concept cells for many different
4 concepts than to have multiple neurons spend energy on analyzing the input every time a concept is presented (Roy, 2013).
At first glance, the existence of concept cells alone already seems to contradict distributed representation models. However, it is possible that it only seems like a neuron reacts to a stimulus on its own, while in reality the invariance of its responses reflects the invariance of its contributions to activation of a neuronal ensemble that represents the concept. Some findings on the properties of concept cells indicate that this is the case. Concept cells have often been found to respond in different degrees to concepts related to their most preferred one (e.g. a Luke Skywalker cell might also respond to Darth Vader and princess Leia, see: Quiroga, 2012). This suggests that they are part of a network of cells that can together represent multiple concepts related to one higher order concept, supporting the distributed networks interpretation.
However, it is difficult to test directly whether other neurons close to a concept cell represent other parts of an overarching concept, because concept cells can only be searched for in humans with a very limited amount of electrodes (since electrodes damage tissue). Finding concept cells is
therefore already an operation mainly based on coincidence, and finding more of the same kind is therefore unlikely. A more viable way to test whether concept cells are part of a neuronal ensemble might therefore be investigating whether they are communicating with their direct neighbors in a way that implies a neuronal ensemble is active. To be able to do this, it is necessary to define a plausible model of communication within a neuronal ensemble.
§1.2 Neuronal Coherence
Neurons communicate mainly by synaptic transmission triggered by action potentials (Gazzaniga, Ivry, & Mangun, 2008). Typically, neuronal communication is described as summation of excitatory and inhibitory post-synaptic potentials (EPSPs and IPSPs) that together cause a de- or hyperpolarization of the post-synaptic cell. If this polarization is strong enough, the cell might fire its own action potentials and transmit the information further to other neurons the cell is anatomically connected to (Gazzaniga, Ivry, & Mangun, 2008). Neurons receive thousands of different inputs from many different cells and transmit information to an equally large amount of post-synaptic neurons (White, 1989; Braitenberg and Shüz, 2013). How can neurons handle all this information in a flexible way that allows for discrimination between different situations?
Fries (2005) proposed that neurons are able to communicate in flexible ways because a flexible mechanism determines which anatomical connections between neurons are functional at the same time. This mechanism is what we call neuronal coherence and is based on the combined signal of all neurons in a group. When activity of groups of neurons is measured, voltage changes show an oscillatory pattern (Gazzaniga, Ivry, & Mangun, 2008). These oscillations influence action potential generation. When voltage is low, larger changes are necessary to reach the action potential threshold than when voltage is high. Therefore, EPSPs will have a larger effect when voltage is rising than in the troughs of the signal. Oscillations can adjust excitability of cells this way in a rhythmic manner that creates time windows during which generation of action potentials and EPSPs is likely to affect a postsynaptic neuron and time windows during which generation is unlikely to have an effect (Burchell, Faulkner & Whittington, 1998; Volgushev, Chistiakova, & Singer, 1998).
Neurons of which the membrane potential is fluctuating with the same frequency and phase should influence each other more effectively than neurons that are fluctuating out of phase (or with a different frequency) because their sensitive time windows are in sync (see Figure 1). The binding by synchronization (BBS) hypothesis is based on this idea and proposes that neurons representing the same concept or object form a neuronal ensemble and synchronize their activity when it is present or thought about (Singer& Gray, 1995). By synchronizing they can not only influence each other more easily, but also reduce the effect of activity of unrelated unsynchronized neurons. Fries (2005) extends the notion of BBS by stating that neuronal coherence is also the principal way groups of neurons communicate over large distances.
5 Figure 1. Synchronized versus unsynchronized activity
of three neurons. The BBS and CTC hypotheses both propose that neurons need to be in sync to
communicate properly, but BBS is a model of local communication while CTC is a model of distant communication. Figure adapted from Fries (2005).
Neuronal coherence as a mechanism for effective communication has been supported by research (for recent reviews: see Bosman et. al. (2014) and Fries (2015)). Modelling and animal studies have shown that neuronal ensembles can form to represent different stimuli types as a result of synchronized activity between cells and have demonstrated that synchronized activity increases output rates of post-synaptic cells (Schillen & König, 1994; Salinas & Sejnowski, 2000; Cannon et. al., 2014). Furthermore, neuronal synchronization among cells responsive to task relevant stimulus properties can predict which stimuli were presented to the animals and how they reacted. When inspecting neurons that are nearby each other, these effects are especially apparent in gamma band (30-100+ Hz) synchronization. Long distance communication is typically related to lower frequencies (Canolty & Knight, 2010).
§1.3 ING and PING Models of Gamma Band Synchronization
Though it is clear that neuronal coherence is a viable model of communication between neurons, the mechanisms of the generation of coherence are still hard to pin down. To understand coherence, it needs to be clear how oscillations are generated in the first place. Generation of gamma frequency oscillations has been studied extensively and it has become clear that inhibitory activity from interneurons is crucial for generating and maintaining gamma rhythms (Buzsáki & Wang, 2012). The Interneuron Network Gamma oscillations model (ING) is based on this research and proposes a specific mechanism of gamma oscillation generation (Whittington et. al., 2000). According to this model, inhibitory interneurons (IN) form networks in the cortex that consist of inhibitory connections between the interneurons themselves and with pyramidal neurons (PN). ING proposes that gamma synchronization of the intrinsic activity of IN is an emergent property of the network because the influence they have over each other is inhibitory. The rhythm the IN network creates is then passed on to PN, providing local PN with sensitive time windows (see Figure 2).
