Iterated Insights

Below are anonymized peer-review and editorial responses to earlier versions of my iterative updating article (Reser, 2002). I thought it might be interesting for people, or machines really, to see. This model was my main focus for 15 years, the article was rejected for five years and I couldn’t publish it anywhere, but I didn’t think the reviewers engaged with the main theoretical points at all so I was very confused. I don’t believe I changed the article at all in response to any of the reviews (or anyone else’s input whatsoever). I decided to send it to Arxiv and then backed away from publishing completely.

A Cognitive Architecture for Machine Consciousness and Artificial Superintelligence: Updating Working Memory Iteratively, 2022 arXiv:2203.17255 [q-bio.NC]

xxxx

Cognitive Psychology

Your paper is interesting but it isn’t appropriate for Cognitive Psychology. We publish papers that advance theory, and that requires positing some new theory and testing it empirically, usually in new experiments that test novel predictions. New theories are often compared quantitatively with existing theories, addressing their ability to account for critical phenomena. Your paper doesn’t do this. You present a theory but you don’t evaluate it with the level of rigor we expect. To publish in this journal, you need to test your theory against data. You haven’t done that, so your paper isn’t appropriate for this journal.

I think most psychology journals are going to want to see novel predictions generated and tested and benchmark effects accounted for quantitatively. That would be the standard for Psychological Review and Psychonomic Bulletin & Review. Cognitive Science or Brain and Behavioral Sciences may publish something like this, so you might try them.

You might look at Mike Kahana’s work on the temporal context model, beginning with Howard and Kahana (2002, J Math Psych). It proposes an evolving context like the one you suggest, and it accounts for a lot of data in free recall tasks. I used that idea to explain sequential choices of keystrokes in skilled typing (Logan, 2018, Psych Review), essentially arguing that keystrokes are chosen in the context of the previous keystrokes that have been chosen so far.

I wish you luck in finding an outlet for this work.

xxxx

Ref: BBR_2019_809
Title: Thought is Structured by The Iterative Updating of Working Memory
Journal: Behavioural Brain Research

Dear Dr. Reser,

Thank you for submitting your manuscript to Behavioural Brain Research. I regret to inform you that reviewers have advised against publishing your manuscript, and we must therefore reject it.

Please refer to the comments listed at the end of this letter for details of why I reached this decision.

We appreciate your submitting your manuscript to this journal and for giving us the opportunity to consider your work.

Kind regards,

xxxx
Behavioural Brain Research

Comments from the editors and reviewers:
-Reviewer 1

– The idea of iterative WM updating is interesting, and was not addressed sufficiently in the literature. According to this idea, if I understand it correctly, WM updating is not discrete but relies on residue traces from previous states of the information. While I like this idea, and would like to see evidence showing the relationship between iterative updating and other aspects of updating (e.g., what is the role of removal in iterative vs. non-iterative updating? attentional selection?), I don’t think that this review is suitable for publication in its present form.

The review suffers from over-generalizations and clear errors in many places. For example, “…. which then decay from working memory over the course of seconds or minutes, some quicker than others” (p. 4): many (perhaps, most) theories of WM don’t assume temporal decay. Many of the arguments are too general, e.g. “iterative updating could be achieved using techniques such as multivoxel pattern analysis, which has an advanced capacity to visualize changes in the human brain over time”, or “conscious content may be cobbled together from unconscious content”. They lack details and references to support them. Some of the terms are vague (“methodical updating”). Also, important and relevant aspects of updating such as gating (see work by O’Reilly & Frank, Kessler & Oberauer) and removal (Oberauer, 2001, Lewis-Peacock et al., 2018) are overlooked in the review.

Two final points regarding the writing. First, the review discusses in great length work that is widely discussed elsewhere (work by Baddeley, Cowan, James, for only some examples), making it much longer than needed. Second, I suggest language/style editing – sentences such as “It is possible that this iterative property is not utilized in the brain for information processing; however, it is argued here that it is” or “After preparing an early draft of this article, the author came across a transcript…” are long, unnecessary, and make it very hard to read the paper.

