Fig 1.
The hypothetical neural process of conflict resolution and the antisaccade tasks used in the study.
Two main neural modules, the action selection and the remapping modules, are involved in the antisaccade process. A. At the onset of the sensory stimulus, the sensory signal drives the action selection module through the automatic pathway 1 and the remapping module through the deliberate pathways 2 & 3. When performing an automatic action driven by the stimulus, the subject does not need to apply a strong top-down control and the remapping module is in the baseline state which does not change the default sensorimotor mapping. B. In contrast, if the subject performs a voluntary action against the automatic one, a strong top-down control is required. The top-down control suppresses the automatic response by temporally inhibiting the action selection module through the pathway 5 and promotes the sensorimotor remapping by facilitating the remapping module through the pathway 4. C The antisaccade task. In the task the color of the fixation signal on the center of the screen serves as the cue for the trial type. In prosaccade trials (top), the subject has to make a saccadic eye movement toward the white target as soon as it appears. In antisaccade trials (bottom), the subject has to make a saccade away from the target. In the present study we simulated three types of antisaccade tasks: Overlap, Gap and NoGap. D. In the Overlap paradigm, the fixation signal stays on throughout the entire trial. E. In the Gap paradigm, the fixation signal is turned off 200 ms before the onset of the target. F. In the NoGap paradigm, the fixation signal is turned off at the same time with the onset of the target.
Fig 2.
The detailed model circuits for antisaccade task.
A. The action-selection module. The two excitatory populations of saccade neurons (SacL & SacR) receive sensory input and modulation from the remapping module. The two populations compete with each other through the inhibitory interneuron population (I0) and also form mutual inhibitions with the fixation neuron populations FNL or FNR through inhibitory interneuron populations I1-I4. FNL and FNR are driven by the fixation signal and the top-down gaze-holding control which depends on the rule signal. B. The remapping module. The visual layer in the remapping module consists of neural populations that represent all possible sensorimotor maps including the direct map (the strongest by default) and the inverted map, which can be facilitated after training. The signals from the direct and inverted maps compete with each other in the decision layer and the outcome drives the downstream action-selection module. C. Schematics of anti- and prosaccade responses. In a prosaccade trial, the left visual signal directly triggers a left saccade by activating SacL. The signal from the direct map in the remapping module may also contribute to the generation of saccades. During an antisaccade, the saccade neurons in SacL and SacR are temporally suppressed by the top-down gaze-holding control. In the mean time the top-down control facilitates the visual response of the inverted map, which strongly activates the decision population DecR. As a result, DecR wins the competition and activates SacR. Note that we neglect the decussation of the nervous systems in order to avoid complexity in the graphical representation.
Table 1.
Synaptic conductance (in nS) of connections in the action-selection module.
Letters preceding each number indicate the type of receptor. N: NMDA, A: AMPA, G: GABA. Asterisks indicate the synapses that are endowed with short-term facilitation. See text for detail.
Fig 3.
The reaction time and percentage correct associated with the decision layer.
The decision layer is capable of producing the mean reaction time and the percentage correct observed in prosaccade and antisaccade tasks if the input levels are properly selected. We measured A the percentage of correct decisions and B the mean reaction times of the decision layer under various input levels from the upstream Dir and Inv neurons. Although the two types of neurons receive the same visual stimulus, their firing rates can be adjusted differently by varying their baseline inputs, which are indicated in the abscissa (for Dir) and in the ordinate (for Inv). The region above the diagonal line (where the input from the inverted pathway is stronger than that from the direct pathway) corresponds to the antisaccade condition and the region below corresponds to the prosaccade condition. A decision is counted as correct when the firing rate of the neural population driven by Dir or Inv hits a present decision threshold (50Hz) first in the prosaccade or antisaccade condition, respectively. The reaction time is the time interval between the onset of the input and the decision. We identified specific baseline input levels (green circles for prosaccades and red circles for antisaccades) which produce the mean reaction time and the percentage correct that match the typical values observed in monkey experiments. These specific baseline inputs were used to determined the strength of the top-down influence in the proposed model.
Fig 4.