The Pyramidal-Interneuron Network Gamma oscillations model (PING) is very similar to ING but takes in to account that gamma rhythms are typically short-lived and flexible (Buzsáki & Wang, 2012). In addition to the mechanisms described by ING, PING proposes that the excitatory
connections from PN to IN play a role in gamma synchronization as well (Whittington et. al., 2000). Stimuli from the outside world can activate certain PN, which in turn will activate its neighboring IN with a signal of a certain frequency, phase and amplitude. This input might change the rhythm of firing of IN, causing them to start firing at a new rhythm, which will then again be imposed on the PN they are connected to (see Figure 2). This rhythm will be sustained as long as the PN are active, but when the stimulus is removed IN will return to their intrinsic rhythm. This model is especially well suited to explain temporary synchronization in neuronal ensembles representing the same stimulus.
ING and PING models have recently been investigated using animal models. Vinck et. al. (2013) demonstrated in monkeys that gamma synchronization is present in the visual cortex when no stimuli are presented, but becomes twice as strong in the presence of a stimulus. This coupling was
6 also selective: neurons with strong responses to the stimuli became entrained to the rhythm, while neurons with weak responses became less synchronized. These results are in agreement with both an ING and PING model of neuronal synchronization within neuronal ensembles.
If concept cells are part of neuronal ensembles, it would be reasonable to expect that they communicate with nearby neurons according to the ING and PING models. If concept cells would primarily be communicating with downstream neurons instead, few gamma band synchronization between the concept cell and nearby neurons would be expected. In humans, concept cells and their relationship to the ING and PING mechanisms of generation of gamma oscillations have not yet been investigated. The current study will therefore investigate whether concept cells in humans are part of neuronal ensembles in a way that is consistent with ING and PING models of neuronal communication. For these purposes, data of single unit activity from concept cells in the human hippocampus as well as LFP activity from the hippocampus will be used.
Based on the distributed representation model of neuronal representation of concepts and the ING and PING models of communication within neuronal ensembles, hypotheses on the activity of concept cells can be generated.
1) Concept cells are expected to show an increase in spike-LFP coherence in the gamma band when their concept is presented as compared to when other concepts are presented.
2) If any baseline gamma phase locking is present, the spike-LFP coherence increase is expected to be seen in a different gamma frequency than the phase locking in baseline is observed. This would suggest a switch in neuronal ensemble the cells communicate with.
§2. Methods §2. 1 Participants
Data collection for this project has also been described by Sikkens (2014) and Sikkens et. al. (2014). The data was collected from 16 patients suffering from pharmacologically intractable epilepsy by the Netherlands Institutie of Neuroscience (NIN). Patients were implanted with chronic depth electrodes for 10 – 14 days in order to locate the epileptogenic focus for surgical removal. These electrodes were positioned based solely on medical criteria.
Figure 2. Depiction of ING and PING models. At baseline, a rhythm is maintained by intrinsic activity of inhibitory interneurons (IN) (ING). When a stimulus is presented, a pyramidal neuron (PN) activates its neighboring IN, which then instigates a new rhythm and influences PN as well (PING). Figure adapted from Sikkens (2014).
7 §2.2 Recordings
Each electrode probe contained eight electrode sites used for medical diagnosis that recorded global LFP signals of large groups of cells. The sites were distributed along the shaft of the electrode in a laminar fashion. Electrodes placed in the MTL were also equipped with 9 micro-wires (eight for recording and one as reference) at the tip. These electrodes are able to record single cell activity and local LFP signals of small groups of neurons (see Figure 3). This makes them suitable for recording concept cells if they are present near the electrodes. Since concept cells are found in the MTL, the micro-wires were only added to the electrodes placed in this brain area.
For all analyses performed in this study, only data from the micro-wires was used. Signals from these electrodes were recorded with a sampling rate of 32 KHz using a 64 channel Neuralynx system. This system performs on-line filtering between 1 and 9000 Hz.
§2.3 Task
While the electrodes were present, subjects were asked to participate in sessions during which different computerized tasks originally designed by the NIN for medical examination were presented (see Figure 3). For our analyses, only screening sessions were used. During these sessions, patient viewed pictures of celebrities, famous places or personal pictures of family, friends or specific interests (e.g. cars). To ensure participants were paying attention, questions were asked about the stimuli (e.g. “Is this person Dutch or not Dutch?”) that were answered by button presses. During one recording session, a series of 36 to 51 different pictures was presented, 8 times each. Stimulus
presentation lasted 1 second, and the inter stimulus interval was 0.5 seconds. Pictures were presented in a pseudo-random order. The screening sessions were designed to find responsive units with a minimal amount of trials because these units were necessary for certain examinations the NIN needed to perform. Therefore, no changes to optimize the research design (for example by increasing the amount of trials per picture) could be made.
§2.4 Preprocessing
The continuous data from the micro-wires was split in to individual trials starting 0.5s before stimulus onset and ending 1s after. The resulting traces were filtered using a low-pass filter with a 120 Hz cut-off. Additionally, a notch filter was used to remove 50 Hz line noise from the signal. Finally, linear trends were removed from the data and the mean was subtracted.
§2.5 Data Analysis
To investigate our hypotheses, concept cells were extracted from the data. The behavior of these concept cells themselves as well as the behavior of the surrounding tissue during presentation
Figure 3. Patients implanted with depth electrodes participated in a passive viewing task. A) Schematic overview of the
electrodes with the clinical electrodes placed in laminar fashion on the probe and the micro wires at its tip. B) Cartoon of the patient participating in a passive viewing task. The different electrode types are recorded on two separate systems. This figure was originally produced by Sikkens (2014).
8 of preferred and other pictures were examined separately. Furthermore, the relationship between the behavior of the concept cells and the behavior of surrounding tissue was investigated using a novel approach to estimate spike-LFP coherence. The approaches used to extract concept cells and examine local signal are described in detail here. However, since the method for estimating spike-LFP
coherence was developed to deal with certain behaviors of the concept cells we found, this method will be explained in detail in the results section.
Extraction of concept cells.
To find single cell activity, spikes were extracted from the raw data of the micro-wires using Wave Clus (Quiroga, Nadasdy & Ben-Shaul, 2004) and sorted using an unsupervised algorithm based on principal component analysis (in-house software developed by Umberto Olcese). Clusters of spikes produced by this algorithm were inspected by looking at the average waveform and the inter-spike interval. Clusters behaving in a biologically plausible manner were saved as cells and used for other analyses, others were discarded.
