Research Articles, Systems/Circuits

Monkey Prefrontal Single-Unit Activity Reflecting Category-Based Logical Thinking Process and Its Neural Network Model

Takayuki Hosokawa, Muyuan Xu, Yuichi Katori, Munekazu Yamada, Kazuyuki Aihara and Ken-Ichiro Tsutsui
Journal of Neuroscience 17 August 2022, 42 (33) 6380-6391; DOI: https://doi.org/10.1523/JNEUROSCI.2286-21.2022

Abstract

Category-based thinking is a fundamental form of logical thinking. Here, we aimed to investigate its neural process at the local circuit level in the prefrontal cortex (PFC). We recorded single-unit PFC activity while male monkeys (Macaca fuscata) performed a task in which the category and rule were prerequisites of logical thinking and the outcome contingency was its consequence. Different groups of neurons coded a single type of information discretely or multiple types in a transitional form. Results of time-by-time analysis of neuronal activity suggest an information flow from category-coding and rule-coding neurons to transitional intermediate neurons, and then to contingency-coding neurons. Category-coding, rule-coding, and contingency-coding neurons showed stable coding of information, whereas intermediate neurons showed dynamic coding, as if it integrated category and rule to derive contingency. A similar process was confirmed by using a spiking neural network model that consisted of subnetworks coding category and rule on the input layer and those coding contingency on the output layer, with a subnetwork for integration in the intermediate layer. These results suggest that category-based logical thinking is realized in the PFC by separated neural populations organized for working in a feedforward manner.

SIGNIFICANCE STATEMENT To elucidate the neural process for logical thinking, we combined an in-depth analysis of single-unit activity data with a biologically plausible computational model. Results of time-by-time analysis of prefrontal neuronal activity suggest an information flow from category-coding and rule-coding neurons to transitional intermediate neurons, and then to contingency-coding neurons. Category-coding, rule-coding, and contingency-coding neurons showed stable coding, whereas intermediate neurons showed dynamic coding, as if they integrated category and rule to derive contingency. A spiking neural network model reproduced similar temporal changes of information as the recorded neuronal data. Our results suggest that the prefrontal cortex (PFC) is critically involved in category-based thought process, and this process may be produced by separated neural populations organized for working in a feedforward manner.

Introduction

Category-based thinking is a fundamental form of logical thinking, through which a subject makes an inference according to the similarities and differences of objects (Osherson et al., 1990; Bruner et al., 1966). To date, monkey single-unit studies have shown various types of category representation in the prefrontal cortex (PFC; Freedman et al., 2001; Cromer et al., 2010) and temporal cortex (Freedman et al., 2003; Kiani et al., 2007), and that PFC plays a major role in using abstract categories (Jung et al., 2018; Wutz et al., 2018) and making top-down decisions based on categories (Roy et al., 2014). We have reported that category and rule, as prerequisites of logical thinking, and contingency, as its consequence, were discretely and overlappingly represented by different subpopulations of neurons within the PFC (Tsutsui et al., 2016). Based on these findings, we hypothesized that the neural computation behind logical thinking, that is, the process of deriving a conclusion from prerequisites, takes place in local neural circuits within PFC. In this study, we aimed to test this hypothesis by further analyzing the same set of data with a computational model. In studying the neural process in the PFC local circuits, we combined an in-depth analysis of single-unit activity data with computational modeling, because the basic configuration of the local circuits in the neocortex has not yet been identified up to the level at which we can construct a structure-driven functional model, such as those in the hippocampus and cerebellum (Wolpert et al., 1998; Klausberger and Somogyi, 2008; Bezaire and Soltesz, 2013). Evaluation of the validity of the constructed computational model would be possible by comparing its output with the empirical neuronal data.

We trained Japanese monkeys in the “group-reversal task,” where the subjects were required to make an inference on the upcoming outcome according to the cue stimulus and the current rule (Fig. 1A,B). Establishment of quick adaptations to the rule changes (Fig. 1C) indicated that the monkeys depended on categories and rules, rather than individual stimulus–outcome contingencies, in performing the task. Results of behavioral probe tests further confirmed that the monkeys learned categories of the stimulus group and predicted the outcome on the basis of the current rule and the category of the presented stimulus (Hosokawa et al., 2018). Single-unit recording revealed that different subpopulations of PFC neurons code the task-relevant information (e.g., category and rule) as prerequisites of logical thinking and contingency as its consequence either discreetly or overlappingly (Tsutsui et al., 2016). In this paper, we further analyzed the temporal change in the information contents represented by these groups of neurons in a time resolution on 10-ms order. Neurons coding a single type of information showed stable neural representations. Of these neurons, the rule-coding neurons were active throughout the trial, the category-coding neurons were active during the cue presentation, and the contingency-coding neurons were active during the cue presentation and thereafter. Neurons coding multiple types of information showed dynamic neural representations; they became active at the cue presentation and showed a progressive change in the neural representation, as if they integrated the rule and category information to derive contingency information. From these findings, we constructed a neural network model that consists of subnetworks coding category and rule on the input layer and those coding contingency on the output layer, with a subnetwork for integration in the intermediate layer. The contingency information is a logical derivative of the combination of the rule and category information, and according to the logic, the model can be regarded as functionally equivalent to the XNOR gate in digital circuits (Fig. 2B). The model showed a similar behavior to PFC neurons, demonstrating that the category-based logical thinking can be supported by separated neural populations organized for working in a feedforward manner.

Figure 1.
Figure 1.

A, The monkeys predicted juice as the reward or saline as the punishment according to the cue stimulus presented on the monitor. They showed anticipatory licking when they predicted juice, while they withheld licking when they predicted saline. B, Relationship between the rules and the stimulus categories. Cue stimuli of category A indicated juice under rule X and saline under rule Y, whereas those of category B indicated saline under rule X and juice under rule Y. C, Example of a trial sequence and monkey's typical performance. The figure shows the trial number, the presented cue, monkey's predictive behavior (lick or not), the outcome (juice or saline), and the result (correct or error) in each trial in rows. The rule reversed between trial 320 and 321. At trial 321, the monkey predicted juice, but the liquid delivered was saline; therefore, the performance result was “error.” At trials 322–328, the monkey correctly predicted saline and withheld licking, although the monkey had not experienced the switched contingency for those cue stimuli after the reversal. The monkey continued to perform the task correctly thereafter.

Figure 2.
Figure 2.

A, Time course of the group reversal task. In every trial, a visual stimulus indicated whether the appetitive or aversive type of liquid (fruit juice or saline) was going to be delivered at the end of the trial. The monkeys were required to make an inference which outcome to be delivered based on the current rule and the category of the presented cue. B, A circuit drawing and a table showing the relationship among the category, rule, and contingency factors. The process of deriving the outcome from the category and rule factors can be explained as the XNOR logic.

Materials and Methods

Subjects

Two male Japanese monkeys (Macaca fuscata) were used as experimental subjects. Throughout the experiments, they were treated in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and the Tohoku University Guidelines for Animal Care and Use. This project was approved by the Center for Laboratory Animal Research of Tohoku University. Monkeys were housed individually in a cage in a room with natural lighting.

