Mechanisms of distributed working memory in a large-scale network of macaque neocortex
Curation statements for this article:-
Curated by eLife
Evaluation Summary:
Mejias and Wang propose here the first large-scale model of the brain that actually performs a cognitive task. Previous models have focused on neural dynamics during the so-called "resting state", in which subjects are not performing any cognitive task - thus, resting. This study is therefore an important improvement in the field of large-scale modelling and will certainly become an influential reference for future modelling efforts.
(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
Neural activity underlying working memory is not a local phenomenon but distributed across multiple brain regions. To elucidate the circuit mechanism of such distributed activity, we developed an anatomically constrained computational model of large-scale macaque cortex. We found that mnemonic internal states may emerge from inter-areal reverberation, even in a regime where none of the isolated areas is capable of generating self-sustained activity. The mnemonic activity pattern along the cortical hierarchy indicates a transition in space, separating areas engaged in working memory and those which do not. A host of spatially distinct attractor states is found, potentially subserving various internal processes. The model yields testable predictions, including the idea of counterstream inhibitory bias, the role of prefrontal areas in controlling distributed attractors, and the resilience of distributed activity to lesions or inactivation. This work provides a theoretical framework for identifying large-scale brain mechanisms and computational principles of distributed cognitive processes.
Article activity feed
-
-
Author Response:
Reviewer #1 (Public Review):
The model proposed here is the first large-scale model that actually performs a cognitive task, which in this case is working memory but could easily extend to decision making in general as is acknowledged by the authors. Briefly, each of the 30 areas are simulated as a rate, Wong-Wang circuit (i.e. two excitatory pools inhibit each other through a third, inhibitory population). The authors use previously collected anatomical data to constrain the model and show qualitatively match with the data, in particular how mnemonic activity emerges somewhat abruptly along the brain hierarchy.
Strengths Previous models have focused on neural dynamics during the so-called "resting state", in which subjects are not performing any cognitive task - thus, resting. This study is therefore an important …
Author Response:
Reviewer #1 (Public Review):
The model proposed here is the first large-scale model that actually performs a cognitive task, which in this case is working memory but could easily extend to decision making in general as is acknowledged by the authors. Briefly, each of the 30 areas are simulated as a rate, Wong-Wang circuit (i.e. two excitatory pools inhibit each other through a third, inhibitory population). The authors use previously collected anatomical data to constrain the model and show qualitatively match with the data, in particular how mnemonic activity emerges somewhat abruptly along the brain hierarchy.
Strengths Previous models have focused on neural dynamics during the so-called "resting state", in which subjects are not performing any cognitive task - thus, resting. This study is therefore an important improvement in the field of large-scale modelling and will certainly become an influential reference for future modelling efforts. As typically done in large-scale modelling, some anatomical data is used to constrain the model. The model shows several interesting characteristics, in particular how distributed working memory is more resilient to distractors and how the global attractors can be turned off by inhibition of only top areas.
Weaknesses Some of these results are not clear how they emerge, and some "biological constraints" do not seem to constrain. Moreover, some claims are slightly exaggerated, in particular how the model matches the data in the literature (which in some cases it does not) or how somatosensory working memory can be simulated by simply stimulating the "somatosensory cortex".
This paper has two different models, one being a simplified version of the main model. However, it is not very clear what the simplified model adds the main findings, if not to show that the empirical anatomical connectivity does not constrain the full model.
We thank the reviewer for this evaluation, and for appreciating the innovative character of our study in implementing a cognitive function in a data-constrained large-scale brain model. We hope that it will be useful for future studies planning to add cognitive functions to their large-scale models, and also for experimentalists who might benefit from this insight.
In response to the detailed comments of the reviewer, and to address the weaknesses identified above, we have rewritten parts of the text, clarified important concepts and included a new simulations. Briefly:
-We have clarified the nature and effects of the ‘biological constraints’ that we use. The full model that we use is indeed data-constrained, in the sense that we use real data to determine the values of many parameters. Having a data-constrained model, however, does not mean that all the results will be equally constrained. Some model results will critically depend on (some) data used to constrain the model, while other results will be more robust to changes in these parameters. We have highlighted this point and we also added explanations for each of the results presented.
-We have corrected several claims along the text to make it more in line with experimental evidence, and included the new references suggested by the reviewer to this effect. For example, for the case of somatosensory WM mentioned by the reviewer, we have indicated that the existence of a ‘gating’ mechanism (explored in a supplementary figure) is important for achieving an accurate match with the experimentally observed effects of somatosensory stimulation.
-Finally, we have highlighted the complementary benefits of the full and simplified models, and improved our motivation for the latter. Briefly, the simplified model allows us to identify the key ingredients needed for distributed WM (useful to generalize to other animal models), while the full model ensures that the main findings are still present when more realistic assumptions are made. A good example is the counterstream inhibitory bias, which is in principle not necessary for a simplified model but becomes a crucial factor to implement the distributed WM mechanism in our macaque model.
