Efficient value synthesis in the orbitofrontal cortex explains how loss aversion adapts to the ranges of gain and loss prospects
Curation statements for this article:-
Curated by eLife
Evaluation Summary:
This work has potential value for researchers in several areas of cognitive and systems neuroscience. Range adaptation is a widespread property in neuronal circuits, and a network mechanism that relates neuronal adaptation to behavioral outputs is a valuable addition to the literature. However, limitations in the current framing and analyses leave some uncertainty about the interpretation of the results and their broader applicability.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Is irrational behavior the incidental outcome of biological constraints imposed on neural information processing? In this work, we consider the paradigmatic case of gamble decisions, where gamble values integrate prospective gains and losses. Under the assumption that neurons have a limited firing response range, we show that mitigating the ensuing information loss within artificial neural networks that synthetize value involves a specific form of self-organized plasticity. We demonstrate that the ensuing efficient value synthesis mechanism induces value range adaptation. We also reveal how the ranges of prospective gains and/or losses eventually determine both the behavioral sensitivity to gains and losses and the information content of the network. We test these predictions on two fMRI datasets from the OpenNeuro.org initiative that probe gamble decision-making but differ in terms of the range of gain prospects. First, we show that peoples' loss aversion eventually adapts to the range of gain prospects they are exposed to. Second, we show that the strength with which the orbitofrontal cortex (in particular: Brodmann area 11) encodes gains and expected value also depends upon the range of gain prospects. Third, we show that, when fitted to participant’s gambling choices, self-organizing artificial neural networks generalize across gain range contexts and predict the geometry of information content within the orbitofrontal cortex. Our results demonstrate how self-organizing plasticity aiming at mitigating information loss induced by neurons’ limited response range may result in value range adaptation, eventually yielding irrational behavior.
Article activity feed
-
-
Evaluation Summary:
This work has potential value for researchers in several areas of cognitive and systems neuroscience. Range adaptation is a widespread property in neuronal circuits, and a network mechanism that relates neuronal adaptation to behavioral outputs is a valuable addition to the literature. However, limitations in the current framing and analyses leave some uncertainty about the interpretation of the results and their broader applicability.
(This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.)
-
Reviewer #1 (Public Review):
The study uses public behavioral and fMRI data to study the range adaptation properties of the orbitofrontal cortex (OFC) during risky choice that requires integrating potential gains and losses. The authors demonstrate how spill-over effects from the range of gains to the sensitivity to losses, cannot be explained by simple efficient coding accounts. The authors construct an artificial neural network (ANN) and show that Hebbian plasticity between attribute-specific and integration units can account for the context-dependent effect in behavior and fMRI data.
This is an interesting study that discusses a potential mechanism for context effects often seen in decision-making. A major concern is that the manuscript focuses on Hebbian plasticity as the key mechanism, whereas the results show that the choice of …
Reviewer #1 (Public Review):
The study uses public behavioral and fMRI data to study the range adaptation properties of the orbitofrontal cortex (OFC) during risky choice that requires integrating potential gains and losses. The authors demonstrate how spill-over effects from the range of gains to the sensitivity to losses, cannot be explained by simple efficient coding accounts. The authors construct an artificial neural network (ANN) and show that Hebbian plasticity between attribute-specific and integration units can account for the context-dependent effect in behavior and fMRI data.
This is an interesting study that discusses a potential mechanism for context effects often seen in decision-making. A major concern is that the manuscript focuses on Hebbian plasticity as the key mechanism, whereas the results show that the choice of activation functions (sigmoidal vs. gaussian) has a comparable contribution to explaining behavior but is not discussed. In addition, the performance of even the best model is not very convincing for extreme ranges of expected value. There are additional moderate and minor concerns with result presentation and interpretation.
-
Reviewer #2 (Public Review):
This work investigated the properties of range adaptation in value-based decision-making using a combination of behavioral analysis, BOLD data, and computational modeling. The authors built on an openly available dataset that included behavior and fMRI data from subjects performing a gambling task. In this task, subjects were given a series of offers with a chance to win or lose varying amounts of money, which they could accept or reject. The authors found that the range of potential gains in a session influenced subjects' sensitivity to losses as well as gains. To examine the source of this behavioral pattern, the authors constructed artificial neural networks (ANNs), each comprised of an input layer with "gain" and "loss" populations and an output "integrator" layer than combined these components. In …
Reviewer #2 (Public Review):
This work investigated the properties of range adaptation in value-based decision-making using a combination of behavioral analysis, BOLD data, and computational modeling. The authors built on an openly available dataset that included behavior and fMRI data from subjects performing a gambling task. In this task, subjects were given a series of offers with a chance to win or lose varying amounts of money, which they could accept or reject. The authors found that the range of potential gains in a session influenced subjects' sensitivity to losses as well as gains. To examine the source of this behavioral pattern, the authors constructed artificial neural networks (ANNs), each comprised of an input layer with "gain" and "loss" populations and an output "integrator" layer than combined these components. In networks with Hebbian plasticity, the range of one input variable influenced sensitivity to the other, producing a choice pattern qualitatively similar to subjects' behavior. Using representational similarity analysis, the authors then showed that the cross-trial patterns of activity in these networks resembled cross-trial patterns of BOLD activation in the orbitofrontal cortex and other value-associated regions during the gambling task. They further showed that the integrator layer of the ANNs shows correlates with value-related variables in a way that qualitatively resembles known responses in orbitofrontal cortex neurons.