Apparatus

In the laboratory, a monkey sat on a home-made primate chair to which a touch-key sensor (Supertech) was attached. Visual stimuli were presented at the eye level on a liquid-crystal display (LCD; Prolite E431S, Iiyama) placed 35 cm in front of the monkey, and two types of liquid were delivered as reward and punishment through a double-spout device placed in front of the monkey's mouth (Fig. 1A). An infrared sensor (Supertech) was attached to the spout device to monitor the monkey's spout-licking behavior. All of the abstract figures used as visual stimuli had been generated as computer graphics files (bitmap format, 142 × 142 pixels) and stored on a hard disk. When a visual stimulus was presented on the LCD, its size was extended to 6° × 6° in visual angle. Orange juice (Toris Conc, Suntory) was used as a reward and concentrated saline (7%) was used as a punishment. Visual stimulus presentation was controlled by graphics presentation software (Presentation, Neurobehavioral Systems) on a personal computer (xSeries100, IBM Japan) that was synchronized with and controlled by a home-made host computer. Juice/saline delivery was controlled by the opening of solenoid valves (CKD Corporation) remotely controlled by the host computer.

Behavioral task

Monkeys performed a “group reversal” task, in which they needed to make an inference on the contingency (juice reward or saline punishment) based on the category group of the presented cue stimulus and the current task rule (Fig. 1A,B). The stimulus–outcome associations with eight distinct stimuli and two distinct outcomes (juice and saline) were kept unchanged within a session and reversed between sessions. In every trial, a visual stimulus was presented to the subject, and then its associated outcome (juice or saline) was delivered through the spout in front of the mouth. Therefore, the stimulus served as a cue to predict the outcome of the trial. The correct response at the time of liquid delivery was to lick if the preceding stimulus had been associated with juice (reward acquisition by “go” response) and not to lick if it had been associated with saline (punishment avoidance by “no-go” response).

The precise time sequence of the task events was as follows (Fig. 2A). When a red fixation spot (4 mm in diameter) appeared at the center of the LCD, the monkey touched the key and fixated on the spot. After a precue period (either variable between 1.0 and 1.5 s or fixed to 1.25 s), a visual stimulus was presented at the center of the LCD for 0.75 s. After cue extinction, a delay period lasted for either a variable duration between 0.75 and 1.5 s or a fixed duration of 1.25 s. When the fixation spot turned from red to green, the monkey released the key, and the fixation spot disappeared. After the key release, another delay period lasted for 0.5 s (or 0 s in some training sessions), then either juice or saline was delivered. For the precue and first delay period, variable timing was used during the training sessions and initial three months of the recording sessions. Later, fixed timing was used. We counted a trial as correct if a monkey licked the spout within a certain time window including the juice delivery time (from 200 ms before juice onset to 500 ms after juice onset) in juice trials and did not lick the spout within the time window in saline trials. We repeated the same trial (correction trial) if a monkey terminated a trial erroneously by fixation break or early key release so that the monkeys learned to complete a trial even if the predicted outcome was saline.

In each trial, a stimulus was chosen pseudorandomly in such a way that each of the eight stimuli in a set was used once in a block of eight trials. The rule of stimulus–outcome associations, four out of eight stimuli being associated with juice and the rest being associated with saline, was kept unchanged within a session, and between sessions they were reversed without any explicit cue. One block consisted of eight trials, and one session consisted of 6–12 blocks, namely, 48–96 trials. The first trial after the rule reversal was always a saline trial so that a monkey would notice the reversal. Therefore, the punishment, an unexpected delivery of saline in the first trial after the rule reversal, served as a cue for the rule reversal. Since the outcome associated with each stimulus switched according to the current rule, and the monkeys were required to infer the outcome based on the current task rule and the category group of the presented stimulus. This inference process was logical because the outcome (juice or saline) was determined by the combination of the presented cue and the current rule (Fig. 2B).

Eye movements were monitored using an infrared eye movement recording system (ETL-200, I-scan). The trial was canceled immediately if the eye position exceeded the limit of 1° from the fixation spot. The eye movements during each trial were also examined offline to confirm eye fixation.

Single-unit recording

Before single-unit recording, a stereotaxic magnetic resonance imaging (MRI) scan of the brain was taken for each monkey. Then a head-fixation device and a recording chamber were implanted in a standard surgical procedure using pentobarbital (Nembutal, Dainippon Sumitomo Pharma) for general anesthesia. The recording chamber was cylindrical in shape with an inner diameter of 18 mm. The skull over the lateral PFC (LPFC) was removed giving a skull opening size 18 mm in diameter, and the cylindrical recording chamber was implanted at 45 degrees of angle over the opening of the skull. After the monkey had recovered from surgery, extracellular single-unit recording was performed in the PFC during the performance of the task by using an Epoxylite-coated or glass-coated tungsten microelectrode (impedance: 1.5 MΩ at 1 kHz; FHC) attached to a hydraulic x–y stage microdrive (MO-97S, Narishige) to penetrate and advance into the brain. Electrophysiological signals were amplified (10,000 times), bandpass filtered (low cut: 100 Hz; high cut: 10,000 Hz) with a standard biophysical amplifier (BioAmp A2-v6, Supertech), and displayed on an oscilloscope (CS-4125A, Kenwood). The amplified electrophysiological signals were also audibilized and presented to the experimenter through a speakerphone. The action potentials of isolated neurons were sorted by a window-discriminator (DDIS-1, Bak Electronics) and displayed on a digital storage oscilloscope (DCS-7040, Kenwood). The recorded electrophysiological signals were digitized at 25 kHz by an analog-to-digital conversion interface (Power 1401, CED) and then stored on the hard disk of the personal computer. Moreover, we stored the time of the detected action potentials together with those of key touch/release, licking responses, eye movements, and task events. Rastergrams and histograms showing the neuronal activity in each cue condition were displayed online on an LCD video screen. Before single-unit recording began, the monkeys had been trained with three different regular stimulus sets. In the daily recording routine, we randomly selected one of the three stimulus sets at the beginning. We searched for a neuron while a monkey performed the task with the stimulus set. After isolating a neuron, we recorded the neuronal activity with the same stimulus set, during which the rule was normally reversed from three to five times. After finishing the recording of neuronal activity, we usually changed the stimulus set to another one by random selection, then began to search for another neuron. The authors were not blinded to group allocation during the experiment and in assessing the data.

Offline analysis of recorded neuronal activity

The recorded neuronal data were analyzed offline using a data analysis program with MATLAB (MathWorks). Details of the specific analyses are provided below. Randomization tests for the coefficient of partial determination (CPD) were one-tailed, while the other statistical tests were two-tailed.

We analyzed neuronal activity by using the following multiple regression equation with three variables (Draper and Smith, 1998): y=b0 + b1x1 + b2x2 + b3x1x2 + b4x3,(1) where y denotes the spike rate during the cue period, b0 the offset, b1, b2, b3, and b4 the coefficients for the explanatory variables x1,x2, and x3. x1: rule (1: A+/B−, −1: A−/B+), x2: category group (1: stimulus in group A, −1: stimulus in group B), x1x2: interaction term between rule and category factors, which also corresponds to contingency (1: juice, −1: saline), and x3: cumulative trial number (to account for a gradual activity change over trials). According to this regression model, we analyzed how PFC neurons code the task-relevant information in the group reversal task: category, rule and contingency information. We classified neurons into four types: rule-coding, category-coding, contingency-coding, and intermediate-type neurons. The category-coding type showed significance in the regression coefficient of the category factor only, the rule-coding type showed significance in the regression coefficient of the rule factor only, the contingency-coding type showed significance in the regression coefficient of the contingency factor (interaction term) only, and the intermediate-type showed significance in the regression coefficients of multiple factors. Significance was assessed at p < 0.05 (two-tailed). To minimize the influence of neuronal activity change because of the reversal, we excluded from the analyses one block (eight trials) of data just after the reversal.