All saved clusters were tested using Wilcoxon signed-rank tests comparing the firing rate during baseline (0.5-0.1s before stimulus onset) to the firing rate during the response period (0.2 -0.7s after stimulus onset), for every picture separately (8 trials) and for all pictures together (+/- 400 trials). A cluster was considered a potential concept cell when it spiked significantly more during the response period of one or several related pictures than during baseline (α=0.01, minimum of 5 spikes in the response period), but didn’t spike more during the response period than during baseline in general (α=0.001).
Examination of local LFP signals.
To examine the responses of the tissue directly surrounding the concept cells, power analyses were performed. However, because absolute power values differ drastically between patients, time points and frequencies, taking a simply average over the powerspectra calculated for every cell is not meaningful. Instead, for every cell, the average power during baseline was calculated (per frequency bin) and used to transform all power values to z-values. For all power analyses performed, z-scored powerspectra like these were used to calculate an average spectrum representing change in the signal relative to baseline rather than absolute power.
To inspect the effects of presentation of different stimuli on a larger region around the concept cells, the Weighted Phase Lag Index (WPLI) was used. The WPLI is a measure of LFP -LFP coherence designed to avoid problems such as volume conduction and noise (Vinck, Oostenveld, Van Wingerden, Battaglia & Pennartz, 2011). The WPLI reflects how much the distribution of observed phase differences of two LFP signals in a given frequency band differs from a distribution where observation of every phase difference is equally likely, giving more weight to larger differences. A WPLI of 1 indicates that the two LFP signals have a phase lag that is perfectly consistent, a WPLI of 0 indicates that no consistent phase lags are present. For every cell, the WPLI was calculated for all possible channel combinations on the probe the cell was found on, except for combinations with the channel the cell was recorded from. To create WPLI spectra, the results were averaged over all channel combinations. For all WPLI analyses performed, average spectra over cells were calculated in the same way the average z-scored powerspectra were calculated.
9 §3. Results
§3.1 Extraction of Concept Cells
Using the criteria described in the method section, a total of 13 unique concept cells was extracted from the data of five different patients (Figure 4A). A permutation test showed that, on the population level, the difference in firing rate between the response period and baseline was larger for the preferred pictures than for the other pictures (Figure 4B).
§3.2 Examination of the Local LFP Signals
To examine whether the neurons surrounding the concept cells are affected by presentation of preferred stimuli as well, a power analysis over the LFP signals from the channels the cells were recorded on was performed (see method section for details). Average powerspectra were computed twice, once for the powerspectra calculated using only the trials where the preferred picture(s) of the cells were presented and once for the powerspectra calculated over all other trials. A strong power increase in the low gamma band (+/- 30-40 Hz) and weaker increases in the higher gamma frequencies (>40 Hz) about 0.2-0.6s after stimulus presentation were observed for the preferred pictures but not for the other pictures (see Figure 5A).
To inspect the effects on a larger region around the concept cells, a WPLI analysis was performed (see method section for details). When non-preferred pictures were presented,
synchronization among LFP signals on the probe was observed for the entire response period in the low gamma band (35-60 Hz) as well as brief desynchronization in the high gamma band (+/- 85 Hz, 0-0.2s after presentation) and a late onset synchronization in theta (+/- 4-8 Hz, 0.5-0.7s after presentation). None of these effect were observed when preferred pictures were presented (see Figure 5B).
§3.3 Spike-LFP coherence
LFP analyses provide insights about how concept cells are affected by presentation of preferred stimuli. However, purely LFP-based analyses do not provide enough resolution to examine whether and how concept cells change the way they communicate with other neurons. We therefore expanded on the LFP analyses by investigating spike-LFP coherence.
Estimating spike-LFP coherence.
The Pairwise Phase Consistency (PPC, see Figure 6A) is a measure of spike-LFP coherence (Vinck et. al., 2010). The PPC quantifies whether a cell consistently spikes during the same LFP phase (of a given frequency band) by taking the average of the cosine of the angular distance between the LFP phases observed at the occurrence of two different spikes, for every pair of spikes present in the data. If spikes consistently occur during the same LFP phase, the angular distances are close to zero and the PPC takes on a value close to 1 (because the cosine of an angle of 0 is 1). If spikes consistently occur in anti-phase, the PPC is close to -1. If there is no consistency between spike times and LFP phase, the PPC is 0. The PPC is an improvement over earlier developed methods of calculating spike-LFP coherence because it’s not sensitive to differences in spike count. Comparison between PPC estimations calculated for different conditions or groups is therefore relatively straightforward. However, PPC based analyses cannot readily be used for our dataset without adaptations. Though PPC is not influenced by differences in spike count, a minimum amount of spikes is needed for accurate estimations. When comparing responses to preferred pictures to responses to other
pictures, the estimates of responses to the preferred pictures are based on only eight trials. Moreover, cell responses were extremely sparse, each trial contains little information (about 3 -4 spikes, see PSTHs in Figure 4). Simply comparing the PPC calculated over the 8 relevant trials and the PPC calculated over all other trials is therefore not possible. A seemingly obvious solution to this problem would be to calculate the PPC for the preferred pictures over the total population of relevant trials
10 A.
B.
Figure 4. Examiniation of Concept Cell Population. A) Three example cells (one cell per row) and their response to their preferred picture(s) (red) and to other pictures (blue). In the top part of the graph, the spike density function calculated over the eight trials corresponding to the picture above is plotted against time. In the lower part, the peristimulus time histogram of the same trials is shown. B) Examination of average firing rate (FR )for all 13 cells. For every cell seperately, the average FR during response is plotted against the average FR during baseline, for the preferred pictures (red) and for the other pictures (blue).