1 Simulated neural activity in prosaccade and antisaccade tasks.
The model exhibits distinct neuronal activity (population firing rate) between prosaccade and antisaccade and between fast and slow errors, giving insight into how errors are produced. Here we display population firing rates from four sample trials in the Gap task and the visual target is on the left in all panels. Neural activity in the NoGap or Overlap tasks is qualitatively similar with that of the Gap task displayed here. A. Neurons in the direct map (Dir neurons) are dominant by default. Therefore they exhibit stronger response to the target signal in prosaccade trials. As a result, the corresponding downstream decision neurons (in DecL) on the left win the competition against the decision neurons (in DecR) on the right. DecL then activates the downstream saccade neurons (SacL) and triggers a prosaccade. B. In antisaccade, the top-down control suppresses neurons in the direct map pathway (DirL) while facilitates neurons in the inverted pathway (InvL). The strongly responded InvL neurons drive downstream DecR decision neurons which in turn activate saccade neurons (in SacR) and trigger an antisaccade. The model exhibits two types of errors, fast and slow, in antisaccade trials. C Fast errors are produced due to the subjects being unable to withhold a saccade against the direct target signal input. The decisions were not yet reached in the decision layer in the remapping module when the erroneous saccades were triggered by the action-selection module. D Slow errors originate from wrong decisions made in the decision layer (middle panel). In these trials the subjects were able to withhold a saccade initially against the strong input from the target signal. But the subsequently arrived signal from the decision module carries the wrong information.
Fig 5.
Fast errors in simulated antisaccade result from weak gaze-holding controls in the action-selection module.
A. Distributions of the strength of gaze-holding control in three types of antisaccade responses: correct (black), fast errors (light grey) and slow errors (dark grey). Arrows indicate the mean of the corresponding distributions. Fast errors are significantly associated with weaker control levels. B. Spike rasters (top, thick lines denote saccade onsets) and trial averaged firing rate (bottom) from a simulated build-up neuron recorded in antisaccade trials. Trials with erroneous responses (dark grey) exhibit elevated activity prior to the stimulus onset comparing to the activity in correct antisaccade trials (black). The trend becomes more significant if we only select trials with fast errors (light grey). All activity shown here were recorded from the simulated Gap task.
Fig 6.
The model reproduces the observed diversity in the neuronal responses between prosaccade and antisaccade across brain regions.
In all panels the target stimulus was presented on the left. Each panel displays spike rasters (top) and trial-averaged firing rates (bottom) from sample neurons. In the visual layer of the remapping module, depending on the type of recorded visual neurons we observe A, stronger prosaccade than antisaccade responses (Dir neurons) or B, stronger antisaccade than prosaccade responses (Inv neurons). C & D. Most movement neurons in the remapping module exhibit stronger antisaccade responses than prosaccade responses (C, align to the stimulus onset. D, align to the saccade onset). E, In the action-selection module, saccade neurons exhibit two waves of activity during antisaccade. Neurons receiving the direct visual stimulus develop a fast but weak response which is followed by a strong movement response on the correct side (right). F, if we compare the same neurons (SacR and SacL) between prosaccade and antisaccade, the neuronal responses in the prosaccade trials are stronger than those of the antisaccade trials prior to the saccade onset. All activity shown here were recorded from the simulated Gap task. Activity exhibited in the NoGap or Overlap tasks is qualitatively similar.
Fig 7.
The model produces reaction time distributions in consistent with those observed in experiments.
Here we plot the distributions for antisaccade (correct: white, error: black) and prosaccade (shaded) in the A Gap, B NoGap and C Overlap tasks. In the Gap task, the offset of the fixation signal prior to the onset of the target weakens the subject’s ability to withhold the gaze, which is reflected in the large number of express saccades (RT < 125ms, indicated by arrows) both in the prosaccade and in the erroneous responses in antisaccade trials. This trend is less significant in the NoGap task and disappears in the Overlap task. Due to the relatively small percentage of erroneous responses, the heights of their distributions have been magnified by three-fold for easy viewing.
Fig 8.
Gaze-holding control and remapping control both influence the behavioral outcome but with different effects.
A. Reaction time distributions for the antisaccade trials with correct (white) and erroneous (black) responses in the Gap task. The distributions of erroneous responses in antisaccade have been magnified vertically by three-fold as in Fig 7. The black number to the right of the distribution of correct responses represents the percentage of correct trials. The grey number below the abscissa indicates the ratio between the numbers of fast and slow errors. B-E, same with A but with different strengths of gaze-holding control and remapping control. The strength of the remapping control affects the percentage correct, the mean reaction time and the ratio of fast/slow errors while the strength of the gaze-holding control mainly affects the ratio of fast/slow errors. Arrows indicate the mean RT of correct antisaccades in each panel.
Table 2.
Synaptic conductance (nS) of connections in the remapping module.
Letters preceding each number indicate the type of receptor. N: NMDA, A: AMPA, G: GABA
Table 3.
Synaptic conductance (in nS) of connections between modules.
Letters preceding each number indicate the type of receptor. N: NMDA, A: AMPA, G: GABA.
Table 4.
Levels of the background noise for each neural populations in the model.
Values are given in firing rate (x1000 Hz) / synaptic conductance (nS). All noise inputs are added to the populations as synaptic input with Poisson statistics through AMPA mediated currents.