We calculated the CPD for the category factor in the regression model and compared the CPD calculated from the original data with that calculated from the randomized data. The CPD is defined as follows: CPD=(SSEr–SSEf)/SSEr,(2) where SSEr is the sum of squared errors in the regression model that excludes one of the factors and SSEf is the sum of squared errors in the regression model with all factors. A CPD ranges from 0 to 1. We plotted CPD as percentage (i.e., multiplying the original values by 100) on a logarithmic scale.

To determine whether CPD at a given time point is higher than the chance level at the population level, we conducted a randomization test (n = 1000, p < 0.05, one-tailed). First, we shuffled the spike data (y in the Eq. 1) between trials while keeping the explanatory variables (x1, x2, and x3 in the Eq. 1) the same as the original data, and calculated CPD using the shuffled spike data for each neuron. Then, we calculated the mean CPD in each time bin for each type of neuron. We repeated this procedure 1000 times and constructed the distribution of mean CPDs from the randomized data in each time bin. We then determined whether CPD was significant depending on whether the mean CPD from the observed data exceeded 95% of CPDs from the randomized data.

Spiking neural network model

We constructed a spiking neural network model consisting of four subnetworks. Each subnetwork was composed of recurrently connected excitatory and inhibitory neurons. The task-relevant information (i.e., the category, rule and contingency information) was assumed to be encoded by selective neurons in the excitatory populations of the corresponding subnetwork. Specifically, each of the category-coding, rule-coding, and contingency-coding subnetworks contained two selective populations, whereas the intermediate subnetwork contained four selective populations. For the sake of simplicity, we assumed that the selective populations were nonoverlapping. In each subnetwork, nonselective neurons and inhibitory neurons were randomly connected to the selective neurons. The synaptic connections in the nonselective population were assumed to have a baseline value (which may change across different subnetworks). Compared with this, the synaptic connections within each selective population were assumed to have potentiated values, while the synaptic connections between different selective populations, or between the selective and nonselective population were assumed to have depressed values (except for the intermediate subnetwork as explained below). This structure could result from long-term potentiation (Bliss and Collingridge, 1993) and long-term depression (Artola and Singer, 1993) under correlated and anti-correlated neural activity, respectively. Additionally, in the intermediate subnetwork, we assumed that the synaptic connections between selective populations corresponding to the same contingencies had potentiated values. Furthermore, the selective populations were organized through feedforward connections. Such connections may be established naturally when neurons in the category-coding and rule-coding subnetworks competed to innervate neurons in the intermediate subnetwork (Song and Abbott, 2001).

In order to compare the model with recorded data, we simulated the dynamics of the network using noisy leaky integrate-and-fire neurons. The membrane potential of each neuron evolved according to the following equations: CmdVk,idt=−Ik,iL−Ik,iE−Ik,iI + Ik,ibias + σ⋅ξk,i,(3) Ik,iL=gL⋅(Vk,i−EL),(4) Ik,iE=gk,iE⋅(Vk,i−EE),(5) Ik,iI=gk,iI⋅(Vk,i−EI),(6) where k denoted the population and i the neuron. In addition, Cm denoted the membrane capacitance, and Ik,iL, Ik,iE and Ik,iI denoted the leak current and synaptic currents from excitatory and inhibitory neurons, respectively. gL, gk,iE, and gk,iI denoted the conductance that induced these currents, and EL, EE, and EI denoted the corresponding reversal potentials (Wang, 1999). Furthermore, the time-independent bias current Ik,ibias and the Gaussian white noise term ξk,i were added to each neuron. When the membrane potential Vk,i hit the threshold value Vth, the neuron emitted a spike. Vk,i was then reset to the resting potential Vr and remained insensitive to synaptic currents during a refractory period τr.

The synaptic activity triggered by the presynaptic neuron was described by the variable sk,i as follows: dsk,idt=−sk,iτs + (1−sk,i)∑mδ(t−tk,im),(7) where δ was the Dirac δ function, and tk,im denoted the m-th spike time of the ith presynaptic neuron in population k. The summation ran over all spikes arriving at the synapse. When a spike arrived at the presynaptic axon terminal, synaptic activity was raised to unity instantaneously, and then decayed toward zero between successive spikes with a time constant τs.

The conductance induced on the postsynaptic membrane was assumed to be proportional to the strength of synaptic connection and the synaptic activity: gk,in=∑lpopulationgkl∑jneuronssl,j(t−di,jk,l) + gkext(8) where n=E (excitatory), I (inhibitory) denoted the type of the induced conductance. The first summation ran over presynaptic populations, and the second ran over individual neurons. For simplicity, the synaptic connections within/between populations were assumed to have uniform values. gkl denoted the strength of synaptic connections from population l to population k. Since neurons in each population received distinct numbers of synapses, the strength of synaptic connections was normalized by the total number of afferent synapses to the neurons. Specifically, gkl was given as follows: for excitatory synapses, gkl=Gkl/NkE, whereas for inhibitory synapses, gkl=Gkl/NkI, where NkE and NkI were the numbers of excitatory and inhibitory synapses received by each neuron in population k, respectively. Gkl denoted the total synaptic strength from population l to a neuron in population k. Additionally, di,jk,l denoted the transmission delay. Furthermore, an external input was applied to each neuron. The external inputs were also assumed to have uniform values across neurons in each population and were composed of two parts: the constant background input gkb and the time-dependent activation input gka described as follows: gkext=gkb + gka.(9) The activation input was set to zero unless otherwise mentioned.

Since ∼60% recorded neurons showed significant response to at least one type of task-relevant information, we assumed that the proportion of selective neurons in the excitatory population was 60% in each subnetwork. On the other hand, the proportion of inhibitory neurons in each subnetwork was assumed to be 20% (Abeles, 1991). Each of the category-coding, rule-coding, and contingency-coding subnetworks had two selective populations and the intermediate subnetwork had four selective populations. In the category-coding and rule-coding subnetworks, each selective population was composed of 300 neurons, the nonselective population was composed of 400 neurons, and the inhibitory population was composed of 250 neurons. In the contingency-coding subnetwork, the size of each population was doubled since the number of contingency-coding neurons was about twice as much as the numbers of category-coding or rule-coding neurons. Therefore, each selective population was composed of 600 neurons, the nonselective population was composed of 800 neurons, and the inhibitory population was composed of 500 neurons. Finally, in the intermediate subnetwork, each selective population was composed of 300 neurons, the nonselective population was composed of 800 neurons, and the inhibitory population was composed of 500 neurons. In total, 7500 neurons were used in simulations.

Populations in the category-coding, rule-coding, contingency-coding and intermediate subnetworks were denoted by C, R, O, and I, respectively, as shown in Figure 6A. The selective, nonselective and inhibitory populations were indicated by numbers, and subscripts ns and inh. Specifically, we assumed that the selective populations C1 and C2 corresponded to category groups A and B, R1 and R2 corresponded to rules A+/B− and A−/B+, and O1 and O2 corresponded to contingency of juice and saline. The selective populations I1, I2, I3 and I4 corresponded to the combination of category group A and rule A+/B−, of category group B and rule A+/B−, of category group A and rule A−/B+, and of category group B and rule A−/B+, respectively.

