Balancing true and false detection of intermittent sensory targets by adjusting the inputs to the evidence accumulation process
Curation statements for this article:-
Curated by eLife
eLife assessment
This research provides convincing evidence that standard behavioral modeling and EEG-derived signatures of the decision process may not agree on mechanisms underlying changes in decision strategy. The authors make a strong case for the importance of informing behavioral modeling with putative neural signatures of the corresponding decision processes. The assumptions of this neurally-informed modeling approach should be further explored and clarified to highlight not only its benefits but also potential caveats.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Neuroscience (eLife)
Abstract
Decisions about noisy stimuli are widely understood to be made by accumulating evidence up to a decision bound that can be adjusted according to task demands. However, relatively little is known about how such mechanisms operate in continuous monitoring contexts requiring intermittent target detection. Here, we examined neural decision processes underlying detection of 1 s coherence targets within continuous random dot motion, and how they are adjusted across contexts with weak, strong, or randomly mixed weak/strong targets. Our prediction was that decision bounds would be set lower when weak targets are more prevalent. Behavioural hit and false alarm rate patterns were consistent with this, and were well captured by a bound-adjustable leaky accumulator model. However, beta-band EEG signatures of motor preparation contradicted this, instead indicating lower bounds in the strong-target context. We thus tested two alternative models in which decision-bound dynamics were constrained directly by beta measurements, respectively, featuring leaky accumulation with adjustable leak, and non-leaky accumulation of evidence referenced to an adjustable sensory-level criterion. We found that the latter model best explained both behaviour and neural dynamics, highlighting novel means of decision policy regulation and the value of neurally informed modelling.
Article activity feed
-
-
eLife assessment
This research provides convincing evidence that standard behavioral modeling and EEG-derived signatures of the decision process may not agree on mechanisms underlying changes in decision strategy. The authors make a strong case for the importance of informing behavioral modeling with putative neural signatures of the corresponding decision processes. The assumptions of this neurally-informed modeling approach should be further explored and clarified to highlight not only its benefits but also potential caveats.
-
Reviewer #1 (Public Review):
This manuscript examines how subjects change their decision strategy in a visual motion change detection task between blocks of trials where they sought to detect stronger versus weaker signals. The authors hypothesized that decision bounds would be reduced for the weaker signal condition. While behavioral changes were reasonably consistent with this hypothesis, it was challenged by EEG measures that have been previously found to relate to decision variables. In particular, a Beta-band EEG measure suggested decision bounds being reduced for the stronger signal condition, in distinct contrast to the initial hypothesis. Based on this, the authors developed an alternative behavioral model that could explain behavioral adjustments while having decision bounds that were constrained by the putative signature of …
Reviewer #1 (Public Review):
This manuscript examines how subjects change their decision strategy in a visual motion change detection task between blocks of trials where they sought to detect stronger versus weaker signals. The authors hypothesized that decision bounds would be reduced for the weaker signal condition. While behavioral changes were reasonably consistent with this hypothesis, it was challenged by EEG measures that have been previously found to relate to decision variables. In particular, a Beta-band EEG measure suggested decision bounds being reduced for the stronger signal condition, in distinct contrast to the initial hypothesis. Based on this, the authors developed an alternative behavioral model that could explain behavioral adjustments while having decision bounds that were constrained by the putative signature of the decision variable derived from the EEG Beta amplitude. This alternative model has two central features: 1) Sensory evidence is referenced to an adjustable criterion before accumulation such that evidence below the criterion provides negative input to the decision variable. 2) There is a lower reflecting bound on the decision variable such that it cannot go to negative values.
This experiment makes a strong case for the benefit of reconciling behavioral modeling with ideas in the literature about neural signatures of the corresponding decision processes. In this work, the standard behavioral modeling and the EEG-derived putative signatures of the decision process did not initially agree. This is an important finding. The authors also go further. Something must give. Either the standard modeling approach for this type of task provides an incorrect account or the putative mapping of EEG Beta amplitude to the decision process is incorrect. The authors argue for the former. They build from the starting point that the neural signatures are correct and develop alternative behavioral models that they argue to be consistent with these while also explaining behavior. One of these models (described above) fits the data as well as the standard model. I think this approach and the resulting model are interesting, but I have concerns.
