Activity in orbitofrontal neuronal ensembles reflects inhibitory control

Abstract

Stopping, or inhibition, is a form of self-control that is a core element of flexible and adaptive behavior. Its neural origins remain unclear. Some views hold that inhibition decisions reflect the aggregation of widespread and diverse pieces of information, including information arising in ostensible core reward regions (i.e., outside the canonical executive system). We recorded activity of single neurons in the orbitofrontal cortex (OFC) of macaques, a region associated with economic decisions, and whose role in inhibition is debated. Subjects performed a classic inhibition task known as the stop signal task. Ensemble decoding analyses reveal a clear firing rate pattern that distinguishes successful from failed inhibition and that begins after the stop signal and before the stop signal reaction time (SSRT). We also found a different and orthogonal ensemble pattern that distinguishes successful from failed stopping before the beginning of the trial. These signals were distinct from, and orthogonal to, value encoding, which was also observed in these neurons. The timing of the early and late signals was, respectively, consistent with the idea that neuronal activity in OFC encodes inhibition both proactively and reactively.

1 |. INTRODUCTION

Inhibition is a key function of the brain’s executive control system (Aron, Cai, Badre, & Robbins, 2015; Hampshire & Sharp, 2015; Logan, Yamaguchi, Schall, & Palmeri, 2015). The process of inhibitory control over prepotent behavior is often studied with the stop signal task, in which the subject must deliberately withhold a planned action in response to a specific stop-related signal (Eagle et al., 2007; Li, 2006; Logan, 1994; Schall, 2001). Performance in this task predicts dysregulated executive function in psychiatric conditions like drug addiction, obsessive–compulsive disorder and obesity (Chamberlain, Fineberg, Blackwell, Robbins, & Sahakian, 2006; Iacono, Malone, & McGue, 2008; Logan, Schachar, & Tannock, 1997; Nederkoorn, Braet, Eijs, Tanghe, & Jansen, 2006; Schachar, Tannock, Marriott, & Logan, 1995). One convenient feature of the task is that it is validated in several species, meaning that animal models can be used to provide insight into the mechanisms of human inhibition as suggested by earlier studies (Hanes & Carpenter, 1999; Hanes & Schall, 1995; Logan & Irwin, 2000) and reviews (Eagle, Bari, & Robbins, 2008; Pouget, Murthy, & Stuphorn, 2017). Understanding those mechanisms holds promise in developing rational treatments for psychiatric diseases and may also help address philosophical questions about self-control and the nature of volition (Schall, Stuphorn, & Brown, 2002).

In order to stop effectively, our brains must monitor both the sensory world and the internal milieu for information indicating that planned actions have become disadvantageous and should be cancelled. When this control is directly driven by external signals, such as a stop signal, it is known as reactive control (Braver, 2012; Braver, Gray, & Burgess, 2007; Chen, Scangos, & Stuphorn, 2010). In the stop signal task (Logan, 1994; Logan & Cowan, 1984), reactive control can be identified because it occurs after the presentation of a stop signal and before the inferred behavioral response to it, the stop signal reaction time (SSRT; see Schall, 2001; Schall et al., 2002 for review). Stopping decisions can also be influenced and in some cases are entirely determined by internal processes, which can begin even before the start of a trial. This form of inhibitory control is known as proactive control (Braver, 2012; Braver et al., 2007; Chen et al., 2010).

Signatures of reactive control have been observed in the frontal eye fields (FEF) and the primary motor cortex, as well as in midbrain structures like the superior colliculus (SC). Many medial prefrontal structures such as supplementary motor area (SMA), pre-SMA and SEF show signatures of proactive control (Chen et al., 2010; Chikazoe et al., 2009; Majid, Cai, Corey-Bloom, & Aron, 2013; Stuphorn, Brown, & Schall, 2010; Stuphorn & Emeric, 2012). In the case of eye movements, control is most directly determined by processes occurring in the FEF and SC. In these regions, inhibition is driven by a rapid rise in firing rates of a specific subpopulation of neurons—fixation neurons—that gate the activity of another subpopulation—movement neurons (Hanes & Schall, 1996; Logan et al., 2015; Schall, 1991).

What is the source of these inhibition signals? We propose that inhibition likely reflects the integration of diverse forms of information, at varying levels of abstraction, bearing on the need to stop (Eisenreich, Akaishi, & Hayden, 2017; Hampshire & Sharp, 2015). Such inhibition-related signals are likely to be especially prominent throughout the prefrontal cortex, which, directly and indirectly, is positioned to regulate motor processes (Aron, 2007; Duncan, 2001; Hampshire & Sharp, 2015; MacLeod, Dodd, Sheard, Wilson, & Bibi, 2003). We were particularly interested in the orbitofrontal cortex (OFC), a region on the orbital surface that is closely associated with value and reward processing (Padoa-Schioppa, 2011; Rudebeck & Murray, 2014; Schoenbaum, Roesch, Stalnaker, & Takahashi, 2009; Wallis, 2007). The OFC is the major input for sensory information into the PFC: It receives strong visual, auditory, gustatory and olfactory inputs. It also has access to signals relating to internal states, via the medial network (Öngür & Price, 2000). It is also often proposed to occupy an early position in PFC processing hierarchies (Carmichael & Price, 1994, 1996; Fuster, 1988, 2001; Rushworth, Kolling, Sallet, & Mars, 2012; Rushworth, Noonan, Boorman, Walton, & Behrens, 2011). These facts raise the possibility that OFC serves as a first stage (or at least a relatively early stage) for computing preliminary executive signals including, potentially, inhibition ones (Yoo & Hayden, 2018). Multiple studies give the OFC a prominent role in inhibition (Bryden & Roesch, 2015; Chikazoe et al., 2009; Dias, Robbins, & Roberts, 1996; Eagle et al., 2007; Horn, Dolan, Elliott, Deakin, & Woodruff, 2003; Majid et al., 2013; Mishkin, 1964; Roberts & Wallis, 2000). On the other hand, a good deal of work argues against such a direct inhibitory role of OFC (Chudasama, Kralik, & Murray, 2006; Ghods-Sharifi, Haluk, & Floresco, 2008; Rudebeck & Murray, 2014; Schoenbaum, Setlow, Nugent, Saddoris, & Gallagher, 2003). In other words, the OFC has become something of a battlefield in debates about the origins of stopping. Surprisingly, then, very little work has measured its contributions to stopping by examining its responses in stopping tasks. It is particularly important to do so in non-human primates because of uncertainties about the homology between rodent and primate OFC (Heilbronner, Rodriguez-Romaguera, Quirk, Groenewegen, & Haber, 2016).