Figure 4. Investigation of Concept Cell Population. A) Three example cells and their response to their preferred picture(s) (red) and other pictures (blue). In the top part of each graph, the spike density function calculated over the eight trials corresponding to the picture displayed above is plotted against time. In the lower part, the peristimulus time histogram is shown. B) Examination of the firing rate (FR) of all 13 cells. Per cell, the average FR during the response period is plotted against the average FR during baseline, both for the preferred picture(s) (red) and for other pictures (blue). Permutation testing showed that the difference in FR is significantly higher for the preferred pictures than for the other pictures (p < .001).
11 Figure 5. Examination of Local LFP signals. A) The average z-scored powerspectra. To be able to create a meaningful average over the 13 cells, the powerspectrum of each local LFP was transformed to z -scores using the average power in the baseline. The averages of the z-scored powerspectra calculated over trials belonging to other pictures (left) and over trials belonging to preferred pictures (right) are shown here. B) The average z-scored WPLI spectra. Calculations as in (A).
A.
B.
(8 trials x 13 cells). However, it is not meaningful to calculate the PPC that way because spikes and LFP phases of one cell are not related to spikes and LFP phases of other cells (especially when recorded from different patients and at different times). Because the data is independent, one would expect PPC estimates of 0. Combining the data of all the cells is therefore not a solution.
To deal with this problem, we developed a new procedure using the PPC as its base but expanding the analysis by making use of pseudovalues (Womelsdorf, Fries, Mitra, & Desimone, 2006; Richter, Thompson, Bosman & Fries, 2015). Pseudovalues are single trial estimates of parameters that can only be calculated over multiple trials (like the PPC) and are calculated as
(1) where N is the total sample size, F is the formula for calculation of the parameter, S is the total sample and S(i) is the sample with the i-th observation left out. A pseudovalue represents the difference
between an estimation containing the information of interest (e.g. a trial of interest i) and an estimation containing everything but the information of interest (e.g. all other trials). Conceptually, such a difference expresses how much influence the information of interest has on the estimation. Positive pseudovalues represent a positive influence on the estimation, negative pseudovalues
12 A.
B.
represent a negative influence and pseudovalues of 0 represent equal influence compared to the other trials. In other words, a pseudovalue represents whether the estimate of the trial under investigation would be larger, smaller or equal to the estimate of the parameter for other trials if it were possible to estimate the parameter for that trial by itself.
Figure 6. Estimating spike-LFP coherence. A) The PPC is a measure of spike-LFP coherence. First, LFP phases are extracted at all times where spikes occur (left, simulated data). Next, the cosine of the angular distance
between the LFP phases of the spikes is calculated for every pair of spikes in the signal (two examples shown on the right). The average cosine over all pairs of spikes constitutes the PPC. For a more thorough discussion, see Vinck et. al. (2010). B) Using the PPC to estimate pseudovalues. The PPC cannot be estimated reliably based on the 8 trials corresponding to a cell’s preferred picture. To deal with this problem, the influence of the 8 relevant trials on PPC estimations calculated over a larger total amount of trials can be used instead of the PPC itself. This influence is expressed by pseudovalues. Pseudovalues are calculated by subtracting PPC estimates calculated over a subset of the total amount of trials from PPC estimates calculated over a subset of trials containing the eight relevant ones.
13 Using the PPC as our parameter, we applied the general principle behind pseudovalues to obtain spike-LFP coherence estimates (see Figure 6B). For every cell, PPC estimations containing the information of the 8 relevant trials and PPC estimations that did not contain the information of the 8 relevant trials were calculated. To avoid the problem of spectral leakage, the LFP signal of a random electrode on the same probe as the electrode that measured the cell was used for the calculation of the PPC. Since the information of interest in our situation is not individual trials but the set of trials belonging to the preferred picture of a cell, we treated every set of trials belonging to one picture as one observation instead of every individual trial. To create pseudovalues, the PPC estimations that did not contain the relevant information were subtracted from the PPC estimates that did contain the relevant information. These pseudovalues represent the influence of presentation of the preferred picture on the PPC estimations and were therefore treated as estimations of spike-LFP coherence for the preferred pictures. This procedure was repeated for the trials of a random non-preferred picture to create spike-LFP coherence estimates for non-preferred pictures.
Comparing spike-LFP coherence.
To investigate whether the concept cells communicate in a different way with the cells around them when a preferred picture is presented compared to when other pictures are presented, a way to compare pseudovalues for preferred pictures with pseudovalues for non-preferred pictures is needed. Since significance testing requires populations of observations with variance, it is not possible to do this when only one estimate (per frequency bin) per group is available. Therefore, two populations of pseudovalues were needed. These populations were created for every cell by using a subset of pictures that was changed iteratively instead of the full N-1 sample for the calculation of the PPC (see Figure 7A). For both the preferred pictures and the non-preferred pictures, 1000 different subsets of pictures were used to create two populations of 1000 pseudovalues.
A sample size of 6 other pictures (7 pictures in total, 56 trials) was chosen for this analysis instead of the largest possible sample size that allowed for 1000 unique iterations. Normally, the largest possible sample size would be preferred because estimations of parameters contain the least amount of noise when based on the largest amount of information. In this case however, we are not actually interested in the PPC estimations based on all trials but in the influence of a subset of trials on those estimations. Because we only have 8 trials of interest (2% of the total amount), and these 8 trials contain very little information, this influence is “drowned” when using the largest possible sample size. This produces two pseudovalue populations with a variance of 0, which are therefore not useful for significance testing. To find a compromise between accuracy of the PPC estimations and room for the 8 relevant trials to have influence over the estimations, we repeated the analysis for one cell with 1-9 added other pictures (see Figure 7B). Noise reduction seemed to hit a plateau at about 6 added other pictures. We therefore chose to use this sample size.