Reviewer #2 (Public Review):
There is a lot to like about this manuscript. It provides a large-scale model of a well-known phenomenon, the "delay activity" underlying working memory, our oldest and most enduring model of a cognitive function. The authors correctly state that despite the ubiquity of delay activity, there is little known about the macro and micro circuitry that produces it. The authors offer a computational model with testable hypotheses that is rooted in biology. I think this will be of interest to a wide variety of researchers just as delay activity is studied across a variety of animal models, brain systems, and behavior. It is also well-written.
My main concern is the authors may be self-handicapping the impact of their model by not taking into account newer observations about delay activity. For a number of years now, evidence has been building that working memory is more complicated than "persistent activity" alone. Stokes, Pasternak, Dehaene, Miller and others have been mounting considerable evidence for more complex dynamics and for "activity-silent" mechanisms where memories are briefly held in latent (non-active) forms between bouts of spiking. There is also mounting evidence that the thalamus plays a key role in working memory (and attention). In particular, higher thalamic nuclei are critical for regulating cortical feedback. Cortical feedback plays a central role in the model presented here. The model presented in this manuscript just deals with persistent attractor states and the cortex alone.
This is not to say that this manuscript does not have good value as is. No one disputes that some form of elevated, sustained, activity underlies working memory. This work adds insights into how that activity gets sustained and the role of, and interactions between, different cortical areas. The observation that the prefrontal and parietal cortex are more critical than other areas, that there are "hidden" attractor states, and "counterstream inhibitory bias" are important insights (and, importantly, testable). They will likely remain relevant even as the field is moving beyond persistent attractor states alone as the model for working memory. The new developments do not argue against the importance of delay activity in working memory. They show that it is more to the story, as inevitably happens in brain science.
The authors do include a paragraph in the Discussion referencing the newer developments. Kudos to them for that. However, it presented as "new stuff to address in the future". Well, that future is now. These "newer" developments have been mounting over the past 10 years. The worry here is that by relying so heavily on the older persistent attractor dynamics model and presenting it as the only model, the authors are putting an early expiration date on their work, at least in terms of how it will be received and disseminated.
We thank the reviewer for a careful and positive evaluation of our work. We consider that the main point raised here is indeed crucial: classical explanations of WM based on elevated and constant firing are an important part of the story, however other alternative or complementary approaches developed in the past years also deserve attention. These approaches include, to name a few, activitysilent mechanisms (Mongillo et al. 2008, Trübutschek et al. 2017), dynamic hidden states (Wolff et al. 2017), persistent activity without feedback (Goldman 2009), and paradigms relying on gamma bursts (Miller et al. 2018).
It’s important to highlight, however, that our approach is “attractor network theory” not “persistent activity theory”, and an attractor does not have to be a steady state (tonic firing) but may display complex spatiotemporal patterns (fluid turbulence with tremendously rich temporal dynamics and eddies on many spatial scales is an attractor). We now have largely eliminated the use of “persistent” in the manuscript. On the other hand, for lack of a better word it’s fine to still use that term, if it is understood in a more general sense, which also includes stable representations in which the activity of individual neurons varies along the delay period (Goldman, 2009; Murray et al. 2017) or rhythmic activity which persists over time (Miller et al. 2018). The attractor network theory should be contrasted conceptually with mechanisms based on intrinsically transient memory traces (see Wang TINS 2021 for a more elaborated discussion on this).
Our proposal for distributed WM has a general aim and it’s not restricted to the classical ‘elevated constant firing’ scenario. Following the reviewer’s suggestion, we have rewritten the text to make sure that multiple mechanisms of WM are acknowledged in different parts of the text, not only on a paragraph in the discussion. We have also acknowledged the importance of thalamocortical interactions and cited previous relevant studies in this sense (such as Guo et al. 2017), also as a response to comments from Reviewer 1.
In addition, we have attempted to go beyond a simple rewriting and, using a variation of our simplified model, we now show that distributed WM representations can also happen in the context of activitysilent models (Figure 3 –figure supplement 1). In particular, we use a simplified network model with reduced local and long-range connectivity strength and incorporate short-term synaptic facilitation in synaptic projections. Our model results show that, while activity-silent memory traces can’t be maintained when areas are isolated from each other, inter-areal projections reinforce the synaptic efficacy levels and lead to a distributed representation via activity-silent mechanisms.
We hope that this result serves to prove the generality of our distributed WM framework, and opens the door to subsequent studies focusing not only on distributed activity-silent mechanisms, but in distributed frameworks relying on other WM mechanisms as well.
-
Evaluation Summary:
Mejias and Wang propose here the first large-scale model of the brain that actually performs a cognitive task. Previous models have focused on neural dynamics during the so-called "resting state", in which subjects are not performing any cognitive task - thus, resting. This study is therefore an important improvement in the field of large-scale modelling and will certainly become an influential reference for future modelling efforts.