I think the authors should give greater consideration to the possibility that the relationship between the decision process and EEG-derived neural signatures - despite having a basis in previous results - is either incorrect in some ways, not the full story, or might not fully apply to this paradigm. The authors include important analyses that already suggest that point to some degree, such as the analysis of the raw Beta amplitude in Figure 4. I give the authors credit for including that analysis and the critical discussion they have surrounding it. I suggest that the authors should include further discussion on this general point. What is the main evidence for the relationship of Beta amplitude to decision bound? Are there any differences from the previous studies that might break the relationship in this paradigm? With the approach taken here so heavily dependent on that relationship, discussing how it might not be correct seems of utmost importance.
My biggest concern revolves around where the Beta amplitude measurement fits into the model. Figure 3 (neurally informed modeling) shows it informing an evidence-independent component that adds to the decision variable, and it is described as such in the methods. However, Beta amplitude is clearly affected by the evidence (Figure 2A). And in the results, it is described as best corresponding to a decision variable, which would include the influence of the sensory evidence. If Beta amplitude depends on evidence, then an adjustment of the evidence criterion would influence Beta amplitude, even during the ITI. So, I don't see how it can be properly used as the constraining factor for the evidence-independent urgency in neurally-informed modeling.
To illustrate my concern, consider the evidence criterion adjustment model. In this model, the average DV value during the ITI will be closer to the decision bound in the weaker signal condition compared to the stronger signal condition - that is how the model fits the false alarm rate differences. This should be reflected in Beta amplitude if the latter reflects the DV. However, the Beta-derived urgency is highest for the condition with the lowest false alarm rate and lowest for the condition with the highest false alarm rate. It seems there is an inescapable conclusion that Beta amplitude does not fully reflect the main behavioral-driving features of the DV in this paradigm, even in the criterion-adjustment model. That said, I do think the development and consideration of the criterion-adjustment model is an important contribution, even with this shortcoming.
Additional comments:
The criterion-adjustment model appears equivalent to one with a constant negative drift added to the decision variable in addition to the contribution of evidence (and a lower reflecting bound). If true, it seems that there could be an alternative equivalent account that does not involve regulation of the transfer of incoming evidence.
I understand why one of the model parameters (e.g., noise or bound) must be fixed, but I don't understand why the authors didn't keep it the same parameter across all models. Couldn't they have fixed the noise parameters for all the models? Having it different makes it difficult to compare parameter values across the models, especially because the fitted noise term in the neurally-informed models appears dramatically reduced compared to the fixed value of noise used for the bound-adjustment model. On the topic of parameters, can the leak parameter be reported in more intuitive units? It seems to be parameterized as a fraction leak per time step, and I couldn't easily find the time step used. Reporting the leak as a time constant would be immediately understandable.
-
Reviewer #2 (Public Review):
Geuzebroek and colleagues use computational modeling and EEG to investigate how people adjust continuous decision-making across different contexts. By neurally informing computational models of decision-making, they reject models in which in contexts with weaker sensory evidence a lower decision threshold or greater leak is applied, in favor of a model implementing a novel control mechanism, in which an adjustable sensory criterion determines which samples are considered evidence to be accumulated. This work was rigorously performed and in a compelling manner teases apart competing mechanisms to reveal a significant novel one.
The contributions of this work are at least two-fold: First, the work outlines a novel mechanism by which decision-makers adjust to different environments by taking expectations about …
Reviewer #2 (Public Review):
Geuzebroek and colleagues use computational modeling and EEG to investigate how people adjust continuous decision-making across different contexts. By neurally informing computational models of decision-making, they reject models in which in contexts with weaker sensory evidence a lower decision threshold or greater leak is applied, in favor of a model implementing a novel control mechanism, in which an adjustable sensory criterion determines which samples are considered evidence to be accumulated. This work was rigorously performed and in a compelling manner teases apart competing mechanisms to reveal a significant novel one.
The contributions of this work are at least two-fold: First, the work outlines a novel mechanism by which decision-makers adjust to different environments by taking expectations about sensory evidence into account. Second, they demonstrate how behavior alone can be insufficient to tease apart competing models and lead to misattribution of observed behavioral differences and how neural measures can help arbitrate between models and avoid misattribution.
This work is of great relevance for the decision-neuroscience community, calls for a re-examination of previous findings, and opens exciting new avenues for future research.
-