-Reviewer 2

–

The manuscript describes theoretical considerations about the changing working memory states underlying human cognition. In particular, it proposes that working memory is updated incrementally over time, with several concepts remaining active from on point in time to the next and one or a few being updated or replaced, to give rise to sequences of thoughts, and thus cognition and reasoning. While I believe that this is an important and fascinating research topic, I unfortunately don’t see that this paper significantly expands our knowledge or understanding of working memory and cognition. It’s not that I object to anything proposed here – I think most of what is described here wouldn’t be particularly controversial. My criticism is rather that the manuscript in it’s entirety remains very abstract and vague in describing the proposed mechanisms. It does not provide clear steps towards a concrete implementation, nor make any concrete proposal how to test its predictions.

For example, one of the key components of the presented theory is that the updating of working memory is incremental, with only a subset of active concepts changing in every step. This is quite plausible, but the author doesn’t make a compelling argument why it wouldn’t work in a different way – e.g. by fully updating the whole working memory representation in each step (e.g. by association from long-term memory based on the current working memory content). Making a strong argument here would require a deeper level of analysis that considers how the update of working memory content could be realized and what the concrete advantages and problems with different implementations are. The author refers to certain classes of models to explain how some aspects of what he proposes could be realized, but it all remains very vague.

The author remarks that there is a lack of literature on this topic. I partly agree, but there still have been several approaches to model the sequences of activation states (and working memory states in particular) underlying cognition, in a much more concrete form than presented in this manuscript. For instance the SHRUTI model by Shastri and colleagues (e.g. Shastri 1999) describes inference and reasoning in a connectionist architecture, with a limited number of bound entities held in working memory states through sustained, synchronized neural activity, and iterative changes of these memory states while an inference is processed (which I believe comes quite close to what is hypothesized in this manuscript). There is the related work of Hummel and colleagues on inference and analogical mapping (e.g. Hummel & Holyoak 2003). Furthermore, there are models in the framework of dynamic field theory that e.g. incrementally build representations of visual scenes in working memory (in a kind of visuo-spatial sketchpad), either from sensory or from verbal description, and test the match of further spatial descriptions to this scene, shifting the focus of attention from one item to another and sequentially exploring different hypotheses about the mapping of a phrase to a scene (e.g. Richter et al. 2014, Kounatidou et al, 2018).

These works span the fields from psychology to AI and robotics, and are aimed at various level of neural realism and detail. What they have in common is that they provide a concrete realization of the mechanisms underlying the selection of concepts into working memory or the focus of attention (either from long-term memory or sensory input) and the iterative updating of these representations. They have to address problems like how to select an appropriate next item to be activated, and how to move forward toward some goal (rather than e.g. just running in circles). What they also have in common is that they are generally very complex models with a quite limited scope, which limits their impact in the research community. A generalization of such approaches would certainly be desirable, but it still needs to retain a certain level of specificity and detail in order to move the field forward.

References:

Shastri, L. (1999). Advances in Shruti—A neurally motivated model of relational knowledge representation and rapid inference using temporal synchrony. Applied Intelligence, 11(1), 79-108.

Hummel, J. E., & Holyoak, K. J. (2003). A symbolic-connectionist theory of relational inference and generalization. Psychological review, 110(2), 220.

Kounatidou, P., Richter, M., & Schöner, G.. (2018). A Neural Dynamic Architecture That Autonomously Builds Mental Models. In T. T. Rogers, Rau, M., Zhu, X., & Kalish, C. W. (Eds.), Proceedings of the 40th Annual Conference of the Cognitive Science Society (pp. 643–648).

Richter, M., Lins, J., Schneegans, S., Sandamirskaya, Y., & Schöner, G.. (2014). Autonomous Neural Dynamics to Test Hypotheses in a Model of Spatial Language. In P. Bello, Guarini, M., McShane, M., & Scassellati, B. (Eds.), Proceedings of the 36th Annual Conference of the Cognitive Science Society (pp. 2847–2852). Austin, TX: Cognitive Science Society.

Manuscript Number: BRCG_2019_317

Thought Is Structured by the Iterative Updating of Working Memory

Dear Dr Reser,

Thank you for submitting your manuscript to Brain and Cognition.

