Nonlinear feedback modulation contributes to the optimization of flexible decision-making
- Peking-Tsinghua Center for Life Sciences, Peking University, China
- School of Psychological and Cognitive Sciences, Peking University, China
- PKU-IDG/McGovern Institute for Brain Research, Peking University, China
- Department of Neurobiology, The University of Chicago, United States
Figures
Figure 1
Behavioral task.
(A) The flexible visual-motion discrimination (FVMD) task. Monkeys needed to report their decision about the direction of the visual motion stimuli by choosing either the green or red saccade target. The appearance of the two color targets preceded that of the visual-motion stimulus, and the target positions were randomly chosen on each trial to avoid fixed mapping between motion stimulus and saccade direction. Monkeys could initiate their saccade as soon as they had made their decision. (B–C) Psychometric curves for the two monkeys. The averaged performance accuracy from all recording sessions (N=125) for each monkey is plotted as the proportion of trials in which 315° was chosen as a function of the directions and coherence levels of the motion stimuli. Error bar denote ± SEM. (D–E) Chronometric curves are shown separately for the two monkeys. (F–G) Two trial conditions, contralateral (F) and ipsilateral (G), defined according to the spatial configurations of task stimuli during neural recording.
Figure 2 with 5 supplements
Nonlinear feedback modulation of saccade choice on sensory evaluation in LIP.
(A) The activity of one example neuron in the CT condition of the FVMD task is shown for each motion-coherence level. The zero-coherence trials were grouped based on the monkey’s choices. The two vertical dashed lines denote the time of target and motion stimulus onset, respectively. (B) The same example neuron’s activity in the IT condition of the FVMD task. (C–F) The activities of two more example neurons. (G–H) The averaged population activities of all direction-selective neurons (N=83) that were collected during the recording sessions in which the saccade targets were arranged in either horizontal or oblique directions. The activity to each motion direction and coherence level is shown separately for the CT condition (G) and IT condition (H). (I) The activity differences between CT and IT conditions (CT minus IT) of single LIP neurons were plotted for both preferred and nonpreferred motion directions. Each dot represents the activity of a single neuron. The histograms in the horizontal and vertical axes represent the distribution of activity difference between CT and IT conditions for preferred and nonpreferred motion directions, respectively. (J) An ROC analysis was used for quantifying the motion DS for both CT (solid) and IT (dashed) conditions. The colored dots denote the time points for which there was a significant difference between the CT and IT conditions (paired t-test: p<0.01). (K) The average DS for low- and zero-coherence trials is shown as in (J). (L) The mean activity (averaged across all motion directions and coherence levels, shaded area denotes ± SEM) was compared between the two saccade directions (CT vs. IT) at the population level. Activity was aligned to either motion stimulus onset (left panel) or saccade onset (right panel). There was no significant difference between CT and IT conditions before monkeys made saccade choices. (M) Variance in LIP population activity as explained by the individual demixed principal components. Each bar shows the proportion of total explained variance that was contributed by the four task variables. The pie chart shows the total variance explained by each task variable. H, high; M, medium; L, low; 0, zero; P, preferred; NP, nonpreferred.
Figure 2—figure supplement 1
The comparisons of LIP activity between the CT and IT conditions.
The averaged population activity responded to motion stimuli in both CT and IT conditions is shown separately for different motion coherence levels. The averaged LIP activity to the preferred (A–B) and nonpreferred (C–D) motion directions was shown separately. Different colors denote different motion coherence levels. The two vertical dashed lines denote the time of target and motion stimulus onset, respectively.
Figure 2—figure supplement 2
There was no systematic relationship between direction preference and saccade-related modulation in LIP neurons’ responses to motion stimuli.
A modulation index was computed for each neuron to quantify differences in motion direction selectivity between the CT and IT conditions. The distribution of modulation indices was then compared between neurons preferring 315° motion direction (red) and those preferring 135° motion direction (blue).
Figure 2—figure supplement 3
The comparison of LIP activity between the CT and IT conditions.
This figure only includes the data sessions in which the saccade targets were aligned close to the vertical direction. (A–B) The averaged population activities in both the CT (A) and IT (B) conditions are shown separately for each motion direction and coherence level. (C–D) The comparisons of LIP activity responded to the preferred motion direction between the CT and IT conditions are shown for different coherence levels. (E–F). The comparisons of LIP activity responded to the nonpreferred motion direction between the CT and IT conditions.
Figure 2—figure supplement 4
The motion and saccade representations in LIP shown by dPCA analysis.