We tested the hypothesis that OFC carries information relevant to stopping decisions. Specifically, we predicted that successful versus failed stopping would be correlated with different neuronal ensemble patterns with specific time courses. We therefore examined ensemble states of OFC neurons recorded in a stop signal task. We found a significant coding pattern difference that emerged following the stop signal but before the stop signal reaction time. We also found a distinct (i.e., statistically orthogonal) ensemble pattern difference that was observable before the trial onset and that was derived from the same set of neurons. Both of these patterns were distinct from (also orthogonal to) economic (i.e., value) signals, which were also carried by the same neurons. Together, these pattern differences provide evidence that OFC carries information sufficient to influence inhibition and suggest it may do so both reactively and proactively.

2 |. METHODS

2.1 |. Surgical procedures

Two male rhesus macaques (Macaca mulatta, subject J and subject T) served as subjects. All animal procedures were approved by the University Committee on Animal Resources at the University of Rochester and were designed and conducted in compliance with the Public Health Service’s Guide for the Care and Use of Animals.

We used standard techniques as described previously (Strait, Blanchard, & Hayden, 2014). Animals were habituated to laboratory conditions and then trained to perform oculomotor tasks for liquid reward. A Cilux recording chamber (Crist Instruments) was placed over the OFC and attached to the calvarium with ceramic screws. Appropriate anesthesia was used at all times; induction was performed with ketamine, and isoflurane was used for maintenance. For surgical induction, we used 10–15 mg/kg of ketamine, 0.25 mg/kg of midazolam and 2–4 mg/kg of propofol. For maintenance, we used isoflurane, ad-lib level, set depending on active monitoring procedure. For systemic antibiotics, we used cefazolin, and for topical application, we used standard veterinary triple antibiotic. For analgesics, we used meloxicam and, when judged necessary by veterinary staff, buprenorphine.

Post-operative care included close monitoring and restoration of fluid intake. Animals received appropriate analgesics and antibiotics after all procedures. Position was verified by magnetic resonance imaging with the aid of a Brainsight system (Rogue Research Inc.). Recording locations are shown in Figure 1b. Throughout both behavioral and physiological recording sessions, the chamber was kept sterile with regular antibiotic washes and sealed with sterile caps.

FIGURE 1.

Stop signal task and subject behavior: (a) Task framework (b) recording site—Area 13 of the OFC (scan from subject J shown) Behavioral results for subject J are presented in panels (c–e) and for subject T in (f–h) (c, f) Go trial reaction time distributions (d, g) Inhibition function varied as a function of SSDs (e, h) Previous trial had effects in reaction time behavior. Red lines in (c, e) denote the summation SSD-50 + SSRT (Logan 1994; Logan & Cowan, 1984), and red lines in (d, g) denote SSD-50. Error bars in (e, h) represent SEM, and * denotes t test significance with p < .05

Subjects had never previously been exposed to decision-making tasks designed to test stopping decisions. Previous training history for these subjects included several foraging tasks (Blanchard & Hayden, 2014; Calhoun & Hayden, 2015; Hayden, 2018; Pirrone, Azab, Hayden, Stafford, & Marshall, 2018), two types of gambling tasks (Azab & Hayden, 2017, 2018; Farashahi, Donahue, Hayden, Lee, & Soltani, 2019), an attentional task similar to the one used in (Hayden & Gallant, 2013) and two economic choice tasks (Heilbronner & Hayden, 2016; Wang & Hayden, 2017).

2.2 |. Recording site

A Cilux recording chamber (Crist Instruments) was placed over Area 13 of the OFC (Figure 1b). The targeted area expands along the coronal planes situated between 28.65 and 33.60 mm rostral to the interaural plane with varying depth. Position was verified by magnetic resonance imaging with the aid of a Brainsight system (Rogue Research Inc). Neuroimaging was performed at the Rochester Center for Brain Imaging, on a Siemens 3T MAGNETOM Trio Tim using 0.5-mm voxels. We confirmed recording locations by listening for characteristic sounds of white and gray matter during recording, which in all cases matched the loci indicated by the Brainsight system.

2.3 |. Electrophysiological techniques

Single electrodes (Frederick Haer & Co., impedance range 0.8–4 MOhm) were lowered using a microdrive (NAN Instruments) until waveforms between one and five neuron(s) were isolated. Individual action potentials were isolated on a Plexon system. Neurons were selected for study solely based on the quality of isolation; we never preselected based on task-related response properties.

We use single tungsten electrodes to record from a total of 96 neurons in two monkeys (n = 52 in subject J and 44 in subject T). The maximum number of neurons that we ever recorded simultaneously was 7, while the minimum was 1. All neurons are grouped together for analysis with no preselection based on functional or electrophysiological properties.

2.4 |. Eye tracking and reward delivery

Eye position was sampled at 1,000 Hz by an infrared eye-monitoring camera system (SR Research). Stimuli were controlled by a computer running MATLAB (MathWorks) with Psychtoolbox (Brainard & Vision, 1997) and Eyelink Toolbox (Cornelissen, Peters, & Palmer, 2002). A standard solenoid valve controlled the duration of water delivery. The relationship between solenoid open time and water volume was established and confirmed before, during and after recording.

2.5 |. Task paradigm

The task followed standard stop signal task paradigm (Logan, 1994; Logan & Cowan, 1984). Subjects were placed in front of a computer monitor (1,920 × 1,080 px) with black background. Following a brief (300 msec) central fixation on a white circle (radius 25 px, Figure 1), the fixation spot disappeared on the appearance of eccentric saccade target (90 px white square, 2.38°, positioned at 288 px in left or 1,632 px in right of screen, 50% chance). A go trial (67% of trials, randomly selected) was indicated by a go signal which is the peripheral target, whereas a stop trial (33% of trials, randomly selected) was indicated by an additional appearance of a stop signal—a central gray square (90 px square, 2.38°) delayed relative to the go signal presentation. Stop signal delays (SSDs) in the task were set to stabilize at a delay causing approximately 50% successful stopping out of all stop trials recorded for the task in that day; SSDs were modulated through a staircase procedure with intervals of 16 msec. On go trials, subjects were rewarded for a saccade to the go signal and fixating on it for 200 msec, and on stop trials, subjects were rewarded for inhibiting their saccade to go signal and fixating at the stop signal for 400 msec. Water rewards were provided as feedback, and they were contingent on subject’s performance. Rewards were always 125 μl. The inter-trial interval was 800 msec.

The economic choice task had a similar task framework to stop signal task, and they interleaved randomly in an interval of 1–3 trials. In go trials (random 67% of the total), a peripheral target called go offer (offer 1, 90 px white square, 2.38°, positioned at 288 px in left or 1,632 px in right of the screen, 50% chance) was presented, and it was randomly associated with low (15 μl), medium (125 μl) or high (250 μl) reward offers, as indicated by yellow-, blue- and magenta-colored squares (offer 1), respectively. These were called forced choice trials. In stop trials (random 33% of the total), in addition to the go offer, we presented a center stop offer (offer 2, 90 px square, 2.38°) that is delayed with respect to the appearance of go offer (offer 1) with the current stop signal delay computed from the stop signal task. The stop offer (offer 2) was also randomly associated with yellow, blue and magenta colors to indicate low, medium and high reward sizes. The go offer in stop trials was always in blue color to represent medium reward sized offer. The trials in which the subjects chose the go offer are called as choose offer 1 trials, and the ones in which they chose stop offer are called choose offer 2 trials. This setup allowed the subject to make a choice through reward comparison in case of stop trials (choose offer 2 trials) and through a forced choice in case of go trials (choose offer 1 trials). All other parameters were the same as stop signal task.