It has been difficult to find reviewers for this manuscript. A large number of invitations to review were sent out. A number of invitations were declined and others did not get a response. I have, however, received one competent review. I have also read the manuscript for continuity and find myself in agreement with the comments in the review.

We find the ideas presented in the paper of interest. But we also find that there are several assumptions that are not explicitly supported, that the involvement of the neural substrates with respect to the central issue developed here is, unfortunately, absent. You do mention neuronal activity and such but this is rather general and nonspecific. I would like to reiterate, as mentioned in the Aims and Scope of the journal, that contributors must address an aspect of the interaction between brain mechanisms (networks, structural or functional anatomy, processes) and cognitive function or behavior. Although the brain is often mentioned, the specific aspects of that involvement are not found here. In addition, the reviewer raises a number of concerns that seriously constrain the import of this submission.

In light of the comments in the review, I very much regret that we cannot consider the manuscript for publication. I hope that you will find this evaluation helpful in your future research endeavours.

We appreciate your submitting your manuscript to this journal and for giving us the opportunity to consider your work.

Kind regards,

xxxx

Brain and Cognition

Editor and Reviewer comments:

Reviewer #1: Review of BRCG_2019_317: “Thought is structured by the Iterative updating of working memory” by Jared Edward Reser

This theoretical paper argues for a new model of working memory that is based on the concept of iterative updating. My impression of the current paper is rather mixed: while I find the ideas presented here interesting, I also found that the whole framework appears to be built upon assumptions for which little to no empirical evidence exists as of yet. This, in combination with the fact that I found the paper to be lacking a clear focus, regrettably leads me to recommend rejection of this paper in its present form.

My main objections to the paper in its present form are the following. Firstly, the main argument for iterative updating forming the structure of thought appears to assume that iterative or recursive updating processes exist in the human brain. While I do not find this assumption to be implausible by itself, there is currently no direct neural measurement of iterative updating that I am aware of. This of course makes the foundation of the current paper rather shaky. For this reason, the conclusions of the current paper are not very well supported by the existing evidence.

Secondly, in addition to presenting new theoretical models of working memory, consciousness, and thought, it also appears to be a philosophical treatment of William James’ work in relation to current neuroscientific theory, and a recommendation for a novel computing architecture. Should the author wish to further expand the ideas presented in the current manuscript, I would strongly advise focusing on the neuroscience data and remove all the James quotes (the guy’s always good for a nice opening quote, but the current comparison feels forced and distracting).

Thirdly, I believe that the paper should be considerably shortened. Instead of starting by a complete historical treatment of memory research, it would be much more efficient when the paper would start with a brief outline of the proposed model, after which evidence (in favor, or against) this model would be integrated with a careful treatment of the detailed aspects of the model. Please note that the current introduction is not particularly inviting to continue reading.

Lastly, please take note of the formal requirements for submitting a paper to this journal. Specifically, the reference list is full of stylistic inconsistencies that could easily have been rectified before submission.

Signed,

xxxx

Ms. No.: NSY-D-20-00415

Title: Thought Is Structured by the Iterative Updating of Working Memory

Corresponding Author: Dr. Jared Edward Reser

Authors:

Dear Dr. Reser,

Thank you for your submission to Neuropsychologia.

As you may know, less than 25% of submissions to Neuropsychologia are accepted for publication. Your paper has received an initial review by myself and another member of the journal’s editorial team, and I am sorry to inform you that in our judgment the paper does not have sufficient potential impact to justify proceeding to full review. We both felt that the strengths of your submission lay more with cognitive considerations rather than with the links between neural and cognitive levels, which tends to be a focus for our readership.

Whilst I appreciate your disappointment at this decision, it will enable you to have the paper reviewed (and, hopefully, published) by another journal faster than would otherwise have been the case.

I hope that you will consider Neuropsychologia as a potential platform for your future studies.

Sincerely,

xxxxx

Neuropsychologia

Manuscript Number: PHB-D-20-00654

Thought Is Structured by the Iterative Updating of Working Memory

Dear Dr. Reser,

Thank you for submitting your manuscript to Physiology & Behavior.

I regret to inform you that the reviewers recommend against publishing your manuscript, and I must therefore reject it. My comments, and any reviewer comments, are below.