The parameter values were given as follows. When the two populations were connected, each pair of neurons had a probability p=0.2 to establish a synaptic connection. For the category-coding subnetwork, GC1C1=GC2C2=2.0 nS, GCnsCns=1.5 nS, and GC1C2=GC2C1=GC1Cns=GC2Cns=GCnsC1=GCnsC2=1.0 nS; for the rule-coding subnetwork, GR1R1=GR2R2=3.4 nS, GRnsRns=1.5 nS, and GR1R2=GR2R1=GR1Rns=GR2Rns=GRnsR1=GRnsR2=1.0 nS; for the contingency-coding subnetwork, GO1O1=GO2O2=3.5 nS, GOnsOns=1.5 nS, and GO1O2=GO2O1=GO1Ons=GO2Ons=GOnsO1=GOnsO2=1.0 nS; and for the intermediate subnetwork, GI1I1=GI2I2=GI3I3=GI4I4=5.0 nS, GI1I4=GI2I3=GI3I2=GI4I1=9.0 nS, GI1I2=GI2I1=GI3I4=GI4I3=GI1I3=GI3I1=GI2I4=GI4I2=GI1Ins=GI2Ins=GI3Ins=GI4Ins=GInsI1=GInsI2=GInsI3=GInsI4=2.4 nS, and GInsIns=2.5 nS. The total synaptic strength involving inhibitory populations was assumed to be 5.0 nS. Finally, the total synaptic strength for the feedforward connections from the category-coding and rule-coding subnetworks to the intermediate subnetwork was assumed to be 5.0 nS, and the total synaptic strength for the feedforward connections from the intermediate subnetwork to the contingency-coding subnetwork was assumed to be 2.5 nS.

The dynamics of the network was simulated by the Euler-Maruyama method, with a time step Δt=0.5 ms and following parameter values (Wang, 1999): For both excitatory and inhibitory neurons, Vth=−52 mV, Vr=−60 mV, EE=−5 mV and EI=−75 mV; for excitatory neurons, τr=2 ms, Cm=0.5 nF, gL=25 nS, and EL=−70 mV; and for inhibitory neurons, τr=1 ms, Cm=0.2 nF, gL=20 nS, and EL=−65 mV. All the excitatory synapses were assumed to be of the NMDA type, with time constant τE=100 ms. The voltage-dependence of NMDA synapses was not modeled in this study. The time constant of inhibitory synapses was assumed to be τI=20 ms. In addition, the time-independent bias current for each neuron was determined randomly according to a Gaussian distribution with zero mean and standard deviation 0.02⋅σ⋅Δt, where σ=0.1. At each time step, a Gaussian noise with zero mean and standard deviation σ⋅Δt was added independently to each neuron. Furthermore, the transmission delay di,jk,l for each synapse was chosen uniformly in the range from 1 to 5 ms.

The magnitude of the background input was set to allow the neurons to fire in a biologically plausible range. For neurons in the category-coding, rule-coding and contingency-coding subnetworks, gC1b=gC2b=gCnsb=gR1b=gR2b=gRnsb=gO1b=gO2b=gOnsb=9.1 nS and gCinhb=gRinhb=gOinhb=4.3 nS. For neurons in the intermediate subnetwork, the background input was set in a way that without recurrent connections, only neurons in the population receiving largest feedforward input were active: gI1b=gI2b=gI3b=gI4b=gInsb=9.0 nS and gIinhb=5.4 nS. On each trial, information about the category and rule was provided to the network via activation inputs. At the beginning of the trial, the rule was loaded into the network by applying an activation input with magnitude 0.5 nS and duration 100 ms to the corresponding selective population in the rule-coding subnetwork. The simulation was then run freely for 400 ms before we began to record the states of neurons at time zero. Between 1000 and 1750 ms, which correspond to the cue period, the category information was presented by applying an activation input to the corresponding selective population in the category-coding population. The magnitude of the activation input increased linearly from 0 to 0.5 nS over 200 ms and was maintained at 0.5 nS for 150 ms before decaying back to 0 nS over another 400 ms. On the other hand, an activation input with magnitude 0.1 nS was applied to the inhibitory neurons in the intermediate subnetwork between 1000 and 1200 ms to force the competition between the selective populations, and was applied to all excitatory neurons in the rule-coding subnetwork between 600 and 1600 ms, corresponding to the slight enhancement of the rule information observed around the cue onset. We note that these activation inputs are not essential for our analyses. In total, each trial was simulated for 3000 ms, with the states of neurons recorded during the last 2500 ms. CPDs were then calculated using the simulated data from 100 trials.

Results

In the group-reversal task, the monkeys were required to make an inference of the outcome (juice or saline) based on the category of the presented cue stimulus and the current task rule. Thus, we first investigated how PFC neurons coded those task-relevant information [category, rule, and contingency (the outcome the presented cue predicted)]. We analyzed the neuronal data by a multiple regression model with explanatory factors of the task rule, the category group, and the contingency (see Materials and Methods). We recorded 225 neurons from the lateral and orbital parts of the PFC, and classified those neurons into four types based on the results of the multiple regression analysis: the category-coding type that showed significance in the category factor only, the rule-coding type that showed significance in the rule factor only, the contingency-coding type that showed significance in the contingency factor (interaction term) only, and the intermediate-type that showed significance in multiple factors. Figure 3A shows the proportion of each type of neuron. Out of 225 neurons, we found 19 (8.4%) category-coding neurons, 24 (10.7%) rule-coding neurons, 46 (20.4%) contingency-coding neurons, and 47 (20.9%) intermediate-type neurons. Figure 3B shows the locations of each type of neuron. Category-coding neurons were mainly found in the ventral part of the LPFC. Rule-coding neurons were mainly found around the principal sulcus. Contingency-coding and intermediate-type neurons were found in wide areas over the PFC.

Figure 3.
Figure 3.

Proportions and locations of neurons. A, Proportion of each type of neuron. Neurons were classified into four types according to the significance of CPDs for the category, rule, and contingency factors during the entire cue period. B, Recording locations for each type of neuron. The upper row shows the lateral view of the PFC, and the lower row shows the bottom view of it. AS, arcuate sulcus; PS, principal sulcus; LOB, lateral orbital sulcus; IOB, intermediate orbital sulcus; MOB, medial orbital sulcus.

To illustrate how much information each type of neuron coded during the cue period, we calculated the CPD, which indicates the amount of variance in the neuronal activity that is accounted for by each factor (category, rule, and contingency; see Materials and Methods). We plotted CPDs of category, rule, and contingency information coded by each neuron during the cue period in a three-dimensional space (Fig. 4A). As expected, category-coding, rule-coding, and contingency-coding neurons, respectively, scattered along each corresponding axis, while intermediate-type neurons scattered in the middle part of the three-dimensional space. Then we calculated the temporal change of CPD for category, rule, and contingency information for each type of neuron from 1000 ms before the cue onset to 1000 ms after the cue offset (Fig. 4B). In category-coding neurons, category information became active immediately after the cue onset and inactive shortly after the cue offset. In rule-coding neurons, rule information was active even before the cue onset, enhanced around the cue onset, and kept active throughout the trial. In contingency-coding neurons, contingency information became active shortly after the cue onset, and was kept active until the response (key release) or the juice/saline delivery. Intermediate-type neurons coded all types of information (i.e., category, rule, and contingency), and the time courses of category, rule, and contingency information were similar to those of category-coding, rule-coding, and contingency-coding neurons, respectively.