LIP population activity was decomposed into four task-related variables: motion direction, saccade direction, motion-saccade interaction, and timing (condition-independent). (A) Cumulative variance explained by PCA (black) and dPCA (red) for LIP population activity responded to motion stimuli. Only the first 15 principal components (PCs) were shown. dPCA explains almost the same amount of variance as standard PCA. (B–E) Demixed principal components. The upper row: the first demixed PCs of LIP population activity corresponding to the four variables. The lower row: the second demixed PCs of LIP population activity. In each subplot, LIP population activity is projected onto the respective dPCA decoder axis, so that there are 16 lines corresponding to 16 conditions (2 motion directions x 4 coherence levels x 2 saccade directions). Different colors represent different motion directions. Solid and dashed lines represent CT condition and IT condition, respectively. The two vertical dashed lines mark the time of target and motion stimulus onset, respectively.
Figure 2—figure supplement 5
Modulation of sensory evaluation by saccade selection is unlikely to be an artifact of saccade direction selectivity.
(A) The modulation index is plotted against the RF position for each LIP neuron with identifiable RF, the size and the color of each dot denotes the value of the modulation index. (B) The average of the mean firing rate divided by saccade direction for all neurons that did not show clear RF in the MGS task, aligned to the direction with the strongest neural response. Error bar indicates standard deviation. (C) Comparison of the mean saccadic responses to the MGS targets that are closest to the contralateral and ipsilateral targets in the mean task of all LIP neurons.
Figure 3 with 2 supplements
Feedback modulation specifically impacted the decision-correlated activity.
(A–B) Averaged population activities on low-coherence trials in the CT condition (A) and IT condition (B) are shown separately for correct (corr) and error (err) trials. (C–D) An ROC analysis quantified the stimulus-related and decision-related LIP DS on low-coherence trials. The correlations between LIP neural activity and the monkeys’ decisions about motion direction (red) or the physical direction of the motion stimulus (blue) in both the CT (C) and IT (D) conditions are shown over time. (E–F) Partial correlation analysis revealed the decision-related and stimulus-related components of LIP activity. The values for r-decision (E) (the partial correlation between neuronal activity and monkeys’ choice, given the stimulus direction) and r-stimulus (F) (partial correlation between neuronal activity and stimulus direction, given the monkeys’ choices) were compared between the IT and CT conditions. (G–H) Correlation between LIP DS and the monkeys’ RTs on zero-coherence trials. The choice selectivity on zero-coherence trials is shown for the faster RT and slower RT trials. Shaded areas denote ± SEM. The black stars indicate time periods in which there was a significant difference (paired t-test: p<0.01).
Figure 3—figure supplement 1
The comparison of LIP activity to the preferred motion direction between the faster and slower RT trials in the CT condition.
Data from each coherence level is shown separately. Each dot denotes the averaged activity of a single neuron.
Figure 3—figure supplement 2
The comparisons of LIP activity to the preferred motion direction between the faster and slower RT trials in the IT condition, which are shown in the same format as in Figure 2—figure supplement 4.
Figure 4 with 4 supplements
Multi-module RNNs trained with the FVMD task.
(A) Model schematic of the RNNs. Each RNN consists of nine motion direction tuned input units, eight color tuned target input units, 200 hidden units, and two response units. The hidden layer of each RNN consists of two main modules. Each main module consists of two nominal modules, each of which receives either the visual motion input (motion module) or the target color input (target module). Only the target modules project to the two response units, and each main module projects primarily to one response unit. All four nominal modules were assigned with an equal number of units (25%) in the network, which consisted of 80% excitatory and 20% inhibitory units. (B–C) The performance accuracies (B) and RTs (C) of all 50 trained RNNs are shown separately for each motion coherence level. (D–E) Two example units from an example RNN. (D) The neural activity of an example unit from the motion module is shown for each motion direction and coherence level. (E) The neural activity of an example unit from the target module is shown for each saccade direction and coherence level. (F–G) The motion DS for the motion module (F) and saccade DS for the target module (G) for the example RNN. (H–I) The averaged population activities of all direction-selective units in the motion module of the example RNN are shown for the CT condition (H) and IT condition (I). (J–K) The averaged motion DS in the motion module of the example RNN for both the CT (solid) and IT (dashed) conditions was quantified by ROC analysis. (L–M) Partial correlation analysis. The values for r-decision (L) and r-stimulus (M) are compared between the IT and CT conditions for the example RNN. (N) A comparison of the motion DS in the motion module of all trained RNNs between the CT and IT conditions is shown for each coherence level. (O) A comparison of the r-decision and r-stimulus between the IT and CT conditions is shown for all trained RNNs. (Paired t-test: ***, p<0.001).