For alternative journals that may be more suitable for your manuscript, please refer to our Journal Finder (http://journalfinder.elsevier.com).

We appreciate you submitting your manuscript to Physiology & Behavior and thank you for giving us the opportunity to consider your work.

Kind regards,

xxxx

Physiology & Behavior

Editor and Reviewer comments:

Dear Authors

your paper cannot be published in PHB and I have to reject it. Please see comments below. A second Reviewer found the too long, non-focused, and too speculative, using outdated literature.

Thank you for your understanding.

Kind regards

xxxxx

Reviewer #1: This is a peculiar article. The value of this opinion piece depends on the targeted audience. If this is intended for cognitive psychologists or cognitive neuroscientists who study working memory, it is difficult to understand the value added of this think piece aside from the abundant material that has already been written on these topics. If instead the intended audience is a non-expert but scientific crowd, this could serve as a thoughtful and accurate historical primer on conceptualizations of working memory.

To improve impact, we recommend that the author clarify up front in the article WHOM this opinion paper is intended for (and for whom it may not be intended). Likewise, the author should clarify the intention of the paper – it is providing historical perspective and commentary on the evolution of working memory theory. No new ideas are being synthesized or proposed here, as far as we can tell. Instead, a collection of reasonable ideas about what working memory is and how it works are pieced together into a coherent and enjoyable-to-read narrative.

Here are some more specific comments.

1) The representative William James quote for each section highlights nicely how old these ideas are. It is nice to see updated references and models applied to these ideas. That said, we are concerned that the model descriptions provide anything other than a visual translation of James’ words. Often unclear as to what is the value added.

2) This manuscript would benefit from a marked shortening. There is a lot of redundant text. Other areas spend too much time explaining straightforward concepts (like neural ensemble activity). Many single sentences seem to reduce to “neurons fire and spread through the brain”, albeit in trickier phrasing. Perhaps a text reduction would also make the article more accessible.

3) We recommend the author drops the AI stuff altogether. This section feels even more speculative than the rest and would be best condensed into a single section, if kept at all. The contents are vague in that they amount to re-describing a neural system and suggesting that AI should ‘follow’ that structure (eg, “An analogue of the dopaminergic system’s network would be needed to recognize groupings of appetitively stimulating items and prioritize them by sustaining their activity” and “programmed instincts would be built into direct connections between sensory and motor areas”). This seems inherent in any proposed neural structure — ie, that it would be nice to have an artificial model of it. The proposal of these ideas is obvious, it is the implementation that eludes us, which isn’t proposed here.

4) Perhaps most importantly (but see our opening remarks about the intended audience), we are not convinced that anything being proposed by this model is novel. A stronger argument is needed as to how items entering/leaving WM continuously _isn’t_ the current thought. Don’t we already know that WM updates continuously? A clearly presented counterargument would be valuable. The paper claims on page 26 that the present model proposes that the engram for a WM item is a cell assembly. The loaded term “engram” aside, is the field not in agreement that a neural population activity is responsible for a WM representation? While this might be part of the current model (as it should be of any model), this model definitely is not “proposing” that idea. Same goes for “Reoccurring examples of coactivity would lead to the formation of heavily encoded associations (Asok et al., 2019), which would persist as procedural and semantic knowledge”, which is a fancier way of describing Hebbian

learning. Our takeaway is that updating as presented here is simply the excitatory and inhibitory effects of neurons that amount to spreading of information. The probabilistic aspect is simply that the current state of the brain influences the next. These are such obvious truths that they serve as the foundation of current neuroscience and don’t provide any real insight.

Behavioral Brain ResearchEditor and Reviewer comments:

We regret to tell you we were not able to find the required number of reviewers to evaluate your manuscript; we don’t want to delay the process for you any longer.

