(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Sequential neuronal processing of number values, abstract decision, and action in the primate prefrontal cortex [1]

['Pooja Viswanathan', 'Animal Physiology', 'Institute Of Neurobiology', 'University Of Tuebingen', 'Tuebingen', 'Anna M. Stein', 'Andreas Nieder']

Date: 2024-02

Decision-making requires processing of sensory information, comparing the gathered evidence to make a judgment, and performing the action to communicate it. How neuronal representations transform during this cascade of representations remains a matter of debate. Here, we studied the succession of neuronal representations in the primate prefrontal cortex (PFC). We trained monkeys to judge whether a pair of sequentially presented displays had the same number of items. We used a combination of single neuron and population-level analyses and discovered a sequential transformation of represented information with trial progression. While numerical values were initially represented with high precision and in conjunction with detailed information such as order, the decision was encoded in a low-dimensional subspace of neural activity. This decision encoding was invariant to both retrospective numerical values and prospective motor plans, representing only the binary judgment of “same number” versus “different number,” thus facilitating the generalization of decisions to novel number pairs. We conclude that this transformation of neuronal codes within the prefrontal cortex supports cognitive flexibility and generalizability of decisions to new conditions.

Despite its significance, the study of decision-making at the neuronal level without considering the subsequent action or choice has received limited attention. When the forthcoming action or choice remains unknown, the granularity of a decision becomes pertinent to investigate. We hypothesized that the neuronal representation of such a decision may encompass either the complete content of the perceptual options under consideration, a condensed version emphasizing the most relevant features, or an entirely abstract other category—an abstract decision. To investigate this, we developed a sequential numerical decision task. If the decision were dependent on the precise sequence of numbers, as outlined in the first hypothesis, we predicted number or sequence selectivity to be privileged over the selectivity for decision. In contrast, an abstract decision would be reflected by a separation of the decision representation both from the retrospective number or sequence representations and the prospective motor plan. We trained 2 monkeys to view 2 dot displays in a sequence and evaluate whether the numerical values of the 2 displays were the same or different. Only after a delay period and in conjunction with a delayed motor response rule could the monkeys plan and execute a response. This task sequence allowed us to investigate the contents and transformation of neuronal codes during the course of an abstract decision process in the primate prefrontal cortex (PFC).

In the field of perceptual decision-making, decisions are traditionally thought of as a plan of action or making a choice [ 6 , 7 ]. In this framework, the decision is equated with motor preparation. This viewpoint is backed by studies that demonstrate a strong connection between decisions and associated motor actions or choices [ 8 ]. In these studies, association areas in the brain have been said to represent many cognitive processes in terms of embodied motor actions [ 9 ]. But decisions may sometimes be made as a result of an action-independent process. Such a dissociation of the decision and associated action has also been reported during tasks that enforce a rule-based report of the decision, where it is considered an abstract decision [ 10 – 13 ]. In one recent study, the action-independent decision process has been conceived as a state of perception or memory sampling [ 14 , 15 ].

Making decisions involves collecting sensory evidence, careful deliberation of options, making a judgment, and the corresponding choice. When a judgment, such as “same” or “different,” is based on 2 sequential stimuli, sensory information must be first represented and stored. Then, a comparison can be performed, that, in turn, prompts the preparation and execution of a response. Thus, any decision is preceded by the representation of stimuli and followed by an action indicating the response. This sequence of events requires that the information carried by the neurons in a given brain area changes with task events which might prioritize one function over another. While aspects of sensory and working memory representations of individual stimuli [ 1 – 3 ], as well as the execution of motor actions are well explored [ 4 , 5 ], how neurons efficiently and flexibly represent the entire sequence of a decision-making process is not known. In particular, whether this sequence of events consists of dissociable processes remains a matter of debate.

Results

Disentangling decision-making from aspects of perception and action has posed a challenge. We developed a sequential numerical decision-making task to separate the processes in time (Fig 1A). Two monkeys learned to compare the number of items (“numerosity”) in 2 sequentially shown displays, to make a decision on whether the numerosity in both displays was same or different, and to respond according to an impending action rule presented at the end of the trial. The displays contained 1, 3, or 9 dots. To rule out the effects of lower level visual features that tend to co-vary with the numerosity of the displays, we used 2 sets of stimuli, one with randomly chosen radii and inter-dot distance (“standard”), another with controlled total area and inter-dot distance (“control”) (S1 Fig). The rule cue instructed the monkeys on the correct response in conjunction with the monkey’s previous decision on whether the dot displays contained the same number of dots or different (Fig 1A): If the numerosities of the first and second stimulus differed, a red rule cue required the monkey to release a response bar, whereas a blue rule cue required the monkey to maintain their hold of the bar. In contrast, if the numerosities matched, the rule contingencies were inverted. Trials were counter-balanced for decision, number of dots used in first or second position, stimulus protocol and rule cue.

PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 1. Numerical decision task and behavior. (A) The behavioral task is depicted by the screen displays and the response bar held by the monkey. Each trial begins with a fixation period requiring the monkey to look at the central fixation spot on the screen and hold the response bar (up to the rule period). Dot displays are shown in sequence, interspersed by a delay of 1 s. Following the second delay, the monkey can respond during the rule period by either releasing the bar or maintaining their hold. Top row shows the task contingencies for “different” numerosities (50% of the trials), bottom row for “same” numerosities. Correct combination of the “same/different” decisions and motor rules (orthogonalized for the 2 decisions) are rewarded by fluid. (B) Performance of monkeys M1 and M2. Top panel, percent correct responses to the 3 different numerosities. Solid lines for standard numerosity trials and dashed lines for control numerosity trials, mean ± SEM. Bottom panel, performance split by motor rule or by decision. Individual circles are behavioral sessions, with the black bar showing the mean ± SEM, bottom panel. (C) Reaction time is plotted as the time from rule cue onset to monkey’s release on trials when decision matched the motor rule. Box plots depict the median across trials with 25%–75% confidence intervals. (D) Recording area in the primate brain shown in gray, sagittal view with some anatomical landmarks. ls, lateral sulcus; sts, superior temporal sulcus; ps, principal sulcus. The data underlying this and all other figures is available at https://doi.org/10.6084/m9.figshare.25046987. https://doi.org/10.1371/journal.pbio.3002520.g001

Behavior in numerical decision task Both monkeys performed well above chance (i.e., 50% performance) in more than a hundred sessions per individual. Overall, discrimination of the 3 numerosities was close to perfect and performance (correct responses) of both monkeys to the 2 rule cues (red/blue) and the 2 decision options (same/different) was around 90% (Fig 1B). No significant difference of the effect of rule cue on performance in either monkey was observed (Monkey M1, n = 110 sessions: Z = 0.40, P = 0.6880; Monkey M2, n = 133 sessions: Z = −0.35, P = 0.73; Wilcoxon rank sum test). Monkey M2 shows a small effect of decision on performance, performing significantly better on “Same” decisions (Monkey M1: Z = −1.95, P = 0.0513; Monkey M2: Z = −5.36, P = 8.2345E-08). The reaction time for trials in which the response rule required an instant bar release differed significantly for both monkeys across decision and rule cue combinations (Fig 1C). Monkey M1 showed slightly lower times for “Red” cue and “Different” trials (median = 0.26 s) than for “Blue” cue and “Same” trials (median = 0.27 s, Z = −18.79, P = 7.7637E-79, Wilcoxon rank sum test). Monkey M2 showed longer times for “Red” cue and “Different” trials (median = 0.35 s) than “Blue” cue and “Same” trials (median = 0.33 s, Z = 34.99, P = 3.1E-268). Similarly, the stimulus protocol had a small effect on the monkey M2’s discrimination performance (Monkey M1, Z = 2.4889, P = 0.0128; Monkey M2, Z = −4.0537, P = 5.0419E-05). On reaction times, stimulus protocol had a small effect in M1, no effect in M2 (Monkey M1, Z = 4.1198, P = 3.791E-05; Monkey M2, Z = −0.7768, P = 0.4373). Overall, performance in both monkeys was comparable across numerosities and decisions.