To test whether the average pseudovalues for the preferred pictures differed significantly from the average pseudovalues for the non-preferred pictures a bootstrapping procedure was used (Figure 7C and 7D). By swapping randomly chosen values from the two pseudovalue distributions 1000 times, a population of 1000 average difference scores (average pseudovalues for the shuffled
population based on the preferred picture – average pseudovalues for the shuffled population based on the non-preferred picture) could be created. For every iteration, the minimum and maximum difference scores were extracted, creating a population of minimum and a population of maximum differences. The differences between the average pseudovalues for the preferred pictures and the average pseudovalues for the non-preferred pictures were compared to these populations. If the differences were smaller than 97.5% of the minimum population or larger than 97.5% of the maximum population, they were considered to be significant differences.
14
Population preferred picture Population random picture
X 1000
Average shuffled population Average shuffled population
Frequency ->
<
-
Frequency -> Minimum ‘d pseudovalue Maximum ‘d pseudovalue = -< -I te ra tio n s A. B. C. D.15 Figure 7. Calculating Spike-LFP Coherence over a Small Amount of Trials.
A) Creating two populations of pseudovalues for significance testing. To test whether pseudovalues for preferred pictures differ from pseudovalues for non-preferred pictures, two populations of pseudovalues are needed. These were created by iteratively choosing a different subset of pictures to calculate the PPC with. By performing 1000 iterations, a pseudovalue population of 1000 values was created for the
preferred pictures (top) and the non-preferred pictures (bottom).
B) Choosing the right sample size. Normally, pseudovalues are calculated by using N-1 of the available samples to get the most accurate estimates. As the figure shows, PPC estimations contain less noise when more trials of other pictures are used for the calculation. However, adding more pictures also reduces the effect of the eight relevant trials on the estimation. As can be seen in the figure, noise reduction hits a plateau at about six added other pictures. Therefore, a sample size of six other pictures (56 trials in total) was chosen for this analysis instead of a sample size of N-1.
C) Using a bootstrapping procedure to test whether the pseudovalues calculated for the preferred pictures differ significantly from the pseudovalues calculated for the other randomly chosen pictures. By randomly swapping values from the two populations of pseudovalues 1000 times, two distributions of minimum and maximum differences can be created.
D) Test population for one example cell. Values lower than 97.5% of the minimal values and higher than 97.5% of the maximal values (red lines) are considered significant differences.
This test was performed per cell (for an example see Figure 8A), but also for the complete population of cells (Figure 8B) by treating all pseudovalues for the preferred pictures as one
population as well as all pseudovalues for the non-preferred pictures. The test revealed that concept cells desynchronize their activity with the local environment in the alpha band (8 -16 Hz) when presented with their preferred pictures compared to non-preferred pictures but show stronger synchronization in the higher frequencies (>40 Hz). To ensure that these effects were not caused by spike train history effects and dependencies between spike phase and spike rate the entire analysis was repeated using a corrected, less liberal version of the PPC. The same effects were found, but with more specificity in the higher frequencies. Increases were found specifically in two gamma bands, 40 -60 Hz and 75-80 Hz.
It is possible that the found differences between spike-LFP coherence for preferred pictures and for other pictures are already present during baseline. If that is the case, it would be doubtful that the presentation of the preferred pictures is what causes the differences. To check this, ideally the spike-LFP coherence during the baseline of the preferred pictures would be compared to the baseline of the other pictures in the same way as the response periods were compared. This is not possible because the average amount of spikes during baseline is too low to estimate PPC values reliably when taking subsets of the data. Instead, all trials were used to investigate whether there is any phase locking at all during baseline. For every cell, spikes were assigned to random trials before calculating the fourierspectrum and subsequently the PPC for 100 iterations (1000 was computionally not feasable). Using all PPC values calculated this way, a test population was created by taking the minimum and maximum values of each iteration (similar to the test population created before). The PPC values calculated over the original, unshuffled dataset were tested against the lowest and highest 2.5% of the values in the testpopulation (Figure 8C). No phase locking was found in any frequency band, suggesting that the de- and increases in spike-LFP coherence found in the response period of the preferred pictures appeared after stimulus presentation.
16
Figure 8. Spike-LFP Coherence During Presentation of Preferred Pictures vs. Spike-LFP Coherence During Presentation of Other Pictures. A) Results for one example cell. PPC estimations and pseudovalues were obtained using the methods described earlier (see Figure 7). On the left, the average PPC estimations over all iterations are plotted for the preferred (red) and non-preferred pictures (blue). In the middle, the average pseudovalues are plotted. On the right the difference between the pseudovalues for the preferred picture and the pseudovalue for the non-preferred picture is plotted, together with the results of the bootstrapping test. B) Results over the whole population of cells (N=13). A bootstrapping test was performed by treating all pseudovalues as one population. The average pseudovalues and their standard error (shaded areas) are plotted, together with the results of the test. When using the uncorrected PPC (left) a decrease in the alpha band and an increase in higher frequencies were found. When correcting for possible spike train history effects and dependencies between spike phase and rate while calculating the PPC (right), the same results were found but with more specificity. Increases were contained to two gamma bands. C). Phase locking during baseline. To examine whether the effects found in (B) were already present in baseline, a permutation test on the PPC during baseline was performed. No phase locking during baseline was found,
suggesting that the found effects were not already present during baseline.
17 §4. Conclusions and Discussion
In this study, the relationship between activity of concept cells and LFP signals has been investigated using power, WPLI and spike-LFP coherence analyses. Investigation of the local LFP signals revealed a sharp increase in power compared to baseline in the low gamma band (30 -40Hz) and weaker increases in higher gamma frequencies (> 40Hz) 0.2-0.6s after presentation of a preferred stimulus that did not occur when non-preferred stimuli were presented. The WPLI analysis on the other hand showed that LFP signals slightly further removed from the cells show an increase in synchronization in other low gamma frequencies (35-60 Hz) during the entire response period when non-preferred pictures were presented that was absent when preferred pictures were presented, as well as a short lived decrease in high gamma (0-0.2s, 85 Hz) and a late onset synchronization in theta (0.5-0.7s, 4-8 Hz). Finally, spike-LFP coherence decreased in the alpha band (8-16 Hz) and increased in the low gamma (40-60 Hz) and high gamma (75-80 Hz) band when preferred pictures were presented compared to when other pictures were presented.