Figure 4.
Figure 4.

Information coding of each type of PFC neurons. A, CPDs of each type of neuron for the category, rule, and contingency factors calculated from the average firing rate during the entire cue period. Each dot represents a CPD from a single neuron. Different colors are used to represent different types of neurons: red, category-coding neurons; blue, rule-coding neurons; green, contingency-coding neurons; orange, intermediate-type neurons. The CPD (%) is shown on a logarithmic scale. B, Time course of CPDs of each type of neuron for the category, rule, and contingency factors. CPD of each factor was calculated using a 200-ms sliding window at 20-ms steps from 1000 ms before the cue onset to 1000 ms after the cue offset. Averaged CPDs are shown on a logarithmic scale. The thick parts of lines indicate that the value of CPD in the bin is significantly higher than that calculated from randomized data (the randomization test, p < 0.05; see Materials and Methods). C, Cumulative curves of the latency of the information coding for each type of neuron with reference to the cue onset. The latency for each neuron was defined as the time of the first of three consecutive significant bins. The solid red line represents category-coding neurons, the solid blue line rule-coding neurons, and the solid green line contingency-coding neurons. The dashed red line represents the category information coded by intermediate-type neurons, the dashed blue line the rule information coded by intermediate-type neurons, and the dashed green line the contingency information coded by intermediate-type neurons.

To infer the information flow between types of neuron, we analyzed the latency of information coding with reference to the cue onset for each type of neuron (Fig. 4C). More than 50% of the rule-coding neuron (the solid blue line) were active by the time of cue onset. Similarly, intermediate-type neurons showed increasing activation of rule-information coding (the dashed blue line), and ∼50% of them showed significant coding of rule information by the time of cue onset. The latency of category-coding neurons (the solid red line) preceded that of contingency-coding neurons (the solid green line). The latencies of category information (the dashed red line) and contingency information (the dashed green line) coded by intermediate-type neurons were between those of category-coding and contingency-coding neurons.

To quantify the stability/variability of the information represented by each type of neuron, we calculated the temporal change of the mean centers of CPD for category, rule, and contingency information for each type of neuron, and visualized them in the three-dimensional space (Fig. 5A; Movie 1). The mean centers of category-coding and rule-coding neurons were relatively stable during the cue period. On the other hand, the mean center of intermediate-type neurons moved a long distance in the CPD space, starting around that of rule-coding neurons, moving toward that of category-coding neurons for 60 ms and then moving upwards along the z-axis toward that of contingency-coding neurons. Although the mean center of contingency-coding neurons also moved upwards along the z-axis, the moving distance was much less than that of intermediate-type neurons. The averaged distance that the mean center of each intermediate-type neuron moved during the cue period was significantly longer than those of the other type of neuron, suggesting that the information coding by intermediate-type neurons more dynamically changed during the cue period compared with the other types of neurons (Fig. 5B, the Bonferroni-corrected t test, p < 0.05). This result indicates that the intermediate-type neurons may be involved in the dynamic process of information integration.

Figure 5.
Figure 5.

Evolution of information coding during the cue period in each type of neuron. A, Temporal change of the mean centers of CPD of each type of neuron. Three-dimensional (3D) view (left), top view (right). We calculated CPD of each factor for all neurons using a 200-ms window sliding 20-ms steps from the time of cue on to the time of the cue off, and determined the mean center of CPDs for each type of neuron in each time bin by averaging the CPDs of “active” neurons. (“Active” means that the neuron showed a significant CPD for either the category, rule, or contingency factor at the time.) The arrow in each panel indicates the direction of the movement of the mean center of intermediate-type neurons. See also Movie 1. B, Travel distances of CPDs of each type of neuron during the “active” state in the 3D space. *Bonferroni-corrected t test (p < 0.05, two-tailed). Note that the length of the trajectories in A does not necessarily match the distance traveled by each type of neuron in B because of the following reasons: we show the trajectories of the averaged CPDs (i.e., mean center) of all the “active” neurons at each time in A. On the other hand, we show the cumulative sum of travel distances of each neuron in the 3D space while the neuron was “active” in B. Since overlapping but changing populations of neurons became “active” during the cue period, the mean center of CPDs of all the “active” neurons at each time moves much more (as in A) than the point of CPDs of each neuron (as in B).

Movie 1.

Movement of the mean center of CPDs for each type of neuron. We calculated CPD of each factor for each neuron using a 200-ms window sliding 20-ms steps during the time period between 300 ms before the time of the cue on and 50 ms after the time of the cue off. To determine the mean center of each type of neuron, we used the data from the neurons that showed significant CPDs of the corresponding factor (i.e., the category factor for category-coding neurons, the rule factor for rule-coding neurons, the contingency factor for contingency-coding neurons, and any factor for intermediate-type neurons) at each time bin. The size of each dot reflected the number of neurons: a bigger dot indicates that a larger number of neurons were used to determine the mean center.

Together, these results suggest that the category-coding and rule-coding neurons actively maintained the category and rule information, respectively, whereas the intermediate-type neurons used the actively maintained information, performed integration, and passed the result to the contingency-coding neurons. In particular, the computation was accompanied by representational switching (Rao et al., 1997; Rainer et al., 1999; Romo et al., 1999; Saito et al., 2005; Mushiake et al., 2006; Sakamoto et al., 2008; Genovesio et al., 2009; Bartolo and Averbeck, 2020) from the rule to the contingency in the intermediate-type neurons. Previous studies have suggested that representational switching may occur owing to stochastic fluctuation (Mongillo et al., 2003) or short-term synaptic plasticity (Katori et al., 2011). However, representational switching induced by stochastic fluctuation, when averaged over trials, typically occurs over a period of several hundred milliseconds. On the other hand, representational switching induced by short-term synaptic plasticity depends on a specific distribution of synaptic facilitation or depression, whose mechanism remains unclear.

To test whether the computation and representational switching can be implemented by a conventional spiking neural network in a sensory input-driven manner, we constructed a neural network model with hierarchically organized subnetworks (Fig. 6A). Each subnetwork corresponded to a specific type of neuron and was further divided into recurrently connected selective populations, a nonselective population, and an inhibitory population. Each selective population encoded a specific piece or a specific combination of pieces of information. For example, the selective population in the rule-coding subnetwork encoded a specific rule and the selective population in the intermediate subnetwork encoded a specific combination of rules and categories. The selective populations in the intermediate subnetwork enumerated all possible combinations of rules and categories. Each of these selective populations received inputs from the corresponding selective populations in the rule-coding and category-coding subnetworks and in turn sent inputs to the selective population in the contingency subnetwork. Such a feedforward structure between the selective populations could result from the competition among neurons to innervate their downstream counterparts, for example, through spike timing-dependent plasticity (Markram et al., 1997). Finally, we assumed that the connections between selective populations corresponding to the same contingencies are potentiated in the intermediate subnetwork (Fig. 6B). The dynamics of the network was simulated using noisy integrate-and-fire neurons. The rule was first loaded into the network by applying a brief activation input to the corresponding population in the rule-coding subnetwork. Then, during the cue period, a cue stimulus was represented by another activation input to the corresponding population in the category-coding subnetwork.