Encoding of numbers, decision, and action While the monkeys performed this task, we recorded a total of 691 neurons (291 neurons in monkey 1; 400 neurons in monkey 2) from the dorsal bank of the principal sulcus across sessions (Fig 1D). As shown previously [16–19], neurons in the dorsolateral PFC were selective for the number of items, especially during stimulus presentation. We observed that single neurons showed selectivity to different task factors at various times across a single trial. We performed all the subsequent analyses on neuronal responses collected in sliding temporal windows across the trial. Some neurons were selective to numerosity in the first number period (Fig 2A), whereas other discriminated numerosity in both presentation periods (Fig 2B). Yet, other neurons were primarily selective in the delay periods (Fig 2C). Many neurons, however, were selective to more than one task factor, such as both the first and second number (Fig 2B), or both the first number and the decision (Fig 2C). We quantified for each neuron the effect size for stimulus protocol (as a control for visual parameters), first and second number, and decision by calculating the omega-squared, proportion of explained variance (PEV), a measure of information about the respective task factors contained in the neuronal activity (see Methods). Task factors such as the first number and/or the second number, for example, neurons 1 and 2, respectively, (Fig 2D and 2E), or the decision for example neuron 3 (Fig 2F) explained a large proportion of the variance in the trial-by-trial firing rates of single neurons, tested for each neuron against a distribution of PEV values calculated from shuffled labels for each task factor (P < 0.01). Only 13 neurons overall (2%) showed selectivity for the stimulus protocol, 2 neurons for rule cue, and 1 neuron for the monkeys’ action. The temporal PEV-analysis revealed single neurons’ selectivity to multiple task factors; PFC neurons showed mixed selectivity for both numbers and decision (S2 Fig). However, many neurons were selective purely for one factor across time: 58% (68/117) of decision-selective neurons were only selective for decision, whereas 34% (39/115) and 34% (28/83) of number-selective neurons were selective exclusively for the first and second number, respectively. Thirty-five percent (51/147) of neurons were selective for numbers regardless of their order. A large percentage (61%) of these neurons maintained their selectivity across order, i.e., preferred the same number during first and second number presentation (31/51). Thirty-nine percent (20/51) neurons changed their preference. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 2. Single neurons in the prefrontal cortex show selectivity for task factors. (A) Example neuron 1 selective to numerosity in the first number period. The first column shows the same neuron’s responses sorted according to the first number (top), the second number (middle), and the decision (bottom). Each panel consists of a dot-raster histogram where each line is a trial, each dot is an action potential, and the corresponding spike-density histogram below showing time-resolved smoothed average firing rates. In each panel, the respective task factors, the numerosities in the first number period (top) and the second number period (middle), or the decision (bottom) are color coded. (B) Example neuron 2 selective to the first and second numerosity during each number presentation period. This neuron is also selective for the decision. Layout as in (A). (C) Example neuron 3 selective to the decision in the second delay period. Layout as in (A). (D) Information contained in the responses of example neuron 1 (A) about the 4 different task factors expressed as omega-squared percentage explained variance. The dotted lines indicate the 99th percentile of the effect size calculated from the distribution of values obtained from shuffled data. (E) Information contained in the responses of example neuron 2 (B). Layout as in (D). (F) Information contained in the responses of example neuron 3 (C). Layout as in (D). The data underlying this and all other figures is available at https://doi.org/10.6084/m9.figshare.25046987. https://doi.org/10.1371/journal.pbio.3002520.g002

Decoding of task factors across time Across the population, we evaluated how different task factors were organized in time by training support vector machines (SVMs) to decode them. We extracted a population with a minimum number of trials per level for each factor (n = 537 neurons). With this pseudo-population, we asked whether perceptual, cognitive, and motor factors were separable in a time-resolved way. We performed 20 runs of decoding for each task factor by resampling training and test trials to get an average estimate of the pseudo-population. For each factor and run, an SVM was trained on 10 training trials per level with 5-fold cross-validation and its accuracy was tested on 10 held-out trials. The accuracy of decoding was statistically tested against 1,000 permutations of decoding performed with shuffled training trials (P < 0.01). Based on the decoding accuracy, we found that the population initially represented the first number (Fig 3A) and the second number (Fig 3B). The classifier was not able to decode the stimulus protocols we used to control for the area and density of the dot displays, indicating that the neuronal population did not contain information on purely visual and nonnumerical parameters (Fig 3C). Toward the end of the presentation of the second numerosity, the decision (“same” versus “different”) could be decoded with almost 100% accuracy (Fig 3D). In contrast, the impending action rule cue (“red” or “blue”) (Fig 3E) and the subsequent action (“hold” versus “release”) could not be decoded (Fig 3F). This indicates that only the task factors relevant for successful task completion were encoded by the neurons and that the abstract decision rather than rule cue-related and action-related information was encoded by the neuron population. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 3. Relevant task factors can be decoded from the neuronal population in PFC, but not the irrelevant stimulus features nor the impending rule cues or actions. (A) SVM-classifier decoding accuracy for factor “first number” (numerosities 1, 3, 9) plotted against time (mean ± SEM across 20 runs). Dashed horizontal line indicates chance level for the number of classes in each decoder. The true accuracy of decoding on held-out test trials is tested against the 99th percentile of 1,000 permutations of decoding from classifiers trained on shuffled trials. Thickened color lines and red circles plotted along the time axis indicate significant decoding (P < 0.01). (B) Decoding accuracy for factor “second number” (numerosities 1, 3, 9). Same layout as in (A). (C) Decoding accuracy for factor “stimulus protocol” (std/control). Same layout as in (A). (D) Decoding accuracy for the factor “decision” (same/different). Same layout as in (A). (E) Decoding accuracy for factor “rule cue” (red/blue). Same layout as in (A). (F) Decoding accuracy for factor “action” (hold/release). Same layout as in (A). The data underlying this and all other figures is available at https://doi.org/10.6084/m9.figshare.25046987. PFC, prefrontal cortex; SVM, support vector machine. https://doi.org/10.1371/journal.pbio.3002520.g003

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002520

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.