Figure 4—figure supplement 1
Examples of unit activity in the example network.
The two upper rows show the activities of 6 example units in the motion module. Different colors denote different motion directions, and different shades denote different coherence levels. The zero coherence trials were grouped based on the network’s choices. The two lower rows show the activities of 6 example units in the target module. Different colors denote different saccade directions.
Figure 4—figure supplement 2
Population level unit activity across all trained RNNs.
(A) The averaged population activity of all motion direction selective units in the motion module of the example RNN. (B) The averaged population activity of units in the target module of the same example RNN is shown for different saccade directions and motion coherence levels. (C–D) The averaged population activity of all 50 trained RNNs is shown in the same format as in a-b. (E–F) An ROC analysis was used for quantifying the motion DS for the motion module (h) and saccade DS in the target module for all 50 trained RNNs.
Figure 4—figure supplement 3
The motion direction selectivity in the motion module was significantly modulated by later saccade choice in the RNNs.
Only the units in the motion module were included, and data from all the trained RNNs was pulled together. (A–B) The averaged population activities of all direction-selective units in all the trained RNNs are shown separately for CT condition (A) and IT conditions (B). (C–D) The motion DS in the motion module of all the trained RNNs was quantified by ROC analysis for all four motion coherence levels. Solid and dashed lines denote data in the CT and IT conditions, respectively. The color dots denote the time points for which there was significant difference between CT and IT conditions (p<0.01, paired t-test). (E–F) Partial correlation analysis. The averaged value of r-decision (E) and r-stimulus (F) across all the trained RNNs is compared between IT and CT conditions. Shaded areas denote ± SEM.
Figure 4—figure supplement 4
The comparison of unit activity in the motion module of the RNNs between the CT and IT conditions.
(A–B) The averaged population activities of the example RNNs responded to the preferred (A) and nonpreferred (B) motion directions are shown separately for the CT (solid) and IT (dashed) conditions. (C–D) The comparisons of unit activity averaged across all the trained RNNs between the CT and IT conditions are shown for different coherence levels.
Figure 5 with 6 supplements
The circuit mechanisms underlying the nonlinear feedback modulation in RNNs.
(A) The averaged cross-module connection weights are shown for both feedforward and feedback connections. Units in the motion module (M) and target module (T) were grouped based on their preferences for motion DS (315° vs. 135°) and target color selectivity (red [r] vs. green [g]), respectively. (B) The correlation between the match extent of the neural encoding between units in the motion module and those in target modules and the connection weights between them. Each dot denotes data from one RNN (N=50). (C–D) A comparison of the performance accuracy (C) and RT (D) of the full-model RNNs, RNNs without feedback connections, and RNNs with shuffled feedback connections. (E–F) The averaged activity of units in the motion module of the example RNN after the feedback connections were ablated is shown for CT (E) and IT (F) conditions. (G) The comparison of the averaged motion DS between CT and IT conditions for all trained RNNs after feedback connections were ablated. An ROC analysis was used to quantify the motion DS. (H) The differences in motion DS between CT and IT conditions were compared between full-model RNNs and RNNs with shuffled feedback connectivity.
Figure 5—figure supplement 1
The motion DS in the motion module of the trained RNNs did not decrease after inactivating feedback connectivity.
(A) The averaged population activity of units in the motion module of all RNNs (N=50) after ablating feedback connections was shown separately for different motion directions and coherence. (B) The averaged population activity of units in the motion module of all RNNs after disrupting feedback connections was shown in the same format as in a. (C) The averaged motion DS in motion module of the full-model RNNs, RNNs without feedback connections, and RNNs with disrupted feedback connections were compared separately for different motion coherence. An ROC analysis was used for quantifying the motion DS. (Paired t-test: **, p<0.01; ***, p<0.001; n.s., not significant).
Figure 5—figure supplement 2
The averaged population activity of all trained RNNs without feedback connections.
(A–B) The averaged population activity of all direction-selective units in the motion module of all the trained RNNs (N=50) after ablating the feedback connections is shown for different motion directions and coherence levels. Data from both CT condition (A) and IT conditions (B) were shown separately. (C) The averaged population activity of units in the target module of all the trained RNNs after ablating feedback connections is shown for different saccade directions and motion coherence levels.
Figure 5—figure supplement 3
The effects of pattern-specific ablation of the feedback connections on the RNNs’ behavior performance.
