Neural population dynamics underlying evidence accumulation in multiple rat brain regions
Curation statements for this article:-
Curated by eLife
eLife assessment
This valuable paper presents findings showing that different brain regions were best described by a distinct accumulation model, which all differed from the model that best described the rat's choices. These findings are solid because the authors present a very strong methodological approach. This work will be of interest to a wide neuroscientific audience.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Accumulating evidence to make decisions is a core cognitive function. Previous studies have tended to estimate accumulation using either neural or behavioral data alone. Here we develop a unified framework for modeling stimulus-driven behavior and multi-neuron activity simultaneously. We applied our method to choices and neural recordings from three rat brain regions — the posterior parietal cortex (PPC), the frontal orienting fields (FOF), and the anterior-dorsal striatum (ADS) — while subjects performed a pulse-based accumulation task. Each region was best described by a distinct accumulation model, which all differed from the model that best described the animal’s choices. FOF activity was consistent with an accumulator where early evidence was favored while the ADS reflected near perfect accumulation. Neural responses within an accumulation framework unveiled a distinct association between each brain region and choice. Choices were better predicted from all regions using a comprehensive, accumulation-based framework and different brain regions were found to differentially reflect choice-related accumulation signals: FOF and ADS both reflected choice but ADS showed more instances of decision vacillation. Previous studies relating neural data to behaviorally-inferred accumulation dynamics have implicitly assumed that individual brain regions reflect the whole-animal level accumulator. Our results suggest that different brain regions represent accumulated evidence in dramatically different ways and that accumulation at the whole-animal level may be constructed from a variety of neural-level accumulators.
Impact Statement
A computational framework for combining neural and behavioral data to infer latent dynamics underlying decision-making reveals distinct accumulation dynamics in different brain regions in the rat.
Article activity feed
-
-
eLife assessment
This valuable paper presents findings showing that different brain regions were best described by a distinct accumulation model, which all differed from the model that best described the rat's choices. These findings are solid because the authors present a very strong methodological approach. This work will be of interest to a wide neuroscientific audience.
-
Reviewer #1 (Public Review):
The authors use neural recordings from three different brain areas to assess whether the type of evidence accumulation dynamics in those regions are (1) similar to one another, and (2) similar to best-fitting evidence accumulation dynamics to behavioral choice alone. This is an important theoretical question because it relates to the 'linking hypothesis' that relates neurophysiological data to psychological phenomena. Although the standard evidence accumulation dynamic in describing choice has been the gradual accumulation of evidence, the authors find that those dynamics are not represented equally in all brain regions. Such results suggest that more nuanced computational models are needed to explain how brain areas interact to produce decisions, and the focus of theoretical development should shift away …
Reviewer #1 (Public Review):
The authors use neural recordings from three different brain areas to assess whether the type of evidence accumulation dynamics in those regions are (1) similar to one another, and (2) similar to best-fitting evidence accumulation dynamics to behavioral choice alone. This is an important theoretical question because it relates to the 'linking hypothesis' that relates neurophysiological data to psychological phenomena. Although the standard evidence accumulation dynamic in describing choice has been the gradual accumulation of evidence, the authors find that those dynamics are not represented equally in all brain regions. Such results suggest that more nuanced computational models are needed to explain how brain areas interact to produce decisions, and the focus of theoretical development should shift away from explaining behavioral patterns alone and more toward explaining both brain and behavioral interactions. Given that the authors simply test the assumption that the same dynamics that best explain behavior should also explain neural data, they accomplish their objective using a sophisticated methodology and find evidence *against* this assumption: they find that each region was best described by a distinct accumulation model, which all differed from the model that best described the rat's choices.
I thought this was an excellent paper with a clear scientific objective, direct analysis to achieve that objective, and a very strong methodological approach to leave little doubt that the conclusions they drew from their analyses were as reasonable and accurate as possible.