Some limitations of our data need to be kept in mind while interpreting these results. Because the data for this study was recorded using a task designed for medical purposes, the data was not optimal for addressing our hypotheses. Every picture was presented too few times to allow
straightforward analyses like the PPC. Moreover, the interstimulus interval was too short to constitute a decent baseline period. A positive side effect of these problems is that it forced us to develop a method of calculating spike-LFP coherence for very small sample sizes that could make use of
combined data from independent cells. Our method, combining PPC estimates with pseudovalues and bootstrapping, might be applicable to more datasets.
Another consequence of the task design is that it is not clear whether our concept cells respond invariantly to their concept. Often, only one picture of a concept was presented in an individual session. Furthermore, no other stimuli aside from pictures (e.g. written names) were u sed. Finally, we did not investigate whether the extracted concept cells are likely to be putative pyramidal neurons or interneurons. Therefore, it is not certain whether all cells included in the population are concept cells similar to the neurons Quiroga et. al. (2005) studied or cells with a more low-level preference for certain pictures. Though this is a serious limitation, it is likely that the population consists of concept cells because the responses shown in Figure 4 are quite typical and some of th e neurons respond to several related pictures over multiple sessions. A study making use of a task designed for these experiments is needed to ensure the results are replicable.
Finally, no use was made of the electrodes recording larger LFPs. All results mentioned therefore only apply to very local signals. It would be interesting however to examine what influence concept cells have on hippocampal activity as a whole, or which distant areas they communicate with. A study combining micro-wire (LFP and single cell) data with larger and more distant LFP data should therefore be performed.
§4. 1 Increases in Gamma Band Coherence: Evidence of Active ING/PING Mechanisms?
Despite the mentioned limitations, our data and results do provide information on whether our cell population behaves like the cells are communicating with a neuronal ensemble. The spike-LFP coherence results are in line with our hypotheses. Concept cells show an increase in spike-LFP
coherence in the gamma band when their concept is presented as compared to when other concepts are presented, suggesting that the concept triggers a PING mechanism (assuming our concept cells are pyramidal neurons). The results therefore support the idea that concept cells behave like they
18 It was not expected however that a spike-LFP coherence increase would be found in multiple gamma bands, since ING/PING models propose that all cells within one neuronal ensemble use the same frequency band to communicate with one another. Two plausible explanations come to mind. First, the effect might be present in multiple gamma frequencies because concept cells are expected to belong to neuronal ensembles that use different gamma band frequencies to communicate. Possibly, part of the neurons respond in low gamma frequencies, while other neurons respond in higher gamma frequencies. When taking an average, both effects might survive.
Alternatively, the two increases might be present because they represent different types of information being communicated. Recent research has shown that low gamma and high gamma activity in the hippocampus likely have different functions (Colgin, 2016). Low gamma is related to activity in area CA3 and is thought to be involved with memory retrieval. High gamma on the other hand is related to activity in the medial entorhinal cortex (MEC) and is suspected to represent encoding of sensory information. Possibly, concept cells communicate with their ensemble in the low gamma frequencies and with the MEC in the high gamma frequencies. This would explain why the increase in the low gamma is found in a broader band than the increase in high gamma: within the low gamma band, neuronal ensembles communicate using different frequencies, but communication with the MEC occurs in only one set of frequencies.
§4. 2 Is Disengagement from Low Frequency Oscillations Inherently Part of ING/PING Mechanisms? The analyses have provided us with more information than just the predicted gamma
increase. Most interestingly, the found decrease in spike-LFP coherence in the alpha band might be fit to improve ING/PING models. Research has shown that local gamma rhythms are phase locked to lower frequency rhythms (especially theta and alpha) present in a larger area (Von Stein & Sarnthein, 2000; Osipova, Hermes & Jensen, 2008; Colgin, 2016). Possibly, PING mechanisms specifically create gamma rhythms that are not locked to these lower frequency rhythms. This would be beneficial because it would distinguish the activity of neurons in an ensemble more clearly from all other
ongoing gamma activity. Furthermore, it would allow for more neuronal ensembles to be active at the same time with minimal disturbance of the individual communication. In the original ING/PING model, the main variable distinguishing neuronal ensembles is the specific frequency band of the created gamma rhythm. When phase relationships with ongoing slow oscillations are taken in to account as well, multiple neuronal ensembles can use the same frequency bands. Some other studies recording single units in animal models have also found a simultaneous coherence decrease in low frequencies and increase in the gamma frequencies between spikes and local LFP when the neurons were stimulated (Fries, Reynolds, Rorie & Desimone, 2001; Pesaran, Pezaris, Sahani, Mitra & Andersen, 2002), or an increase in alpha coherence in neurons encoding irrelevant information (Buschman et. al., 2012). Possibly, disengagement from low frequencies is part of the PING mechanism of connecting neuronal ensembles.
There are other explanations for the decrease in spike-LFP coherence in the alpha band as well. Research has shown that feedforward and feedback signals in the visual system are
communicated using different frequency bands. Theta and gamma frequencies are involved in feedforward communication, while lower frequencies are involved in feedback communication (Van Kerkoerle et. al., 2014; Bastos et. al., 2015). Possibly, the disengagement of concept cell activity from alpha frequency activity reflects a decrease in feedback signaling to areas lower in the hierarchy of visual processing. This view is not incompatible with the idea that PING mechanisms create gamma frequency activity that is not locked to ongoing slow oscillations. However, if both are true, the
19 question of causality arises. Do PING mechanisms create gamma activity not locked to ongoing slow oscillations and is less feedback signaling a result (or maybe even byproduct) of this process, or do concept cells communicate little with areas lower in the hierarchy in general and are PING
mechanisms present but unrelated to this small amount of feedback communication?