Figure 6.
Figure 6.

Architecture of the neural network model. A, Feedforward structure of the neural network model. The network consists of four subnetworks, each indicated by a different color. The subnetworks are composed of recurrently connected excitatory (including selective and nonselective) and inhibitory neurons. Circles indicate selective neural populations. For simplicity, nonselective and inhibitory neural populations are not shown. B, Recurrent structure of the intermediate subnetwork. The thickness of the arrow indicates the strength of the corresponding synaptic connections.

To compare the model with data recorded during the group reversal task, we calculated CPDs for the simulated data at each time point with instantaneous firing rates estimated by a 20-ms sliding window. The CPDs were then averaged across the cue period (Fig. 7A) and each type of neuron (Fig. 7B,C) and compared with those for the recorded data (Figs. 4A,B, 5A). In general, the model reproduced the distribution and temporal evolution of CPDs (Fig. 7). The category information was activated transiently during the cue period. By contrast, the rule information was maintained throughout the trial because of the reverberation of activity through the recurrent loops among interconnected neurons. Such self-sustained activity manifests when the recurrent connections are sufficiently strong (Amit and Brunel, 1997). Similarly, the contingency information was activated shortly after the stimulus onset and maintained thereafter.

Figure 7.
Figure 7.

CPDs calculated for the simulated data. To calculate the CPDs, 10 neurons from each selective population in the category-coding, rule-coding, and intermediate subnetworks, and 20 neurons from each selective population in the contingency-coding subnetwork are randomly selected. Information coding in the neural network model is illustrated by using the simulated activity of 120 neurons. A, CPDs of each neuron averaged across the cue period (the same format as in Fig. 4A). Each of the category-coding, rule-coding and contingency-coding neurons formed a cluster, and the distribution of the intermediate-type neurons located in the middle. B, Temporal evolution of CPDs in each type of neuron (the same format as in Fig. 4B). C, Temporal change of the mean center of CPDs for each type of neuron during the cue period (the same format as in Fig. 5A).

The intermediate subnetwork demonstrated a rapid representational switching from the rule to the contingency at the stimulus onset. This representational switching depends on the strong recurrent connections between selective populations corresponding to the same contingencies (Fig. 8). Although each group of these selective populations received the same total amount of activation from the rule-coding and category-coding subnetworks, the largest activation was found at the selective population corresponding to the rule and category on a particular trial. The other selective populations were strongly inhibited through mutual inhibition. Therefore, nonlinear response properties of the neurons and strong recurrent connections allowed the activity in the “winning” population group to build up, displaying a representational switching.

Figure 8.
Figure 8.

Comparison of temporal evolution of CPDs. The two columns show the temporal evolution of CPDs in the contingency-coding (left) and intermediate-type (right) neurons, respectively. The strength of recurrent connections between intermediate-type neurons corresponding to the same contingencies increases from top to bottom (top: GI1I4=GI2I3=GI3I2=GI4I1=8.0 nS, middle: GI1I4=GI2I3=GI3I2=GI4I1=8.5 nS, and bottom GI1I4=GI2I3=GI3I2=GI4I1=9.0 nS). Left column, In either case, the contingency information became active in the contingency–coding neurons shortly after the onset of a cue, and was maintained toward the end of the trial. Right column, By contrast, intermediate-type neurons displayed the representational switching from the rule factor to the contingency factor (Fig. 4B, last panel) only when the recurrent connections are sufficiently strong.

Discussion

To investigate neural processes within PFC underlying category-based logical thinking, we combined the empirical study of recording and analyzing the prefrontal neuronal data with computational modeling. The analyses of the empirical neuronal data indicated that different groups or subnetworks of PFC neurons discretely maintained rule and category information, as prerequisites of logical thinking, and contingency information, as its consequence, and another group/subnetwork of PFC neurons integrated the category information with the rule information to generate the contingency information (a prediction about the outcome). A conventional spiking neural network qualitatively reproduced these observations and revealed a rapid, sensory input-driven representational switching caused by nonlinear recurrent interactions between intermediate-type neurons.

We found that different types of neurons showed different temporal patterns of information coding in the analyses of the empirical neuronal data. Rule-coding neurons started coding the rule information before the cue onset and maintained it throughout the trial. Since the rule was kept unchanged for 6–12 blocks (48–96 trials; see Materials and Methods), the rule information (i.e., which category predicted which outcome) was already known for the monkeys even before the cue onset. This observation is consistent with previous studies reporting that PFC is involved in coding of the task rule or context information (White and Wise, 1999; Wallis et al., 2001; Eiselt and Nieder, 2013; Zhang et al., 2013). Category-coding neurons coded the category information only during the cue period: the encoding started after the cue onset and faded around the cue offset. As the monkeys had been trained with the regular stimulus sets for a long time, they recognized the stimuli associated with the same outcome as a category and used the category to perform the group reversal task (Yamada et al., 2010; Tsutsui et al., 2016). The monkeys needed to judge which category the presented cue stimulus belonged to and infer which outcome was going to be delivered in this trial based on the current rule and the judged category. After this judgment, the category information was no longer necessary because, with the contingency information, the monkeys knew whether they should lick or not lick at the end of this trial. Thus, it is reasonable that the encoding of the category information was activated only during the cue period. Contingency-coding neurons also started coding the contingency information after the cue onset but maintained it until the end of trial. As mentioned above, with the contingency information, the monkey could decide whether they should lick or not lick the spout in the current trial. Thus, it should be necessary to maintain the contingency information until the monkey actually conducted the selected behavior. The latency analysis suggested that the coding of rule and category information preceded that of contingency information (Fig. 4C). Furthermore, the analysis of the evolution of CPDs during the cue period revealed that the rule-coding and category-coding neurons showed stable coding, while the contingency-coding neurons gradually increased the proportion of contingency-coding information, during the cue period. On the other hand, the intermediate-type neurons showed dynamic coding, changing the proportion of each type of coding information during the cue period (Fig. 5). The trajectory of intermediate-type neurons suggests that they coded mainly rule information at the beginning of the period, then recruited more category information, and finally coded mainly the outcome information. These results indicate that while category-coding and rule-coding neurons maintain input information, the intermediate-type neurons integrate the rule and category information to generate the contingency information, and the contingency-coding neurons receive and maintain the output information.

Results of the analysis using the spiking neural network model suggests that this computation can be implemented by hierarchically organized subnetworks. In particular, since each subregion contained all four types of neuron and the functional feature of each type of neuron did not appear to differ between subregions (Tsutsui et al., 2016), we focused on the functional roles of the neurons rather than the subregions they belonged to. However, our model does not rule out the possibility that different subregions of the PFC emphasize different types of information (Miller and Cohen, 2001). In fact, the majority of category-coding neurons were found in the ventrolateral PFC (VLPFC). On the other hand, the intermediate-type and contingency-coding neurons were distributed similarly in the PFC, suggesting that these neurons may come from the same population of neurons. This view is supported by the computational model, where the main difference between neurons in the intermediate and contingency-coding subnetworks depends on whether they received feedforward inputs from the rule-coding and category-coding subnetworks. Finally, we note that the main discrepancy between the simulated and recorded data are a reduced variation in the CPDs (compare Figs. 7A and 4A). This discrepancy suggests that the current network, which has a straight feedforward structure along the information processing hierarchy, may reflect an oversimplified model for what is “embedded” in a population of more recurrently connected neurons. Additional recurrent connections may serve as a structural perturbation to the hierarchically organized subnetworks and cause larger variations in the tuning properties of individual neurons.