The feedback connections in either the specialized group or nonspecialized group were ablated separately when tested with the untrained motion stimuli. (A) The comparison of the numbers of feedback connections between selectivity-specialized and nonspecialized groups. The distributions of the total number of feedback connections for all 50 RNNs were shown separately for specialized and nonspecialized groups. The blue and brown vertical dashed lines denote the mean values across all 50 RNNs for specialized connections and nonspecialized connections, respectively. (B) The comparison of the mean weights of feedback connections between selectivity-specialized and nonspecialized groups. (C) The comparison of the networks’ performance accuracy. The performance accuracies of the full-model RNNs, RNNs without selectivity-specialized feedback connections, and RNNs without nonspecialized feedback connections were shown separately for different motion coherence. (D) The comparison of the networks’ RTs. The RTs of the full-model RNNs, RNNs without selectivity-specialized feedback connections, and RNNs without nonspecialized feedback connections were shown separately. (Paired t-test: *, p<0.05; **, p<0.01; ***, p<0.001; n.s., not significant).
Figure 5—figure supplement 4
The averaged activity of RNNs with disrupted feedback connectivity.
(A–B) The averaged activity of units in the motion module of the example RNN after disrupting the feedback connectivity is shown for different motion directions and coherence levels. Data in both IT (A) and CT (B) conditions were shown separately. (C) The averaged activity of units in the target module of the example RNN after disrupting the feedback connectivity is shown separately for different saccade directions and motion coherence levels. (D–E) The averaged population activity of all direction-selective units in the motion module of all the trained RNNs (N=50) after disrupting the feedback connectivity is shown separately for both IT condition (D) and CT conditions (E). (F) The averaged population activity of units in the target module of all the trained RNNs after disrupting feedback connectivity is shown for different saccade directions and motion coherence levels.
Figure 5—figure supplement 5
The behavioral performance and unit activity in the RNNs that were initialized without feedback connectivity.
(A–B) The performance accuracies (A) and reaction times (B) of all 50 trained RNNs are shown separately for each motion coherence level. (C) The averaged population activity of all motion direction selective units in the motion module is shown for each motion direction and coherence level. Data from all 50 networks were averaged. (D) The averaged population activity of all saccade direction selective units in the target module is shown for each saccade direction and coherence level. (E–F) The averaged population activities of all direction-selective units in the motion module of all 50 RNNs were shown separately for CT (E) and IT conditions (F). (G) The comparisons of the motion DS between the CT and IT conditions (quantified by ROC analysis) in the motion module of all 50 RNNs are shown separately for different coherence levels. (H) The comparisons of the performance accuracies between full-model RNNs and RNNs trained without feedback connections (Paired t-test: **, p<0.01; ***, p<0.001; n.s., not significant).
Figure 5—figure supplement 6
The performance accuracies of the additional RNNs trained without feedback connectivity are shown separately for each motion coherence level.
These RNNs were initialized with greater recurrent connection probabilities than the full-model RNNs, so that the number of the total trainable connection weights matched that in the full-model RNNs. (Paired t-test: **, p<0.01; ***, p<0.001; n.s., not significant).
Figure 6
The computational mechanisms underlying the nonlinear feedback modulation in decision-making.
(A) The averaged activity of units in the target module of the example RNN, after the feedback connections were ablated, is shown separately for different saccade directions and coherence levels. (B) The averaged saccade DS in the target module of the example RNN after full feedback ablation. (C) The averaged saccade DS in the target module of the full-model RNNs, RNNs without feedback connections, and RNNs with shuffled feedback connections. (Paired t-test: *, p<0.05; **, p<0.01; ***, p<0.001; n.s., not significant). (D–G) The evolution of the averaged energy landscapes was shown over time. A numerical approximation of the energy landscape in the 1-D decision (saccade choice) subspace is constructed for both full-model RNNs and RNNs with various types of projection-specific inactivation. Each plot represents the averaged results of 50 RNNs, where Position 0 signifies the SVM decision boundary, and the vertical dashed line marks the time of motion stimulus onset. Unvisited portions of the state space are left blank as there is no gradient or potential estimate. (H–K) Averaged numerical estimate of energy landscapes for trials with different task difficulty levels (motion coherence) (N=50). Results from two different saccade choices were averaged together. Only the potential values at positions continuously visited by four or more models were retained. Shaded areas denote ± SEM.
Author response image 1
Author response image 2
Author response image 3
Author response image 4
Author response image 5
Author response image 6
Author response image 7
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Nonlinear feedback modulation contributes to the optimization of flexible decision-making
eLife 13:RP96402.
https://doi.org/10.7554/eLife.96402.3