Finally, the decrease in spike-LFP coherence might be much simpler to explain. Recently, Engel and Fries (2010) proposed that low frequency oscillations maintain the current cognitive state rather than communicating feedforward or feedback signals. A decoupling from these rhythms might simply reflect that a change is present rather than implying a certain type of communication with other cells is performed. Engel and Fries (2010) based their hypothesis mainly on studies of the motor system and the frontal cortex showing a decrease in beta band synchronization and/or power when new information was presented. Our results however show a decrease in the alpha band in the medial temporal lobe. Though the possibility that these ideas can be generalized to other brain areas and more low frequency bands must be taken seriously, evidence for applicability of this hypothesis to our data is insufficient for now. Research combining concept cell activity, activity of areas lower in the hierarchy and paradigms that require subjects to detect changes (which trigger feedback
communication, for an example see: Bastos et. al. 2015) is needed to investigate whether alpha band activity of these cells is related to feedback signaling, maintaining the current state or neither.
§4. 3 Disruption of LFP Coherence in Nearby Tissue: Activity of a Neuronal Ensemble? The results of the power and WPLI analyses add to the conclusion that concept cells
communicate in a way that is consistent with ING/PING mechanisms. The increases in gamma power observed in signals from directly surrounding tissue most likely just represent the sudden input to that area and activity of the concept cells themselves when a preferred picture is presented (since the LFP recorded with a micro-wire reflects activity of only a very small area). The responses of the signals from nearby tissue are more interesting though. The coherence of the LFP signals from tissue surrounding the area containing the concept cells shows a more or less consistent general pattern of changes when stimuli are presented. Apparently, this consistency is disrupted when a preferred stimulus causes concept cells to become active. At first, it seems like ING/PING models would predict an increase in coherence in the same gamma frequencies as the frequencies the spike-LFP coherence increase was observed in. However, concept cells (and therefore the concept cells that belong to their neuronal ensemble as well) are relatively sparsely and randomly distributed throughout hippocampal tissue (Quiroga, 2012). Possibly, some of the seven channels used to calculate the WPLI over are affected while others aren’t because some of them are near neurons that are in the same ensemble as the concept cells while others are not. Though those channels individually might show a power or a coherence increase in the gamma band, these changes would destroy coherence with other channels and are therefore not observable from the WPLI analysis. It must be noted however that it is not clear from these analyses why surrounding tissue shows a typical response to the presentation of any stimulus.
In summary, this study has found evidence suggesting that concept cells communicate with nearby tissue as though they are part of neuronal ensembles. Additionally, the data suggests the ING/PING models of communication might need to incorporate synchronization with low frequencies as well as gamma frequencies. This study therefore supports the idea that concept cells are part of a neuronal ensemble instead of being passive receivers of information that only communicate with distant brain areas. Hence, distributed representation models gain more support from this study than localist models.
20 References
Achard, S., Salvador, R., Whitcher, B., Suckling, J., & Bullmore, E. D. (2006). A resilient, low- frequency, small-world human brain functional network with highly connected association cortical hubs. The Journal of neuroscience, 26(1), 63-72.
Barlow, H. B. (1972). Single units and sensation: a neuron doctrine for perceptual psychology?. Bastos, A. M., Vezoli, J., Bosman, C. A., Schoffelen, J. M., Oostenveld, R., Dowdall, J. R., ... & Fries, P. (2015). Visual areas exert feedforward and feedback influences through distinct frequency channels. Neuron, 85(2), 390-401.
Bosman, C. A., Lansink, C. S., & Pennartz, C. (2014). Functions of gamma‐band synchronization in cognition: from single circuits to functional diversity across cortical and subcortical systems. European Journal of Neuroscience,39(11), 1982-1999.
Bowers, J. S. (2010). More on grandmother cells and the biological implausibility of PDP models of cognition: A reply to Plaut and McClelland (2010) and Quian Quiroga and Kreiman (2010).
Braitenberg, V., & Schüz, A. (2013). Cortex: statistics and geometry of neuronal connectivity. Springer Science & Business Media.
Burchell, T. R., Faulkner, H. J., & Whittington, M. A. (1998). Gamma frequency oscillations gate temporally coded afferent inputs in the rat hippocampal slice. Neuroscience
letters, 255(3), 151-154.
Buschman, T. J., Denovellis, E. L., Diogo, C., Bullock, D., & Miller, E. K. (2012). Synchronous oscillatory neural ensembles for rules in the prefrontal cortex. Neuron, 76(4), 838-846.
Buzsáki, G., & Wang, X. J. (2012). Mechanisms of gamma oscillations. Annual review of
neuroscience, 35, 203.
Cannon, J., McCarthy, M. M., Lee, S., Lee, J., Börgers, C., Whittington, M. A., & Kopell, N. (2014). Neurosystems: brain rhythms and cognitive processing.European Journal of
Neuroscience, 39(5), 705-719.
Canolty, R. T., & Knight, R. T. (2010). The functional role of cross-frequency coupling. Trends in
cognitive sciences, 14(11), 506-515.
Colgin, L. L. (2016). Rhythms of the hippocampal network. Nature Reviews Neuroscience.
Engel, A. K., & Fries, P. (2010). Beta-band oscillations—signalling the status quo?. Current opinion in
neurobiology, 20(2), 156-165.
Fodor, A. (1986). The modularity of mind. Meaning and cognitive structures: Issues in the
computational theory of mind, 3-18.
Fries, P. (2005). A mechanism for cognitive dynamics: neuronal communication through neuronal coherence. Trends in cognitive sciences, 9(10), 474-480.
Fries, P. (2015). Rhythms for cognition: communication through coherence. Neuron, 88(1), 220-235. Fries, P., Reynolds, J. H., Rorie, A. E., & Desimone, R. (2001). Modulation of oscillatory neuronal synchronization by selective visual attention.Science, 291(5508), 1560-1563.