-
Reviewer #2 (Public Review):
The neural dynamics underlying decision-making have long been studied across different species (e.g., primates and rodents) and brain areas (e.g., parietal cortex, frontal eye fields, striatum). The key question is to what extent neural firing rates covary with evidence accumulation processes as proposed by evidence accumulation models. It is often assumed that the evidence-accumulation process at the neural level should mirror the evidence-accumulation process at the behavioral level. The current paper shows that the neural dynamics of three rat brain regions (the FOF, ADS, and PCC) all show signatures of evidence accumulation, but in distinct ways. Especially the role of the FOF appears to be distinct, due to its dependence on early evidence compared to the other regions. This sheds new light and a new …
Reviewer #2 (Public Review):
The neural dynamics underlying decision-making have long been studied across different species (e.g., primates and rodents) and brain areas (e.g., parietal cortex, frontal eye fields, striatum). The key question is to what extent neural firing rates covary with evidence accumulation processes as proposed by evidence accumulation models. It is often assumed that the evidence-accumulation process at the neural level should mirror the evidence-accumulation process at the behavioral level. The current paper shows that the neural dynamics of three rat brain regions (the FOF, ADS, and PCC) all show signatures of evidence accumulation, but in distinct ways. Especially the role of the FOF appears to be distinct, due to its dependence on early evidence compared to the other regions. This sheds new light and a new interpretation of the role of the FOF in decision-making - previously, it has been described as a region encoding the choice that is currently being committed to; this new analysis suggests it is instead strongly influenced by early evidence.
A major strength of the paper is that the results are achieved through joint modelling of the behavioral and neural data, combined with information on the physical stimulus at hand. Joint models were shown to provide more information on the underlying processes compared to behavioral or neural models alone. Especially the inclusion of the neural data seemed to have greatly improved the quality of inferences. This is a key contribution that illustrates that the sophisticated modelling of multiple sources of data at the same time, pays off in terms of the quality of inferences. Yet, it should be added here, that due to the nature of the task, the behavioral data contained only choices, and not response times, which tend to contain more information regarding the evidence accumulation process than choice alone. It would be interesting to additionally discuss how choice decision times can be modeled with the proposed modelling framework.
A main limitation of the paper is that it does not appear to address a seemingly logical follow-up question: If these three brain regions individually accumulate evidence in distinct manners, how do these multiple brain regions then each contribute to a final choice? The joint models fit each region's data separately, so how well does each region individually 'explain' or 'predict' behavior, and how does the combined neural activity of the regions lead to manifest behavior? I would be very interested in the authors' perspectives on these questions.
There are some remaining questions regarding the specific models used, that I was hoping the authors could clarify. Specifically, in equations 10-11, I was wondering to what extent there might be a collinearity issue. Equation 10 proposes that the firing rates of neurons can vary across time due to two mechanisms: (1) The dependence of the firing rate on the accumulated evidence, and (2) a time-varying trial average (as detailed in Equation 11). If firing rates of the neuron indeed covary with the accumulated evidence and therefore increase across time, how can the effects of mechanisms 1 and 2 be disentangled? Relatedly, the independent noise models model each neuron separately and thereby include many more parameters, each informed by less data. Is it possible that the relatively poor cross-validation of the independent noise model may be a consequence of the overfitting of the independent noise model?
Another related question is how robust the parameter recovery properties of these models are under a wider range of data-generating parameter settings. I greatly appreciate the inclusion of a parameter recovery study (Figure S1C) using a single synthetic dataset, but it could be made even stronger by simulating multiple datasets with a wider range of parameter settings. Such a simulation study would help understand how robust and reliable the estimated parameters of all models are. Similarly, it would be helpful if also the \theta_{y} parameters are shown, which aren't shown in Figure S1C.An aspect of the paper that initially raised confusion with me is that the models fit on the choice data and stimulus information alone, make different predictions for the evidence accumulation dynamics in different regions (e.g., Figure 5A, 6A) and also led to different best-fitting parameters in different regions (Figure S9A). It took me a while to realize that this is due to the data being pooled across different rats and sessions - as such, the behavioral choice data are not the same across regions, and neither is the resulting fit models. This could easily be clarified by adding a few notes in the captions of the relevant figures.
Combined, this manuscript represents an interesting and welcome contribution to an ongoing debate on the neural dynamics of decision-making across different brain regions. It also introduced new joint modelling techniques that can be used in the field and raised new questions on how the concurrent activity of neurons across different brain regions combined leads to behavior.