The stop signal task and the economic choice task were stochastically counterbalanced. Specifically, the two tasks were randomly alternated with the constraint that each trial repeated from 1 to 3 trials (number chosen randomly). Therefore, the stop signal and choice trials were presented in a pseudo-alternating fashion. We devised this pattern to increase (relative to strict alternation or random interleaving) because it increased the proportion of trials in which one followed another. The subjects differentiated stopping and choice tasks well and performed each of the tasks with high efficiency (Figure S1).

2.6 |. Statistical methods

Inhibition function related failed inhibitions to stop signal delay (SSD). The delay from the presentation of go signal that caused 50% successful cancellation in stop signal task (SSD-50) was used for computing stop signal reaction time (SSRT). SSRT was usually computed through median and integration methods (Logan, 1994; Logan & Cowan, 1984; Verbruggen & Logan, 2008). Median method computed median of go trials’ reaction time distribution and then subtracted SSD-50 from it to give SSRT. The integration method computed the point in go trials’ RT distribution whose area was half the whole and then subtracted SSD-50 from it to give SSRT. SSRT computed from both of the above methods gave nearly equal results, and they were averaged to obtain the final SSRT estimates reported for both subjects.

Separate PSTH matrices were constructed by aligning spike rasters to the presentation of the go signal and stop signal for every neuron. Firing rates were calculated in 1 msec bins but were generally analyzed in longer epochs. Normalization procedure was carried out by subtracting the mean firing during inter-trial interval (ITI) time period and then by z-scoring each neuron’s data. For display, PSTHs were smoothed using 200 msec running boxcars. Tests used in the study include two-sample t test for parametric analysis, Wilcoxon rank test for non-parametric analysis, chi-square test for comparing decoder’s classification accuracy against baseline (50% classification accuracy) and Pearson correlation method for correlation analysis. To compute population tuning, we picked neurons with significant (p < .05) differences between successful and failed inhibition trials using the Wilcoxon rank test.

2.7 |. Decoding analysis

We chose a neural network-based decoding technique because it could efficiently analyze population responses in frontal cortex that are highly multiplexed and non-linear. The procedure involves generation of pseudo-population activation patterns (Mante, Sussillo, Shenoy, & Newsome, 2013; Stokes et al., 2013) for a multi-dimensional scaling to obtain OFC ensemble activation patterns (e.g., Cunningham & Byron, 2014; Rigotti et al., 2013; Stokes, 2015; Stokes et al., 2013), and then using the OFC activations for binary classification (Zhang, Cheng, Lin, Nie, & Yang, 2018). To generate population activation states as input patterns for the decoding analysis, we first separated all trials of each neuron by trial conditions (successful and failed inhibition trials). Then, we averaged the activity from randomly sampled 10 trials belonging to a condition, with replacement, to form activation state for a neuron in any particular time period. The details on choosing of trials for any neuron are described below. The averaged responses of all 96 neurons were pooled to generate one population activation state for a particular trial condition and for a specific time period. In case of analysis of a single subject or a smaller ensemble, the corresponding neurons out of 96 are only taken forward for generation of population activation state. One hundred and fifty unique activation patterns were used for the network training in any instance.

The network used to study the stopping patterns had 100 hidden nodes and 2 output nodes each representing one target condition for classification. The number of input nodes is equal to the total number of neurons used for analysis (= 96 in case of analysis from two subjects). The network weights were initialized to small random numbers between −0.01 and 0.01.

The following back-propagation algorithm was used for training the decoders (Haykin & Network, 2004; Rumelhart, McClelland, & Williams, 1986; Werbos, 1974). In the below, the input nodes are denoted by subscript, k, hidden nodes by subscript, j, and output nodes by subscript, i. Output error, e, associated with the network’s response for the p’th input pattern was

$$e_i=\text{desired}\text{output}-y_i$$ (1)

where y_i was the i’th output node response, and desired output was 1/0 if the i’th output node was associated with target trial condition for the corresponding input pattern (e.g., successful inhibition, failed inhibition). Total output error over all input patterns was computed by

$$E=\sum _p{E}_p$$ (2)

Network’s objective was to minimize the squared output error (Equation 1) for the p’th pattern as denoted by Equation (3).

$$E_p=\frac{1}{2}\sum _i{e}_i^2$$ (3)

Response of any node was a hyperbolic tangent function (g) of slope = 5 of the total input $\left(h_i^s\right)$ to it. The output node response, y_i, as a function of its input was calculated as

$$y_i=g\left(h_i^s\right)$$ (4)

where net input (h_i^s) to the output layer was

$$h_i^s=\sum _j{w}_{ij}^s{V}_j$$ (5)

In the above, the weights, w_ij, with superscript, s, indicate the second level of the network between hidden layer and output layer. V_j denoted the output of hidden layer, and it was represented as a function of net input to the hidden node $\left(h_i^f\right)$ as follows:

$$V_j=g\left(h_i^f\right)$$ (6)

and

$$h_j^f=\sum _k{w}_{jk}^f{x}_k$$ (7)

The superscript, f, in Equations (6, 7) denotes the first level of the network between input layer and hidden layer, w_jk were their weights and x_k was the input pattern to neural network.

Weight updates were proportional to the negative change in error for the p’th pattern, E_p, on change in weights. All updates happened trial by trial in the training phase. The update used at the second level was by Equation (8) and that in the first level was by Equation (10).

$$\Delta w_{ij}^s=-\eta \frac{\partial E_p}{\partial w_{ij}^s}=\eta {\delta }_i^s{V}_j$$ (8)

where

$${\delta }_i^s=e{g}^{\prime }\left(h_i^s\right)$$ (9)

$$\Delta w_{jk}^f=\eta {\delta }_j^f{x}_k$$ (10)

where

$${\delta }_j^f=\sum _i{\delta }_i^s{w}_{ij}^s{g}^{\prime }\left(h_j^f\right)$$ (11)

η is the learning rate set to .001 for pre-go and post-stop signal decoder and .01 for reaction time decoder, and g′ denotes first-order derivative of hyperbolic tangent function.