Gazzaniga, M. S., Ivry, R. B., & Mangun, G. R. (2008). Cognitive Neuroscience: The biology of the
mind. New York: WW Norton.
Gray, C. M. (1999). The temporal correlation hypothesis of visual feature integration: still alive and well. Neuron, 24(1), 31-47.
21 Hubel, D. H., & Wiesel, T. N. (1965). Receptive fields and functional architecture in two
nonstriate visual areas (18 and 19) of the cat. Journal of neurophysiology, 28(2), 229-
289.
Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J. M. (2011). FieldTrip: open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data.
Computational intelligence and neuroscience.
Osipova, D., Hermes, D., & Jensen, O. (2008). Gamma power is phase-locked to posterior alpha activity. PloS one, 3(12), e3990.
Pesaran, B., Pezaris, J. S., Sahani, M., Mitra, P. P., & Andersen, R. A. (2002). Temporal structure in neuronal activity during working memory in macaque parietal cortex. Nature
neuroscience, 5(8), 805-811.
Plaut, D. C., & McClelland, J. L. (2010). Locating object knowledge in the brain: Comment on Bowers’s (2009) attempt to revive the grandmother cell hypothesis.
Quiroga, R. Q. (2012). Concept cells: the building blocks of declarative memory functions. Nature Reviews Neuroscience, 13(8), 587-597.
Quiroga, R. Q., Nadasdy, Z., & Ben-Shaul, Y. (2004). Unsupervised spike detection and sorting with wavelets and superparamagnetic clustering.Neural computation, 16(8), 1661-
1687.
Quiroga, R. Q., Reddy, L., Kreiman, G., Koch, C., & Fried, I. (2005). Invariant visual
representation by single neurons in the human brain. Nature, 435(7045), 1102-1107.
Richter, C. G., Thompson, W. H., Bosman, C. A., & Fries, P. (2015). A jackknife approach to quantifying single-trial correlation between covariance-based metrics undefined on a single-trial
basis. NeuroImage, 114, 57-70.
Roy, A. (2012). A theory of the brain: localist representation is used widely in the brain. Frontiers in psychology, 3, 551.
Roy, A. (2013). An extension of the localist representation theory: grandmother cells are also widely used in the brain. Frontiers in psychology, 4.
Salinas, E., & Sejnowski, T. J. (2000). Impact of correlated synaptic input on output firing rate and variability in simple neuronal models. The Journal of Neuroscience, 20(16), 6193-
6209.
Schillen, T. B., & König, P. (1994). Binding by temporal structure in multiple feature domains of an oscillatory neuronal network. Biological Cybernetics,70(5), 397-405.
Sikkens, T. (2014). Human MTL neurons selectively phase-lock with different gamma rhythm sources during a face recognition task. Unpublished master thesis.
Sikkens, T., Possel, J., Self, M., Baeyen, H., Claus, S., Roelfsema, P. R., ... & Bosman, C. A. (2014) Low- frequency phase-locking of selective human medial temporal lobe neurons to the local field potential of contralateral lateral prefrontal cortex during visual stimulation. Front.
Neuroinform. Conference Abstract: Neuroinformatics 2014. doi:
10.3389/conf.fninf.2014.18.00046
Singer, W., & Gray, C. M. (1995). Visual feature integration and the temporal correlation hypothesis. Annual review of neuroscience, 18(1), 555-586.
Tulving, E., & Markowitsch, H. J. (1998). Episodic and declarative memory: role of the hippocampus. Hippocampus, 8(3).
22 Van den Heuvel, M. P., Stam, C. J., Boersma, M., & Pol, H. H. (2008). Small-world and scale-
free organization of voxel-based resting-state functional connectivity in the human brain. Neuroimage, 43(3), 528-539.
Van Kerkoerle, T., Self, M. W., Dagnino, B., Gariel-Mathis, M. A., Poort, J., Van Der Togt, C., &
Roelfsema, P. R. (2014). Alpha and gamma oscillations characterize feedback and feedforward processing in monkey visual cortex.Proceedings of the National Academy of Sciences, 111(40), 14332-14341.
Vinck, M., Oostenveld, R., van Wingerden, M., Battaglia, F., & Pennartz, C. M. (2011). An improved index of phase-synchronization for electrophysiological data in the presence of volume- conduction, noise and sample-size bias.Neuroimage, 55(4), 1548-1565.
Vinck, M., van Wingerden, M., Womelsdorf, T., Fries, P., & Pennartz, C. M. (2010). The pairwise phase consistency: a bias-free measure of rhythmic neuronal synchronization.
Neuroimage, 51(1), 112-122.
Vinck, M., Womelsdorf, T., Buffalo, E. A., Desimone, R., & Fries, P. (2013). Attentional
modulation of cell-class-specific gamma-band synchronization in awake monkey area v4. Neuron, 80(4), 1077-1089.
Volgushev, M., Chistiakova, M., & Singer, W. (1998). Modification of discharge patterns of neocortical neurons by induced oscillations of the membrane potential. Neuroscience,
83, 15-25.
Von Stein, A., & Sarnthein, J. (2000). Different frequencies for different scales of cortical integration: from local gamma to long range alpha/theta synchronization. International Journal of
Psychophysiology, 38(3), 301-313.
White, E. L., & Keller, A. (1989). Cortical circuits. Birkhauser.
Whittington, M. A., Traub, R. D., Kopell, N., Ermentrout, B., & Buhl, E. H. (2000). Inhibition - based rhythms: experimental and mathematical observations on network
dynamics. International journal of psychophysiology, 38(3), 315-336.
Womelsdorf, T., & Fries, P. (2006). Neuronal coherence during selective attentional processing and sensory–motor integration. Journal of Physiology-Paris, 100(4), 182-193.
Womelsdorf, T., Fries, P., Mitra, P. P., & Desimone, R. (2006). Gamma-band synchronization in visual cortex predicts speed of change detection. Nature, 439(7077), 733-736.