The question posed by this paper is worthwhile to address, and the model proposed may be a valuable addition to discussions of range adaptation and the mechanisms that mediate it. However, gaps in the analyses and results leave important questions about the interpretation of the work at its current stage. In particular, it would be valuable to see a more direct evaluation of adaptation in the network, a deeper discussion of behavioral predictions of the model, and a discussion of competing models of adaptation. Addressing these gaps will strengthen the work and increase its relevance in neuroeconomics and the broader neuroscience community.
Strengths:
The argument that Hebbian plasticity may provide a mechanism for adaptation in decision-making circuits is a valuable addition to ongoing work on contextual flexibility and neural adaptation and is worthwhile for researchers in the field to consider. The overall framework and modeling approach presented in this work provides a useful tool for researchers to consider when evaluating the relationship between neural adaptation and behavior.By presenting the behavioral results in the framework of neuronal adaptation, the authors incorporated findings from behavioral economics into a framework that is more interpretable and relevant to neuroscience researchers. By combining multiple methods and relating results back to a fundamental question in behavior, this work demonstrates how computational frameworks can be a tool to link neural and behavioral levels of analysis.
Open science provides the ability to directly build on previous work using one's specific expertise and perspective. This paper illustrates the potential value of this approach, producing analyses that could not be done without either the original publicly available work or the unique expertise and contribution of later authors.
Weaknesses:
While the question addressed in this paper and the proposed mechanistic model are potentially of interest to the broader neuroscience community, clearer and more complete analyses are needed to clarify key aspects of the results.The core contribution of this work is the idea that Hebbian plasticity gives rise to neuronal range adaptation and can explain associated effects on value-based decision-making. However, the properties of range adaptation in the model are never tested or shown directly, and it is not clear how well they correspond to known neuronal data. Since a substantial proportion of the paper focuses on the similarity between model activity and neural data, it is worthwhile to examine adaptation directly in the ANN population and compare it to patterns of adaptation previously observed in value-encoding neurons. While RSA results are interpreted as evidence of this type of adaptation, this interpretation is somewhat limited by the indirect nature of RSA, the between-subjects structure of fMRI datasets, and correlations of both fMRI and ANN activity with choice behavior.
In addition, several important aspects of the modeling results are not discussed. For example, the effect of activation functions on ANN output is not mentioned, despite the fact that there are notable differences between networks with different activation functions. Without understanding the reasoning behind the choice of activation functions and their effect on behavior, it is difficult to evaluate the applicability of these models to physiological networks or to predict how robust they are to perturbations. Similarly, while the authors note qualitative similarities between Hebbian network behavior and human choice patterns, they do not discuss points where Hebbian models diverge from human behavior, despite the presence of clear differences at more extreme parts of the value range.
This modeling framework and its application to neuronal adaptation are valuable to explore and have the potential to influence work in neuroeconomics and other fields. However, in the absence of direct comparisons, it is unclear if the proposed model offers greater explanatory power than previous models of adaptation and how well it corresponds to known patterns of adaptation in value-related neurons.
-
Reviewer #3 (Public Review):
The authors investigate range adaptation in the orbitofrontal cortex by taking advantage of an existing data set on willingness to gamble where two different groups experienced a wider or a narrower range of gains but the same range of losses. They find that sensitivity not only to gains but also to losses changes as a function of the gain range, such that for the part that was common to the two groups, people in the wide range group were less willing to gamble than people in the narrow range group. Moreover, a two-layer artificial network with Hebbian plasticity explains the behavioral effects of ranges and multivariate neural representations of value in the orbitofrontal cortex. The authors conclude that range adaptation occurs at the level of the integration layer rather than at the level of the …
Reviewer #3 (Public Review):
The authors investigate range adaptation in the orbitofrontal cortex by taking advantage of an existing data set on willingness to gamble where two different groups experienced a wider or a narrower range of gains but the same range of losses. They find that sensitivity not only to gains but also to losses changes as a function of the gain range, such that for the part that was common to the two groups, people in the wide range group were less willing to gamble than people in the narrow range group. Moreover, a two-layer artificial network with Hebbian plasticity explains the behavioral effects of ranges and multivariate neural representations of value in the orbitofrontal cortex. The authors conclude that range adaptation occurs at the level of the integration layer rather than at the level of the attribute-specific input layer (where gains and losses are separate). The paper provides a welcome addition to the literature on how range adaptations may come about but would benefit from a couple of clarifications.
Major:
It appears like the Gaussian assumption may explain as much or even more of the variance as the plasticity assumption. However, the results do not really address this point. It would be good to provide some information about it for the behavioral findings, check whether the impression also holds for OFC and vmPFC activity, and discuss what the Gaussian assumption implies for the representation of value as such. After all, the monotonicity assumption pervades most previous research on value representation and seems to have been supported reasonably well so far (sometimes with the refinement that positive and negative coding monotonic signals/neurons may be intermixed). Relatedly, one may assume that the Gaussian assumption primarily holds for chosen value cells. But Figure 6 suggests that offer value units are more common in the model. Please explain.
The paper dismisses simplistic efficient coding scenarios that operate on neurons that transmit gain/loss information based on either finding common coding of gain and loss information but no difference between range groups or a difference between range groups but no common coding of gain and loss information. Did the authors also consider common coding of a) expected value, b) gains only, and differences between range groups in (a) and (b) signals, instead of looking at both gains and losses? Because the range manipulation primarily concerned gains rather than both gains and losses, there may be more power in looking at gains only. It may also be worth mentioning that at least for simple reward prediction error signals, a within-subject design, and regions other than the OFC, the simplistic analysis approach can find both effects (Kirschner et al., 2018, Brain). Of course, some of the mentioned or other differences may explain the difference in findings.
-
-
-