We had two different decoders trained on data from (a) pre-go signal and (b) post-stop signal time periods; the former worked on data aligned to presentation of go signal at time = 0, and the latter worked on data aligned to stop signal. For pre-go decoder, the training data were population activation states generated on averaging the signal from the fixation epoch spanning 300 msec before the presentation of go signal. For post-stop decoder, training data were generated on averaging the firing between 100 msec and 250 msec of stop signal presentation. The entire network was run for n = 100 instances with different random weight initializations to obtain average output performance, and new ensemble patterns for training and testing were generated from the trials in every instance. Training procedure in all instances converged to classification accuracy of above 80%, and the converged weights at the end of training were used for testing of decoder. The testing data used were population activation states generated by averaging 100 msec boxcars that slide with step size of 10 msec (a total of 91 boxcars).

All testing was done on an independent test set that was distinct from the one used for fitting. Specifically, we do a 60% training/40% testing split of our population data, and we make sure the trials used for generating the population training and testing states are different. We also ensure each training and testing set is stratified—it has 50% successful and 50% unsuccessful stop activation patterns. One other important detail that makes our training and testing data different is the timing used for our training and testing set generation. The training uses an average of much broader post-stop and pre-go signal epochs, while the testing set uses a boxcar of much smaller size and a total of 100 moving boxcars, jumping at 10 msec interval between −500 and +500 msec with respect to the stop signal, and −500 and 1,000 msec in case of go signal. The difference in timing ensures the distinctness of our training and testing set.

We performed permutation tests as a control for the decoding performance. In the permutation tests, we shuffled the labels for training and testing 6K times for generating a null distribution, and we found its 95 percentile as 70 percent decoding performance. To that end, the results show significant time points with classification percentages >70 (95 percentile null distribution value) in yellow color.

In the case of the decoder used for analyzing the reaction time ensemble patterns, the inputs to the decoder were either the population activation pattern during time periods 200 msec before and, in another case, after the reaction time. The output of the decoder was the number of coarse reaction time bins for classifying the input data (n = 2) in the range 0.1–0.6 s. A sum of hundred ensemble patterns was generated for training for each target class at any instance. The training procedure was similar to pre-go and post-stop signal decoders. The results presented were higher than the permutation control threshold that provided a chance decoding of 50% and were significantly different from the null distribution (t test, p < .05).

Similarities in the functioning and generalization of pre-go and post-stop decoders were analyzed by comparing their converged weights, as well as by comparing their classification accuracy. The similarity index was computed by cross-correlating converged hidden layer weight vectors (with zero lag) of two decoders of interest. The index was averaged across n (=100) instances of networks with different weight initializations. The similarity index obtained from autocorrelating the weight vectors was used to statistically compare and cross-validate the results from cross-correlation, and the results were significant using t test (t test, t-stat = 210, p < .001). Similarities in classification accuracy at pre-go or post-stop signal period were found by using t test on average performances of the two decoders during n instances (with different random weight initializations).

Cancellation time was defined by the size of test-boxcar window positioned at first instance of at least four consecutive test boxcars (100 msec window moving in intervals of 10 ms) in a row, whose performance was significantly higher than 50% using chi-square test (p < .05). The method avoids false positives that otherwise appear by 99% chance when considering just any one single significant instance of 91 total boxcars. With simulations using Markov chains, we found that at least 4 consecutive significant windows were needed in a row for the claim of significance with p < .001; so the criteria to find at least 4 consecutive significant bins were used to find pre-go and post-stop decoder results (Figure 4) as well as cancellation time. Average latency of cancellation signals to SSRT was found by subtracting SSRT of each subject from the mean cancellation time (90 msec).

FIGURE 4.

Ensemble analysis informs about stopping: Performance of pre-go signal and post-stop decoders to distinguish successful versus failed inhibition pattern in subject J (a1 and a2), subject T (b1 and b2) and both subjects together (c1 and c2), respectively. Error bars represent 2.5 and 97.5 confidence intervals of the data. Time points highlighted through yellow shading denote start time of 100 msec boxcars having percent accuracies of classification above chance of 50% (chi-square test, p < .05), and above the red line that indicates the 95 percentile value from permutation control test. Significant time periods of decoding that pass the permutation test control and the chi-square tests are thereby indicated through yellow shading in the pre-go signal period (a1, b1, c1) and the post-stop signal period (a2, b2, c2), respectively. The results suggest OFC ensembles successfully distinguish pre-go and post-stop coding patterns

The decoder performance is invariant and independent of the “simultaneous” recording property. We use neural networks as decoders that take in neural ensemble activation state as input and compute the behavior (go/stop) as output. Specifically, the input to the decoder is the set of all single units we recorded (asynchronously). The activation state of each neuron is used as input for network training and testing. Importantly, the input activation state is generated by random sampling with replacement on the trials and averaging n such sampled trials from each neuron. The shuffling ensures the absence of any temporal relationship between the trials associated with every neuron in the ensemble. Hence, the generated activation state is independent of any trial-to-trial temporal relationship between neurons that are representative of “simultaneous” recording. Therefore, it is valid for our method to be applied for ensemble decoding irrespective of its simultaneous recording nature, that is, for both simultaneous and non-simultaneous recordings.

The decoder rather is sensitive to the temporal activations of each single neuron independently in an ensemble. Our decoder works on the activations of single neurons, in an independent fashion, and relates the ensemble activation state to behavior. That is, the decoder cares about the temporal activity within a neuron, that is, the event locked response activity of every independent single unit. In that light, our decoder results suggest a relationship between higher decoding efficiency of ensemble activity and time, t, from the presentation of the stop signal (the event of interest, Figure 4 and Figures S2–S4).

Our approach of using a multi-dimensional state decoding of an ensemble with “non-simultaneously” recorded cells has been used by multiple other studies (Gochin, Colombo, Dorfman, Gerstein, & Gross, 1994; Thomas, Hulle, & Vogel, 2001, and discussed in Averbeck, Latham, & Pouget, 2006). Specifically, studies constructed the multi-dimensional ensemble state (Gochin et al., 1994) and used neural networks (Thomas et al., 2001) for their findings.

This is a subtler point, but important. Our method is logically equivalent, in many ways, to several other dimensionality reduction approaches that are used on any neuronal ensemble (for a review, see (Cunningham & Byron, 2014), and other studies (Rigotti et al., 2013; Stokes, 2015; Stokes et al., 2013). It is well known that these methods can be done with no problems on asynchronously collected cells, for the same reason that our similar methods can.

Indeed, the positive effect on the pseudo-population that we observe suggests that we would observe the same effect with simultaneously collected cells. We also apply our methods even to simultaneously recorded smaller ensembles (Figure S3)—letting our decoding method to be general enough to track the activation dynamical state pattern of neural ensembles recorded simultaneously or constructed as pseudo-population (Averbeck et al., 2006; Gochin et al., 1994; Thomas et al., 2001). The results of that analysis (Figure S3) suggest—the same story—that OFC possesses significant information (permutation test control, chi-square test, p < .05, see Methods) of both proactive and reactive stopping codes at the level of ensembles (Figure S3). We also find diversity at the level of response patterns for different OFC ensemble responses.