In conclusion, our study suggests that the PFC actively maintains relevant information and dynamically changes its internal representations in accordance with behavioral needs. A spiking neural network model with hierarchically organized subnetworks reproduced the computation and was readily scalable to larger numbers of rules and categories, providing a general framework for understanding category-based information processing.

Footnotes

  • This work was supported by Ministry of Education, Culture, Sports, Science and Technology of Japan Grants-in-Aid for Scientific Research on Priority Areas 17022009 and 18020005 and Grants-in-Aid for Scientific Research on Innovative Areas 26112009, 26120704, and 19H05725 (to K.-I.T.), and Grant 26115501 (to T.H.); Japan Society for Promotion of Science Grants-in-Aid for Scientific Research 24243067, 24223004, and 20H00104 and Grants-in-Aid for Young Scientists 17680027 and 19673002 (to K.-I.T.), Grants 15H05707 and 20H05921 (to K.A.), Grant 26750377 (to T.H.), and Grant 21H05163 (to Y.K.); the Strategic Research Program for Brain Sciences from Japan Agency for Medical Research and Development (AMED) Grant Number 17dm0107043h0005 (to K.-I.T.); the AMED Grant Number JP21dm0307009 (to K.A.); Moonshot R&D Grant Numbers JPMJMS2021 (to K.A.) and JPMJMS2292 (to K.-I.T.); and Institute of AI and Beyond of the University of Tokyo (K.A.). We thank Prof. Wolfram Schultz for providing his visual stimulus set.

  • The authors declare no competing financial interests.

  • Correspondence should be addressed to Ken-Ichiro Tsutsui at tsutsui{at}tohoku.ac.jp or Kazuyuki Aihara at kaihara{at}g.ecc.u-tokyo.ac.jp