-
Author response
Reviewer #1 (Public Review):
The authors use neural recordings from three different brain areas to assess whether the type of evidence accumulation dynamics in those regions are (1) similar to one another, and (2) similar to best-fitting evidence accumulation dynamics to behavioral choice alone. This is an important theoretical question because it relates to the 'linking hypothesis' that relates neurophysiological data to psychological phenomena. Although the standard evidence accumulation dynamic in describing choice has been the gradual accumulation of evidence, the authors find that those dynamics are not represented equally in all brain regions. Such results suggest that more nuanced computational models are needed to explain how brain areas interact to produce decisions, and the focus of theoretical development …
Author response
Reviewer #1 (Public Review):
The authors use neural recordings from three different brain areas to assess whether the type of evidence accumulation dynamics in those regions are (1) similar to one another, and (2) similar to best-fitting evidence accumulation dynamics to behavioral choice alone. This is an important theoretical question because it relates to the 'linking hypothesis' that relates neurophysiological data to psychological phenomena. Although the standard evidence accumulation dynamic in describing choice has been the gradual accumulation of evidence, the authors find that those dynamics are not represented equally in all brain regions. Such results suggest that more nuanced computational models are needed to explain how brain areas interact to produce decisions, and the focus of theoretical development should shift away from explaining behavioral patterns alone and more toward explaining both brain and behavioral interactions. Given that the authors simply test the assumption that the same dynamics that best explain behavior should also explain neural data, they accomplish their objective using a sophisticated methodology and find evidence *against* this assumption: they find that each region was best described by a distinct accumulation model, which all differed from the model that best described the rat's choices.
I thought this was an excellent paper with a clear scientific objective, direct analysis to achieve that objective, and a very strong methodological approach to leave little doubt that the conclusions they drew from their analyses were as reasonable and accurate as possible.
We thank the reviewer for their time and appreciate their generous comments.
Reviewer #2 (Public Review):
The neural dynamics underlying decision-making have long been studied across different species (e.g., primates and rodents) and brain areas (e.g., parietal cortex, frontal eye fields, striatum). The key question is to what extent neural firing rates covary with evidence accumulation processes as proposed by evidence accumulation models. It is often assumed that the evidence-accumulation process at the neural level should mirror the evidence-accumulation process at the behavioral level. The current paper shows that the neural dynamics of three rat brain regions (the FOF, ADS, and PCC) all show signatures of evidence accumulation, but in distinct ways. Especially the role of the FOF appears to be distinct, due to its dependence on early evidence compared to the other regions. This sheds new light and a new interpretation of the role of the FOF in decision-making - previously, it has been described as a region encoding the choice that is currently being committed to; this new analysis suggests it is instead strongly influenced by early evidence.
A major strength of the paper is that the results are achieved through joint modelling of the behavioral and neural data, combined with information on the physical stimulus at hand. Joint models were shown to provide more information on the underlying processes compared to behavioral or neural models alone. Especially the inclusion of the neural data seemed to have greatly improved the quality of inferences. This is a key contribution that illustrates that the sophisticated modelling of multiple sources of data at the same time, pays off in terms of the quality of inferences. Yet, it should be added here, that due to the nature of the task, the behavioral data contained only choices, and not response times, which tend to contain more information regarding the evidence accumulation process than choice alone. It would be interesting to additionally discuss how choice decision times can be modeled with the proposed modelling framework.
We thank the reviewer for their generous views on our work. We agree that adding decision times, which could readily be added to our framework, will likely further constrain the inference of the latent model. We are currently pursuing such topics using this framework and appropriate data. We have altered a passage in our Discussion, where we note the various extensions of our model one could pursue, to include response time within the set of behavioral measurements one might include.
A main limitation of the paper is that it does not appear to address a seemingly logical follow-up question: If these three brain regions individually accumulate evidence in distinct manners, how do these multiple brain regions then each contribute to a final choice? The joint models fit each region's data separately, so how well does each region individually 'explain' or 'predict' behavior, and how does the combined neural activity of the regions lead to manifest behavior? I would be very interested in the authors' perspectives on these questions.