2.8 |. Reward and stopping index

Reward index for every neuron was measured by linearly regressing the firing at outcome epoch (between reaction time and feedback) to the received reward sizes in neuroeconomic trials. The stopping index was measured as the difference in normalized firing rates (FR) of successful and failed inhibition trials divided by their norm.

$$\text{Stopping}\text{index}=\frac{{\text{FR}}_{\text{sc}}-{\text{FR}}_{\text{snc}}}{\sqrt{{\text{FR}}_{\text{sc}}^2+{\text{FR}}_{\text{snc}}^2}}$$ (12)

Cross-validation tests were performed to support the idea that we had sufficient data to detect an effect had it been there and to suggest that our results of lack of a significant correlation between stopping and reward indices were statistically meaningful. For the cross-validation analysis, all trials within a neuron were randomly separated into two groups, A and B. Stopping and reward indices were computed for those two groups of each neuron. We performed correlations between stopping indices of groups A and B and between reward indices of A and B. A total of n (=100) random permutation instances were performed to generate different A and B sets. The test should ideally show high correlations between indices of A and B for any instance, and we indeed saw positive correlations between stopping-index_A and stopping-index_B, and similarly for reward-index_A and reward-index_B. We confirmed that the actual correlation coefficient between stopping and reward indices in OFC fell within bottom 5% of the coefficients computed for n instances of stopping-index_A and stopping-index_B. The above was also confirmed for n coefficients for reward-index_A and reward-index_B. Figure 5c,d show results of no significant correlations between stopping and reward indices with p < .01.

FIGURE 5.

Unrelated reward and stopping codes: (a, b) Illustration of example neurons tuned to reward sizes. Neuron in panel a (panel b) shows significant positive (negative) correlation to reward amounts. Correlations between stopping and reward indices show no significant effect during 100 msec in (c) pre-go signal (0 to −100 msec from the go signal) and (d) post-stop signal time period (0–100 msec from the stop signal)

3 |. RESULTS

3.1 |. Subjects showed unremarkable behavior in the stop signal task

Subjects performed a standard stop signal task (based on Hanes & Schall, 1995, Figure 1a and Methods). On each trial, following a central fixation, monkeys saw an eccentric target (go signal) that, if fixated, provided a juice reward. On a subset of trials (33%, called stop trials), a second signal (stop signal) appeared at fixation and countermanded the previously instructed saccade. Successful inhibition trials were rewarded. Failed trials (trials in which a saccade was made despite a stop signal) were not. Both subjects showed typical behavior in this task; their performance in stop trials varied as a function of time of presentation of stop signals relative to that of go signal (Figure 1d,g). Median reaction time in go trials was 0.41 and 0.27 s in subject J and subject T, respectively (Figure 1c,f).

The delay between the go signal and the stop signal is called the stop signal delay (SSD), and it varied randomly across trials. We estimated the SSD that leads to approximately 50% successful stopping (SSD-50) because it can help in computing the stop signal reaction time, SSRT (Logan, 1994; Logan & Cowan, 1984; Verbruggen & Logan, 2008). The SSD-50 was 0.27 s for subject J and 0.15 s for subject T. SSRT was 0.14 s for subject J and 0.12 s for subject T. These values are typical of rhesus macaques in these tasks, for example (Hanes & Schall, 1995; Ito, Stuphorn, Brown, & Schall, 2003).

Both subjects showed behavioral effects in the reaction times of successful inhibitions as a function of previous trial conditions (Figure 1e for subject J, Figure 1h for subject T). Successful inhibition trials were shorter when following a successful inhibition trial (subject J: N = 328, subject T: N = 357) as opposed to following a failed inhibition (subject J: N = 111, subject T: N = 60). The statistics for subject J was 360 msec shorter, t test, t-stat = 11.33, p < .0001, and for subject T was 290 msec shorter, t-stat = 11.88, p < .0001. Similarly, successful inhibition trials were shorter when following a go trial (subject J: N = 833, subject T: N = 862) as opposed to following a failed inhibition trial (subject J: 310 msec shorter, t-stat = 9.608, p < .0001 and subject T: 210 msec shorter, t-stat = 7.72, p < .0001).

3.2 |. Selectivity for stopping in single neurons

We recorded responses of 96 neurons (52 in subject J and 44 in subject T) in Area 13 of the OFC (Figure 1b). The number of neurons to be collected was determined a priori based on exploratory analyses of previous datasets and was not adjusted during recording based on analyses performed midexperiment. Responses of example neurons are illustrated in Figure 2. We focus on neural responses throughout the trial to make it easy to compare the stopping-related responses at the time periods before and after SSRT. The responses shown in Figure 2a,b are aligned to the go signal (time zero). Note that while these response patterns are conveniently illustrative, they do not necessarily stand in for the properties of the entire population (see below).

FIGURE 2.

Selectivity for stopping in sample neurons: Activity of example neurons during successful inhibition, failed inhibition and go trials is illustrated with respect to (a, b) go signal presentation and (c, d) stop signal presentation time. Time from start of the go (stop) signal to SSRT is shaded in panels a and b (c and d). Neuron in panel a shows significant difference in firing rates of successful and failed inhibition trials before SSRT. Neuron in panel b shows difference even before the beginning of trial. Neuron in panel c is the same as panel a and shows significant difference in firing rates of inhibition trials before stopping response time. Likewise, neuron in panel d shows difference around few msecs after SSRT

In neuron J19, firing rates following the go signal but before the SSRT were lower on successfully inhibited trials (1.8 spikes/s) than on failed inhibition trials (4.1 spikes/s, Wilcoxon rank test, rank sum = 1,480, p < .05, n = 567 trials, Figure 2a). Note that there is a larger and more prominent modulation in firing rate later in the trial. Given its timing, this modulation likely relates to outcome monitoring, is too late to influence stopping, and is not of interest here. Another example neuron, T25, showed distinct patterns for successful and failed inhibition trials even 500 msec before the beginning of the trial (rank sum = 2,080, p < .05, n = 579 trials, Figure 2b).

The responses shown in Figure 2c,d are aligned to stop signal (time zero). Figure 2c illustrates the activity of the same neuron shown in Figure 2a; its response pattern showed significant differences between successful inhibition trials (1.8 spikes/s) and failed inhibition trials (4.4 spikes/s) that begin after the presentation of stop signal but before SSRT (rank sum = 1,340, p < .05). Finally, neuron T10 (Figure 2d) fired more vigorously on successful than on failed inhibition trials at around 100 msec after the SSRT (rank sum = 2,229, p < .05). Simple population analyses suggest that these individual neurons are somewhat atypical, however.

3.3 |. Population averages provide weak information about stopping

We hypothesized that OFC predictively distinguishes successful from failed inhibition. We focused our analyses on two time periods of the trial: (a) the post-stop signal, but pre-SSRT period and (b) the pre-go signal time period. The post-stop epoch is important because it is when inhibition generated in response to countermanding commands would presumably occur and has therefore been the focus of many studies of stopping (Logan et al., 2015; Schall, 2001; Schall et al., 2002). It corresponds to the time during which reactive control occurs (Stuphorn & Emeric, 2012). The pre-go signal epoch corresponds to a time before the trial begins; signal differences here presumably reflect proactive control (Stuphorn & Emeric, 2012).