References

  1. ↵
    2. Abeles M
    (1991) Corticonics: neural circuits of the cerebral cortex. Cambridge: Cambridge University Press.
  2. ↵
    2. Amit DJ,
    3. Brunel N
    (1997) Model of global spontaneous activity and local structured activity during delay periods in the cerebral cortex. Cereb Cortex 7:237–252. doi:10.1093/cercor/7.3.237 pmid:9143444
    OpenUrlCrossRefPubMed
  3. ↵
    2. Artola A,
    3. Singer W
    (1993) Long-term depression of excitatory synaptic transmission and its relationship to long-term potentiation. Trends Neurosci 16:480–487. doi:10.1016/0166-2236(93)90081-v pmid:7507622
    OpenUrlCrossRefPubMed
  4. ↵
    2. Bartolo R,
    3. Averbeck BB
    (2020) Prefrontal cortex predicts state switches during reversal learning. Neuron 106:1044–1054.e4. doi:10.1016/j.neuron.2020.03.024 pmid:32315603
    OpenUrlCrossRefPubMed
  5. ↵
    2. Bezaire MJ,
    3. Soltesz I
    (2013) Quantitative assessment of CA1 local circuits: knowledge base for interneuron-pyramidal cell connectivity. Hippocampus 23:751–785. doi:10.1002/hipo.22141 pmid:23674373
    OpenUrlCrossRefPubMed
  6. ↵
    2. Bliss TV,
    3. Collingridge GL
    (1993) A synaptic model of memory: long-term potentiation in the hippocampus. Nature 361:31–39. doi:10.1038/361031a0 pmid:8421494
    OpenUrlCrossRefPubMed
  7. ↵
    2. Bruner JS,
    3. Olver RR,
    4. Greenfield PM
    (1966) Studies in cognitive growth: a collaboration at the Center for Cognitive Studies. New York: Wiley.
  8. ↵
    2. Cromer JA,
    3. Roy JE,
    4. Miller EK
    (2010) Representation of multiple, independent categories in the primate prefrontal cortex. Neuron 66:796–807. doi:10.1016/j.neuron.2010.05.005 pmid:20547135
    OpenUrlCrossRefPubMed
  9. ↵
    2. Draper NR,
    3. Smith H
    (1998) Applied regression analysis. New York: Wiley.
  10. ↵
    2. Eiselt AK,
    3. Nieder A
    (2013) Representation of abstract quantitative rules applied to spatial and numerical magnitudes in primate prefrontal cortex. J Neurosci 33:7526–7534. doi:10.1523/JNEUROSCI.5827-12.2013 pmid:23616557
    OpenUrlAbstract/FREE Full Text
  11. ↵
    2. Freedman DJ,
    3. Riesenhuber M,
    4. Poggio T,
    5. Miller EK
    (2001) Categorical representation of visual stimuli in the primate prefrontal cortex. Science 291:312–316. doi:10.1126/science.291.5502.312 pmid:11209083
    OpenUrlAbstract/FREE Full Text
  12. ↵
    2. Freedman DJ,
    3. Riesenhuber M,
    4. Poggio T,
    5. Miller EK
    (2003) A comparison of primate prefrontal and inferior temporal cortices during visual categorization. J Neurosci 23:5235–5246. doi:10.1523/JNEUROSCI.23-12-05235.2003
    OpenUrlAbstract/FREE Full Text
  13. ↵
    2. Genovesio A,
    3. Tsujimoto S,
    4. Wise SP
    (2009) Feature- and order-based timing representations in the frontal cortex. Neuron 63:254–266. doi:10.1016/j.neuron.2009.06.018 pmid:19640483
    OpenUrlCrossRefPubMed
  14. ↵
    2. Hosokawa T,
    3. Honda Y,
    4. Yamada M,
    5. Romero MC,
    6. Iijima T,
    7. Tsutsui KI
    (2018) Behavioral evidence for the use of functional categories during group reversal task performance in monkeys. Sci Rep 8:15878. doi:10.1038/s41598-018-33349-3 pmid:30367074
    OpenUrlCrossRefPubMed
  15. ↵
    2. Jung Y,
    3. Larsen B,
    4. Walther DB
    (2018) Modality-independent coding of scene categories in prefrontal cortex. J Neurosci 38:5969–5981. doi:10.1523/JNEUROSCI.0272-18.2018 pmid:29858483
    OpenUrlAbstract/FREE Full Text
  16. ↵
    2. Katori Y,
    3. Sakamoto K,
    4. Saito N,
    5. Tanji J,
    6. Mushiake H,
    7. Aihara K
    (2011) Representational switching by dynamical reorganization of attractor structure in a network model of the prefrontal cortex. PLoS Comput Biol 7:e1002266. doi:10.1371/journal.pcbi.1002266 pmid:22102803
    OpenUrlCrossRefPubMed
  17. ↵
    2. Kiani R,
    3. Esteky H,
    4. Mirpour K,
    5. Tanaka K
    (2007) Object category structure in response patterns of neuronal population in monkey inferior temporal cortex. J Neurophysiol 97:4296–4309. doi:10.1152/jn.00024.2007 pmid:17428910
    OpenUrlCrossRefPubMed
  18. ↵
    2. Klausberger T,
    3. Somogyi P
    (2008) Neuronal diversity and temporal dynamics: the unity of hippocampal circuit operations. Science 321:53–57. doi:10.1126/science.1149381 pmid:18599766
    OpenUrlAbstract/FREE Full Text
  19. ↵
    2. Markram H,
    3. Lübke J,
    4. Frotscher M,
    5. Sakmann B
    (1997) Regulation of synaptic efficacy by coincidence of postsynaptic APs and EPSPs. Science 275:213–215. doi:10.1126/science.275.5297.213 pmid:8985014
    OpenUrlAbstract/FREE Full Text
  20. ↵
    2. Miller EK,
    3. Cohen JD
    (2001) An integrative theory of prefrontal cortex function. Annu Rev Neurosci 24:167–202. doi:10.1146/annurev.neuro.24.1.167 pmid:11283309
    OpenUrlCrossRefPubMed
  21. ↵
    2. Mongillo G,
    3. Amit DJ,
    4. Brunel N
    (2003) Retrospective and prospective persistent activity induced by Hebbian learning in a recurrent cortical network. Eur J Neurosci 18:2011–2024. doi:10.1046/j.1460-9568.2003.02908.x pmid:14622234
    OpenUrlCrossRefPubMed
  22. ↵
    2. Mushiake H,
    3. Saito N,
    4. Sakamoto K,
    5. Itoyama Y,
    6. Tanji J
    (2006) Activity in the lateral prefrontal cortex reflects multiple steps of future events in action plans. Neuron 50:631–641. doi:10.1016/j.neuron.2006.03.045 pmid:16701212
    OpenUrlCrossRefPubMed
  23. ↵
    2. Osherson D,
    3. Smith E,
    4. Wilkie O,
    5. López-Rousseau A,
    6. Shafir E
    (1990) Category-based induction. Psychol Rev 97:185–200. doi:10.1037/0033-295X.97.2.185
    OpenUrlCrossRef
  24. ↵
    2. Rainer G,
    3. Rao SC,
    4. Miller EK
    (1999) Prospective coding for objects in primate prefrontal cortex. J Neurosci 19:5493–5505. doi:10.1523/JNEUROSCI.19-13-05493.1999
    OpenUrlAbstract/FREE Full Text
  25. ↵
    2. Rao SC,
    3. Rainer G,
    4. Miller EK
    (1997) Integration of what and where in the primate prefrontal cortex. Science 276:821–824. doi:10.1126/science.276.5313.821 pmid:9115211
    OpenUrlAbstract/FREE Full Text
  26. ↵
    2. Romo R,
    3. Brody CD,
    4. Hernández A,
    5. Lemus L
    (1999) Neuronal correlates of parametric working memory in the prefrontal cortex. Nature 399:470–473. doi:10.1038/20939 pmid:10365959
    OpenUrlCrossRefPubMed
  27. ↵
    2. Roy JE,
    3. Buschman TJ,
    4. Miller EK
    (2014) PFC neurons reflect categorical decisions about ambiguous stimuli. J Cogn Neurosci 26:1283–1291. doi:10.1162/jocn_a_00568 pmid:24405188
    OpenUrlCrossRefPubMed
  28. ↵
    2. Saito N,
    3. Mushiake H,
    4. Sakamoto K,
    5. Itoyama Y,
    6. Tanji J
    (2005) Representation of immediate and final behavioral goals in the monkey prefrontal cortex during an instructed delay period. Cereb Cortex 15:1535–1546. doi:10.1093/cercor/bhi032 pmid:15703260
    OpenUrlCrossRefPubMed
  29. ↵
    2. Sakamoto K,
    3. Mushiake H,
    4. Saito N,
    5. Aihara K,
    6. Yano M,
    7. Tanji J
    (2008) Discharge synchrony during the transition of behavioral goal representations encoded by discharge rates of prefrontal neurons. Cereb Cortex 18:2036–2045. doi:10.1093/cercor/bhm234 pmid:18252744
    OpenUrlCrossRefPubMed
  30. ↵
    2. Song S,
    3. Abbott LF
    (2001) Cortical development and remapping through spike timing-dependent plasticity. Neuron 32:339–350. doi:10.1016/s0896-6273(01)00451-2 pmid:11684002
    OpenUrlCrossRefPubMed
  31. ↵
    2. Tsutsui K,
    3. Hosokawa T,
    4. Yamada M,
    5. Iijima T
    (2016) Representation of functional category in the monkey prefrontal cortex and its rule-dependent use for behavioral selection. J Neurosci 36:3038–3048. doi:10.1523/JNEUROSCI.2063-15.2016 pmid:26961957
    OpenUrlAbstract/FREE Full Text
  32. ↵
    2. Wallis JD,
    3. Anderson KC,
    4. Miller EK
    (2001) Single neurons in prefrontal cortex encode abstract rules. Nature 411:953–956. doi:10.1038/35082081 pmid:11418860
    OpenUrlCrossRefPubMed
  33. ↵
    2. Wang XJ
    (1999) Synaptic basis of cortical persistent activity: the importance of NMDA receptors to working memory. J Neurosci 19:9587–9603. doi:10.1523/JNEUROSCI.19-21-09587.1999
    OpenUrlAbstract/FREE Full Text
  34. ↵
    2. White IM,
    3. Wise SP
    (1999) Rule-dependent neuronal activity in the prefrontal cortex. Exp Brain Res 126:315–335. doi:10.1007/s002210050740 pmid:10382618
    OpenUrlCrossRefPubMed
  35. ↵
    2. Wolpert DM,
    3. Miall RC,
    4. Kawato M
    (1998) Internal models in the cerebellum. Trends Cogn Sci 2:338–347. doi:10.1016/s1364-6613(98)01221-2 pmid:21227230
    OpenUrlCrossRefPubMed
  36. ↵
    2. Wutz A,
    3. Loonis R,
    4. Roy JE,
    5. Donoghue JA,
    6. Miller EK
    (2018) Different levels of category abstraction by different dynamics in different prefrontal areas. Neuron 97:716–726.e8. doi:10.1016/j.neuron.2018.01.009 pmid:29395915
    OpenUrlCrossRefPubMed
  37. ↵
    2. Yamada M,
    3. Pita MC,
    4. Iijima T,
    5. Tsutsui K
    (2010) Rule-dependent anticipatory activity in prefrontal neurons. Neurosci Res 67:162–171. doi:10.1016/j.neures.2010.02.011 pmid:20193718
    OpenUrlCrossRefPubMed
  38. ↵
    2. Zhang J,
    3. Kriegeskorte N,
    4. Carlin JD,
    5. Rowe JB
    (2013) Choosing the rules: distinct and overlapping frontoparietal representations of task rules for perceptual decisions. J Neurosci 33:11852–11862. doi:10.1523/JNEUROSCI.5193-12.2013 pmid:23864675
    OpenUrlAbstract/FREE Full Text
Back to top

In this issue

Email
Print
View Full Page PDF
Citation Tools
Respond to this article
Monkey Prefrontal Single-Unit Activity Reflecting Category-Based Logical Thinking Process and Its Neural Network Model
Takayuki Hosokawa, Muyuan Xu, Yuichi Katori, Munekazu Yamada, Kazuyuki Aihara, Ken-Ichiro Tsutsui
Journal of Neuroscience 17 August 2022, 42 (33) 6380-6391; DOI: 10.1523/JNEUROSCI.2286-21.2022
Reddit logo Twitter logo Facebook logo Mendeley logo

Jump to section

Keywords

Responses to this article

Respond to this article

Jump to comment:

No eLetters have been published for this article.