(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 model proposed here is the first large-scale model that actually performs a cognitive task, which in this case is working memory but could easily extend to decision making in general as is acknowledged by the authors. Briefly, each of the 30 areas are simulated as a rate, Wong-Wang circuit (i.e. two excitatory pools inhibit each other through a third, inhibitory population). The authors use previously collected anatomical data to constrain the model and show qualitatively match with the data, in particular how mnemonic activity emerges somewhat abruptly along the brain hierarchy.
Strengths:
Previous models have focused on neural dynamics during the so-called "resting state", in which subjects are not performing any cognitive task - thus, resting. This study is therefore an important improvement in the …
Reviewer #1 (Public Review):
The model proposed here is the first large-scale model that actually performs a cognitive task, which in this case is working memory but could easily extend to decision making in general as is acknowledged by the authors. Briefly, each of the 30 areas are simulated as a rate, Wong-Wang circuit (i.e. two excitatory pools inhibit each other through a third, inhibitory population). The authors use previously collected anatomical data to constrain the model and show qualitatively match with the data, in particular how mnemonic activity emerges somewhat abruptly along the brain hierarchy.
Strengths:
Previous models have focused on neural dynamics during the so-called "resting state", in which subjects are not performing any cognitive task - thus, resting. This study is therefore an important improvement in the field of large-scale modelling and will certainly become an influential reference for future modelling efforts. As typically done in large-scale modelling, some anatomical data is used to constrain the model. The model shows several interesting characteristics, in particular how distributed working memory is more resilient to distractors and how the global attractors can be turned off by inhibition of only top areas.
Weaknesses:
Some of these results are not clear how they emerge, and some "biological constraints" do not seem to constrain. Moreover, some claims are slightly exaggerated, in particular how the model matches the data in the literature (which in some cases it does not) or how somatosensory working memory can be simulated by simply stimulating the "somatosensory cortex".
This paper has two different models, one being a simplified version of the main model. However, it is not very clear what the simplified model adds the main findings, if not to show that the empirical anatomical connectivity does not constrain the full model.
-
Reviewer #2 (Public Review):
There is a lot to like about this manuscript. It provides a large-scale model of a well-known phenomenon, the "delay activity" underlying working memory, our oldest and most enduring model of a cognitive function. The authors correctly state that despite the ubiquity of delay activity, there is little known about the macro and micro circuitry that produces it. The authors offer a computational model with testable hypotheses that is rooted in biology. I think this will be of interest to a wide variety of researchers just as delay activity is studied across a variety of animal models, brain systems, and behavior. It is also well-written.
My main concern is the authors may be self-handicapping the impact of their model by not taking into account newer observations about delay activity. For a number of years …
Reviewer #2 (Public Review):
There is a lot to like about this manuscript. It provides a large-scale model of a well-known phenomenon, the "delay activity" underlying working memory, our oldest and most enduring model of a cognitive function. The authors correctly state that despite the ubiquity of delay activity, there is little known about the macro and micro circuitry that produces it. The authors offer a computational model with testable hypotheses that is rooted in biology. I think this will be of interest to a wide variety of researchers just as delay activity is studied across a variety of animal models, brain systems, and behavior. It is also well-written.
My main concern is the authors may be self-handicapping the impact of their model by not taking into account newer observations about delay activity. For a number of years now, evidence has been building that working memory is more complicated than "persistent activity" alone. Stokes, Pasternak, Dehaene, Miller and others have been mounting considerable evidence for more complex dynamics and for "activity-silent" mechanisms where memories are briefly held in latent (non-active) forms between bouts of spiking. There is also mounting evidence that the thalamus plays a key role in working memory (and attention). In particular, higher thalamic nuclei are critical for regulating cortical feedback. Cortical feedback plays a central role in the model presented here. The model presented in this manuscript just deals with persistent attractor states and the cortex alone.
This is not to say that this manuscript does not have good value as is. No one disputes that some form of elevated, sustained, activity underlies working memory. This work adds insights into how that activity gets sustained and the role of, and interactions between, different cortical areas. The observation that the prefrontal and parietal cortex are more critical than other areas, that there are "hidden" attractor states, and "counterstream inhibitory bias" are important insights (and, importantly, testable). They will likely remain relevant even as the field is moving beyond persistent attractor states alone as the model for working memory. The new developments do not argue against the importance of delay activity in working memory. They show that it is more to the story, as inevitably happens in brain science.
The authors do include a paragraph in the Discussion referencing the newer developments. Kudos to them for that. However, it presented as "new stuff to address in the future". Well, that future is now. These "newer" developments have been mounting over the past 10 years. The worry here is that by relying so heavily on the older persistent attractor dynamics model and presenting it as the only model, the authors are putting an early expiration date on their work, at least in terms of how it will be received and disseminated.
-
-
-