Analysis of single neurons did not provide strong evidence for a role for OFC in stopping. The proportion of neurons that individually distinguished successful and failed inhibition trials (regardless of sign) was 8.43% during the 100 msec post-stop signal time period, and it was 10.50% during the 100 msec pre-go signal time period (note that these epochs were selected before analysis in order to reduce the likelihood of inadvertent p-hacking). These proportions were not significantly greater than chance in either of the two key epochs (chi-square stat = 1.22, p = .26 in the post-stop signal time period; chi-square stat = 1.8, p = .17 in the pre-go signal time period). This lack of a detectable effect does not imply that a correlation between stopping and unit activity in OFC does not exist; rather, it suggests that if it does exist it is too weak to detect using conventional methods that focus on single neurons in a sample of the size we collected.

We next tested whether successful and failed inhibition trials have a consistent sign of effect on firing rates. The proportion of significantly positive cells (successful > failed) was 5.40% and was not significantly different from the proportion of significantly negative (successful < failed) cells (3.03% chi-square test, chi-square stat = 0.52, p = .47) in the post-stop signal period. The difference in the sizes of the two cell classes was also not significant before the start of trial at the pre-go signal time period (significantly positive cells 7.55%, significantly negative cells 2.95%, chi-square = 2.40, p = .12).

Next, we looked at grand averages of populations of neurons (Figure 3). We observed no difference between successful and failed inhibition trials either after the stop signal or before the beginning of trial. Specifically, during the post-stop signal time period, responses were slightly less for successful than failed inhibition in subject J (average of 0.3 spikes/s, p = .6, Figure 3b); the opposite pattern was observed in subject T (average of 0.52 spikes/s, p = .53, Figure 3d). Neither effect was statistically significant. Thus, these results suggest that conventional population averages do not reveal information about the pattern of stopping. Together these analyses indicate that if stopping correlates exist in OFC, they are of a different form than they take in regions like FEF and SC.

FIGURE 3.

Population averages provide weak information about stopping: Population activity for successful inhibition and failed inhibition with respect to (a, c) go signal presentation and (b, d) stop signal presentation, for subjects J and T. Time from start of the go (stop) signal to SSRT is shaded in panels a and c (b and d). Data for all SSDs are averaged to present successful and failed inhibition trials. Error bars denote SEM. They do not reveal significant information about the pattern of stopping

3.4 |. Ensemble patterns strongly distinguish successful from failed stopping

Many studies suggest that ensemble patterns possess properties to code for neural information and dynamics, which may not be expressed at the level of single units (Averbeck et al., 2006; Meyers, Freedman, Kreiman, Miller, & Poggio, 2008; Zemel, Dayan, & Pouget, 1998). Some studies extend these thoughts to suggest that information stored in patterns is more important for neural processing than that present in single units (Morcos, Barrett, Rabinowitz, & Botvinick, 2018). Taking inspiration from such recent developments in the theoretical understanding of neural activity, we devised our next analysis on ensemble patterns.

Our decoders take in input from a set of neurons at a time, and the neuronal activation state is computed by averaging 10 randomly sampled trials with replacement for generating a population activation pattern (see Methods). The random shuffling disrupts the trial level temporal relationship between the activation states of every neuronal unit and lets our approach to be applied in general for ensembles irrespective of their simultaneous or synchronous recording nature (Gochin et al., 1994; Thomas et al., 2001). The approach is similar to generation of pseudo-population activation patterns (Mante et al., 2013; Stokes et al., 2013) for a multi-dimensional scaling (Cunningham & Byron, 2014; Rigotti et al., 2013; Stokes, 2015; Stokes et al., 2013), and using the OFC ensemble activations for binary classification (Zhang et al., 2018).

We trained neural network decoders to analyze differences in population activation patterns between successful and failed inhibition that were not measured through unit responses or population averages. We were, again, interested in two time periods: (a) the times after the presentation of the stop signal (which we examined using a decoder trained on post-stop signal pattern, referred to below as post-stop decoder) and (b) the time before the start of trial (which we examined using a decoder trained on pre-go signal pattern, referred to below as pre-go decoder). To ensure we had enough data to detect significant effects, we used 100 msec moving boxcars, and to gain some insight into the time course of effects, we used a 10 msec step size for boxcars. We also ensure equal target class representation in the training and testing datasets (60:40 split, stratified fold and balanced), and the trials used for population state generation for training are different from those for testing dataset in any instance of running the network.

The post-stop signal decoder was able to classify success of an inhibition significantly above chance for 200 msec after the presentation of the stop signal in subject J and 240 msec in subject T. These series are unlikely to occur by chance (p < .001 in all cases, see Methods for specific use of chi-square tests to quantify significance of 4 consecutive bins and the significant bins passed the permutation test control described later in this section, and Figure 4). Notably, the central point of the first bin of the series to reach significance in both subjects occurred before the stop signal reaction time of either subjects (the SSRTs were 140 msec for subject J and 120 msec). We call the central point the cancellation time; it measures the center point latency of first statistically significant difference between successful and failed inhibition trials for the ensemble of neurons. Accounting for the 100 msec window size of each bin led to average cancellation time as 70 msec for subject J and 60 msec for subject T. It preceded the average stopping response by 70 msec in subject J and by 60 msec in subject T, suggesting OFC’s responses may precede the stopping response. We also combined both the subject data and tested the post-stop decoder and found significant decoding for 350 msec from the stop signal for the post-stop decoder (Figure 4).

We then examined the response differences of the pre-go signal decoder. For subject J, high accuracy of decoding was found during the 500 msec before the appearance of go signal. Likewise, it was 240 to 20 msec before the go signal presentation in subject T. We also combined both the subject data and tested the pre-go and found significant decoding for the 500 msec before the go signal for the pre-go decoder (Figure 4). Furthermore, we also tested a case in which we used only the trials whose previous was a go trial. This ensures that the stop trials used for pre-go decoder are uncorrelated in their outcomes to their previous trials, and any significant results in the pre-go signal epoch do not reflect the previous trial outcomes. Its results confirmed our earlier findings, and we found successful differentiation of successful from failed stop trials even before the go signal presentation in the OFC; particularly, high accuracy of decoding was found during the 500 msec before the go signal for subject J, and it was 500 to 20 msec before the go signal in subject T (Figure S2).

These results indicate that the upcoming success or failure of inhibition is decodable from OFC patterns even before the start of the trial (Figure 4, also see Figure S2); our results do not tell us why this correlation exists, although one may infer that it reflects some internal state facilitated by variety of factors such as frequency of task events, frequency of different trial types, motivation, trial sequence, altogether driving successful versus failed inhibition (Chen et al., 2010); thus, it is a likely correlate of proactive control.