We could not share the reviewers view and interest in this question with any more excitement than we already do! Unfortunately, the experiments necessary for answering this question in a satisfying way have not yet been performed (e.g. simultaneous multi-region population recordings). Additionally, our analysis approach, as presented currently, would require some technical alterations to deal with data at that scale. Both efforts are underway, but we feel as though the current manuscript describes the basic modeling framework one would need to use to address these questions if/when such data exists. We have added some text to the Discussion to highlight these exciting future directions:
“An exciting future application of our modeling framework is to model multiple, independent accumulators in several brain regions which collectively give rise to the animal’s behavior. Such a model would provide incredible insight into how the brain collectively gives rise to behavioral choices.”
There are some remaining questions regarding the specific models used, that I was hoping the authors could clarify. Specifically, in equations 10-11, I was wondering to what extent there might be a collinearity issue. Equation 10 proposes that the firing rates of neurons can vary across time due to two mechanisms: (1) The dependence of the firing rate on the accumulated evidence, and (2) a time-varying trial average (as detailed in Equation 11). If firing rates of the neuron indeed covary with the accumulated evidence and therefore increase across time, how can the effects of mechanisms 1 and 2 be disentangled? Relatedly, the independent noise models model each neuron separately and thereby include many more parameters, each informed by less data. Is it possible that the relatively poor cross-validation of the independent noise model may be a consequence of the overfitting of the independent noise model?
Thank you for this important observation. Please see our response to the essential revisions above which addresses this issue. In short, although it is true that firing rates increase with time (with accumulating evidence) they do so in a way that depends on the stimulus, and so just as often as they increase with time, they decrease.
Regarding the poor cross-validation of the independent noise model, we apologize for confusion here — both the shared and independent noise model have exactly the same number of parameters. They only differ in that the latent process for a trial contains unique noise instantiation per trial for the independent noise model and the same instantiating for the shared model. The number of parameters is the same. See above for our response to this issue, and how the manuscript was modified in light of this confusion.
Another related question is how robust the parameter recovery properties of these models are under a wider range of data-generating parameter settings. I greatly appreciate the inclusion of a parameter recovery study (Figure S1C) using a single synthetic dataset, but it could be made even stronger by simulating multiple datasets with a wider range of parameter settings. Such a simulation study would help understand how robust and reliable the estimated parameters of all models are. Similarly, it would be helpful if also the \theta_{y} parameters are shown, which aren't shown in Figure S1C.
We agree that understanding the model fitting behavior under a wider set of parameter settings is valuable. We fit our model to additional sets of parameter settings and included an additional supplemental figure (Figure 1 — figure supplement 2) to illustrate these results. In short, we found that parameter recovery was robust across the different parameter settings we tested. We also updated Figure S1C with the neural parameters. We included the following in the Results to note that parameter recovery was robust:
“We verified that our method was able to recover the parameters that generated synthetic physiologically-relevant spiking and choices data (Figure 1 — figure supplement 1), and that parameter recovery was robust across a range of parameter values (Figure 1 — figure supplement 2)).”
An aspect of the paper that initially raised confusion with me is that the models fit on the choice data and stimulus information alone, make different predictions for the evidence accumulation dynamics in different regions (e.g., Figure 5A, 6A) and also led to different best-fitting parameters in different regions (Figure S9A). It took me a while to realize that this is due to the data being pooled across different rats and sessions - as such, the behavioral choice data are not the same across regions, and neither is the resulting fit models. This could easily be clarified by adding a few notes in the captions of the relevant figures.
Thanks for pointing this out. We agree that this tends to be a point of confusion, and we have added clarification prior to Fig 3, where the choice model is first introduced:
“We stress that because of this, each fitted choice model uses different behavioral choice data, and thus the fitted parameters vary from fitted model to fitted model.”
Combined, this manuscript represents an interesting and welcome contribution to an ongoing debate on the neural dynamics of decision-making across different brain regions. It also introduced new joint modelling techniques that can be used in the field and raised new questions on how the concurrent activity of neurons across different brain regions combined leads to behavior.
We appreciate the very generous views on our work!
-
-