We performed a cross-validation analysis, by testing with a randomly permuted dataset whose labels do not identify correctly to the successful versus failed stopping activity. Particularly, we shuffled the labels for training and testing 6K times for a null distribution, and we found that the distribution’s 95 percentile was 70 percentage of decoding performance. To that end, the results show significant time points with classification percentages >70 (95 percentile null distribution value) in yellow color (Figure 4 and Figures S2, S3) during the post-stop signal time period for the post-stop decoder and the pre-go signal time period for the pre-go decoder. This cross-validation assures the lack of bias in our decoding method. We have also shown that our approach can be applied even on smaller ensembles that we recorded at the same time, and the results suggest that OFC possesses significant information (chi-square test, p < .05, permutation test control, see Methods) of both proactive and reactive stopping codes at the level of smaller ensembles (Figure S3), though diverse at the level of their response patterns.

3.5 |. The post-stop and pre-go decoders are statistically orthogonal

We next examined how the two decoders related to each other. That is, we asked whether the patterns that distinguish successful and failed inhibition after the stop signal are related to those that predict inhibition before the trial begins. We did so by comparing the vector of weights of the post-stop decoder and pre-go decoder. We found a very low similarity between them (similarity coefficient, “r,” obtained at zero lag on cross-correlating weight patterns, r = .02 ± .22). This low correlation may be due to noise in our signal. We therefore performed a cross-validation procedure to measure range of values expected from a true correlation with noise. The measured coefficient fell below first percentile of that obtained from autocorrelation (and thus is significant at p ≤ 0.01; 100 randomizations, average r from the randomized sets = 25.84 ± 0.82). The decoding performances during the time periods after the stop signal (t-stat = 6.0491, p = .003) and before the beginning of trial (t test, t-stat = 8.8874, p < .001) were significantly different between the networks trained on pre-go and post-stop signals. The above results suggest that the two decoders that predict successful versus failed inhibition are statistically orthogonal and thus dissimilar in the early and late epochs.

We now present analyses showing that the input to the decoder (the population activation pattern) represents a more distributed neural activation profile in contrast to the activity being concentrated to just a few units. We perform principal component analysis (PCA) to test OFC’s diversity. If OFC indeed has diverse response patterns, each of the top principal components (PCA) represents only a portion of the variance of the population, a number not even to close to half of the variance = 50%. Our PCA results on a sample instance of neural activation pattern for go and stop behavior suggest diverse and distributed nature of response patterns (Figure S4). Therefore, multiple components (Figure S4a,b top panels) are required to reliably represent the variance observed in the entire OFC population. Moreover, many neurons contribute a non-zero weightage value to the top principal components (Figure S4a,b bottom panels), so multiple neurons significantly contribute toward the diversity observed in the OFC.

3.6 |. Activity of OFC ensembles, but not single neurons, correlates with reaction time

Neurons in prefrontal structures such as SEF show linear correlations between single neural responses and trial reaction time (Stuphorn et al., 2010). We next asked whether single neuron responses from our OFC data around response showed correlations to reaction time, the time taken to saccade to the choice target. We computed reaction time as the time difference between the presentation of go signal and time of saccade to target (on trials with such saccades, go trials and failed inhibition trials (Hanes & Schall, 1995; Stuphorn et al., 2010). We used mean firing rates during 200 msec before and after the reaction time for this analysis; our analytical approach was designed to be similar to that used by Stuphorn et al. (2010). We found no correlations between reaction time and firing rates before the reaction time (Pearson correlation, ρ = 0, p = .41). Likewise, we found no correlation after the reaction time (Pearson correlation, ρ = −.01, p = .09). This analysis suggests that activity of single neurons in OFC, unlike those in SEF, does not scale linearly with reaction time (Stuphorn et al., 2010). This lack of observed correlation raises the possibility that downstream regions may not linearly decode the information from OFC for informing the urgency of action execution.

We hypothesized that OFC ensemble responses predict reaction times. To test this idea, we generated population activation patterns from neurons that contained data for discretized reaction time bins in a range of 0.1–0.6 s with step size 300 msec (2 equally sized bins, see Methods). In particular, we asked whether the ensemble response could be accurately classified to discrete reaction time bins in a non-linear fashion. The results show that neural network decoders were able to classify OFC ensembles to correct reaction time bins with greater accuracy than control distribution (average to chance of 50% decoding), when the population activation pattern was generated from 200 msec time periods before (80.99% decoding accuracy, t test against null distribution generated from permutation control tests gave p < .001) and after (65.29% decoding accuracy, t test p < .05) the reaction time. These results suggest that OFC ensemble responses can predict reaction times.

3.7 |. OFC codes for stopping and reward are unrelated

The reward-encoding role of OFC is a hallmark of its function (Padoa-Schioppa, 2011; Schultz, 2000; Wallis, 2007). We therefore wondered whether the stopping-related activity that we observed might be a side effect of its reward roles. For example, it may be that there is some undetectable natural variation in the relative subjective value of the reward offered for correct performance. On trials in which the reward happened to have a slightly lower value, the subject would be less motivated to perform correctly; this fluctuation would then introduce a correlation between firing rates and successful inhibition.

To test for the possibility that our putative inhibition signals were just reward correlates, we took advantage of a second set of trials, collected in a neuroeconomic stopping task; detailed analysis of the results from that task will be the focus of a later manuscript. In this task, subjects chose or rejected a single reward that had one of three values (low, medium and high rewards, see Methods). The two task types, neuroeconomic and standard stop signal paradigms, were randomly interleaved on a trial-by-trial basis. The data from this task allowed us to assess each neuron’s tuning function for anticipated rewards. Responses to different reward amounts by two example neurons are shown in Figure 5a,b. We found tuning for anticipated reward values in the firing activity during the reward feedback time period. For example, we observed a significant positive correlation between reward amount and firing rate in neuron J19 (ρ = .3138, p < .001, Figure 5a, same as Figure 2a but aligned to feedback) and a significant negative one in neuron T10 (ρ = −.143, p = .04, Figure 5b).

If the stopping-related signals were a consequence of reward encoding, we would see a positive correlation between coding patterns for rewards and stopping. We computed a reward index for all neurons by regressing their responses to outcomes against the outcomes themselves. We computed a stopping index for all neurons by subtracting on their firing rate during successful and failed inhibition before the stop signal reaction time (see Methods). We found no correlations between these indices in the post-stop signal time period (Pearson correlation, ρ = .09, p = .4, Figure 5d). Nor did we find such correlations in pre-go signal time period (ρ = −.02, p = .82, Figure 5c).

This lack of correlation may be a sign that the reward code and the stopping code are different. It may also, in theory, be due to lack of sufficient data to detect a significant effect. To test this idea, we performed a cross-validation analysis (see Methods). Specifically, we reasoned that if insufficient data were the problem then a within-sample correlation would also produce no significant correlation. A positive correlation of a within-sample correlation, using randomly sampled half-sized subsets, then, would indicate that our data have sufficient power to detect a significant effect (Blanchard, Hayden, & Bromberg-Martin, 2015). We thus tested whether the correlation coefficient for stopping and reward indices fell below the bottom 5 percentile of the coefficients obtained for within-group correlations. Indeed, the coefficient fell below first percentile of that obtained for 100 randomizations in cross-validation analysis. Figure 5c,d show no correlations between stopping and reward indices with p ≤ 0.01.

4 |. DISCUSSION

We examined the correlates of successful stopping in ensembles of neurons in OFC. We found that ensembles of neurons readily distinguish successful and failed inhibition. These signals were not consistently associated with a higher or lower firing rate, nor were they associated with two discrete sets of neurons, as in FEF and SC (Hanes, Patterson, & Schall, 1998; Stuphorn, Bauswein, & Hoffmann, 2000; Pouget et al., 2017; for detailed discussion of categorical tuning, see Blanchard, Piantadosi, & Hayden, 2017). Our study shows the presence of stopping-related patterns in OFC at two specific time periods, the first one after the presentation of stop signal and the second one before the beginning of trial. The timing of the two stopping-related patterns is reminiscent of the times associated with reactive and proactive control, respectively (Braver, 2012; Braver et al., 2007; Chen et al., 2010; Chikazoe et al., 2009; Hanes et al., 1998; Majid et al., 2013; Stuphorn et al., 2010; Stuphorn & Emeric, 2012). These response patterns indicate that OFC neurons carry a signal that precedes stopping decisions. They suggest, then, that OFC may be part of the pathway or pathways that determine the success or failure of stopping.

Orbitofrontal cortex is marked by its anatomy: It receives strong and diverse sensory inputs, as well as visceral ones, and it projects to more dorsal prefrontal structures that collectively directly regulate behavior (Cavada, Compañy, Tejedor, Cruz-Rizzolo, & Reinoso-Suárez, 2000; Öngür & Price, 2000; Wallis, 2007). Its connections then mean that it has a nearly ideal anatomy for monitoring sensory and reward information forming a first draft of the type of executive signals that can inform—but not determine—action. Its influence is unlikely to be limited to inhibition; its executive functions likely include contingent (rule-based) decisions, working memory, switching and conflict monitoring (Bryden & Roesch, 2015; Lara, Kennerley, & Wallis, 2009; Mansouri, Buckley, & Tanaka, 2014; Sleezer, Castagno, & Hayden, 2016; Sleezer, LoConte, Castagno, & Hayden, 2017). More broadly, these executive signals likely constitute a component of a larger set of output prediction signals that determine OFC’s major role in cognition (Rudebeck & Murray, 2014; Schuck, Cai, Wilson, & Niv, 2016; Wilson, Takahashi, Schoenbaum, & Niv, 2014).

Another well-known role of the OFC is in signaling value (Padoa-Schioppa, 2011; Wallis, 2007). Given this close association, one might expect that its role in inhibition is to signal the current subjective value of stopping. Our data, which indicate that OFC value responses are orthogonal to its stopping responses, argue against this possibility. They are consistent, however, with the broader theory that OFC carries a suite of signals that regulate the ongoing transformation of stimulus into action and that value is one such signal (Yoo & Hayden, 2018). Other executive signals observed in OFC include spatial information and abstract rules (Lara et al., 2009; Luk & Wallis, 2013; Sleezer et al., 2016; Strait et al., 2016; Tsujimoto, Genovesio, & Wise, 2009; Wallis, 2007; Wallis, Anderson, & Miller, 2001), those related to switching and task adjustment (Chase, Tait, & Brown, 2012; Sleezer & Hayden, 2016), and to conflict adaptation (Bryden & Roesch, 2015; Mansouri et al., 2014; Mansouri, Tanaka, & Buckley, 2009).

In foraging theory, decisions are generally framed as accept–reject. From this perspective, binary choices, the mainstay of behavioral economics and microeconomics, are actually better thought of as two somewhat independent accept–reject decisions (Hayden & Moreno-Bote, 2018; Kacelnik, Vasconcelos, Monteiro, & Aw, 2011; Stephens & Krebs, 1986). Each accept–reject decision, in turn, functions like a classic accept–reject foraging decision, that is, as a choice between pursuing and refraining from pursuit. In other words, what appears to be a binary choice may actually be a pair of stopping decisions. If economic choice ultimately boils down to stopping, there is an opportunity for a “grand unified theory” that can explain the two types of decisions. This possibility would help explain, for example, why many of the same regions are involved in both types of decisions. In particular, OFC has demonstrated importance in both economic decisions and inhibition. Progress in this area promises to help shed light on important debates, such as how economic decision-making relates to self-control (Berkman, Hutcherson, Livingston, Kahn, & Inzlicht, 2016; Shenhav, 2017).

Our results therefore suggest that OFC neurons could be a part of the stream contributing to early inhibitory control signals. Earlier studies have suggested that prefrontal structures such as right inferior PFC, FEF, SC and SMA specifically contribute to inhibitory control; our results rather suggest that the emergent control signals could be distributed across many brain structures including the core reward regions (viz. the OFC). We view these results as supporting a broader notion that cognitive function reflects complex distributed processing rather than simple discrete box-and-line organization (Hunt & Hayden, 2017; Yoo & Hayden, 2018). Nonetheless, we acknowledge that our results are only a start. Critically, they must be tested for functional implications, including lesion studies.

There is a possibility that the stopping code which the post-stop decoder picks up to differentiate failed from successful stopping is a motor code (although it does not explain the proactive control codes). However, if true, it implies that OFC has a premotor role. This view in turn is broadly consistent with our central hypothesis and suggests OFC as ultimate regulator of action rather than as a pure and abstract value area (Yoo & Hayden, 2018). One phenomenon that is at least somewhat related to inhibition is reversal learning. Specific (excitotoxic) lesions to OFC that spare passing fibers indicate that this region is not critical for reversal learning (Rudebeck & Murray, 2011). At the same time, larger-scale regional lesions that include these regions, but that also target white matter, suggest a clear role for OFC in reversal learning (Izquierdo, Suda, & Murray, 2004). Our results suggest that, to the extent that reversal learning depends on inhibition, it may reflect a process that does involve, at least in part, the OFC. However, it is clear that reversal learning is a more complex process—it involves dynamic stimulus–response credit assignment and stimulus–reward association learning in addition to inhibitory control. Therefore, understanding OFC lesion effects on reversal learning and devaluation mechanisms is beyond the scope of our data.

Abbreviations:

FEF

frontal eye fields

OFC

orbitofrontal cortex

PFC

prefrontal cortex

SC

superior colliculus

SEF

supplementary eye fields

SMA

supplementary motor area

SSD

stop signal delay

SSD-50

50% successful stopping

SSRT

stop signal reaction time