Reviewer #1: Major points:

1. My overall impression of the manuscript is that of a chapter in a Master’s/Doctoral thesis rather than that of a review/opinion article in a journal. This is the result of a number of observations:

a. The author spends an inordinate amount of time (i.e., 10 pages) to provide a historical overview of (some) of the more prominent theories of working memory (e.g., Atkinson and Shiffrin, 1968; Baddeley and Hitch, 1974; etc.) as well as some of the neural mechanisms having been proposed to underlie maintenance in working memory (e.g., persistent neural firing). While certainly presented in an engaging and educational manner, it is unclear to me what the relevance of such an in-depth review is for the current article. Clearly, the author’s main point of departure is Cowan’s process model of working memory, which is certainly prominent enough not to need such a detailed historical background. For this manuscript to be suitable for publication as a review article, I believe that it would require a lot of additional work to remove/reduce only tangentially relevant information and focus on the main idea.

b. By contrast, other sections clearly lack required detail – to the extent that the review appears superficial and misrepresents important research contributions. For instance, the section on synaptic potentiation (#5) does not clearly describe the activity-silent mechanisms in question and, importantly, ignores large parts of the literature, including from the pioneers of the phenomenon at hand (e.g., Stokes, 2015; etc.). Moreover, some of the statements made in this section (e.g., “and there is some strong evidence for it.”) are too general and not supported by relevant publications/empirical evidence. This is especially astonishing given that the neural basis of working memory is a current topic of active debate: While persistent neural firing has been considered the (sole) neural correlate of maintenance in working memory for many decades, increasing evidence suggests that this may only represent the tip of the iceberg, with “activity-silent” mechanisms playing a major

role as well. Some of the most important current directions in the field are thus not captured adequately in this review. At the very least, the author would have to present the literature in a way that acknowledges these areas of current debate.

c. When it comes to the author’s actual proposal (i.e., that thought processes are subtended by the partial updating of working memory representations), there is again a lack of substance in which this hypothesis is grounded. In fact, in the entire section in which the model is developed (i.e., section 6 – section 10), the author does not present a single piece of empirical evidence that might support his claim. To develop this theory, I think the author needs to make more of an effort to demonstrate (citing relevant empirical work) that it is more than just an interesting thought experiment based primarily on introspection.

2. It remains unclear to me how some of the literature cited in support of certain claims relate to the claim in question. For instance, to my knowledge, none of the four articles cited to support the claim that most discussions of working memory focus on complete updating of information (i.e., Pina et al., 2018, Niklaus et al., 2019, Miller et al., 2018, and Manohar et al., 2019) rather than partial updating actually discuss updating of information in working memory in a central manner. Perhaps I am missing something, in which case I would encourage the author to spell out the support more directly. In addition, it would be useful to provide direct experimental evidence for this claim, for instance, providing an overview of the paradigms typically used to study updating in working memory.

Minor points:

1. Some of the figures have clearly been adapted from previous publications (e.g., Fig.2), yet no citation and permission to reuse the figure has been provided. Please provide the missing information.

Manuscript Number: ACTPSY-D-21-00223

Thought Is Structured by the Iterative Updating of Working Memory

Dear Dr. Reser,

Thank you for submitting your manuscript to Acta Psychologica.

I regret to inform you that the reviewers recommend against publishing your manuscript, and I must therefore reject it. The reviewer comments are quite thorough and should be helpful in reformulating the ideas and the writing. Although both reviewers found much interesting in your paper, ultimately, it needs to be more grounded in existing theory, existing empirical data, and written with more focus.

For alternative journals that may be more suitable for your manuscript, please refer to our Journal Finder (http://journalfinder.elsevier.com).

We appreciate you submitting your manuscript to Acta Psychologica and thank you for giving us the opportunity to consider your work.

Kind regards,

xxx

Acta Psychologica

Editor and Reviewer comments:

Reviewer #1: The current manuscript outlines an information processing model of working memory. Specifically, the authors propose that information is updated in the different memory stores partially and iteratively. The selection of memory stores follows Cowan’s Embedded Processes Model; focus of attention (FoA), short-term memory (STM), and long-term memory (LTM). According to the model information is able to persist through different memory states by either sustained neuronal firing (for information in the FoA), or synaptic potentiation (for information in STM). Furthermore, iterative updating of working memory is key in higher order cognitive processes such as problem solving, logic, and reasoning. For example, problem solving is afforded by this model through the overlap of items between successive memory states, recursion, and temporarily suspension of information in STM. The model also suggests that new items enter the FoA through “multiassociative search”. Items in

the FoA coactivate other related items in STM or LTM in a Hebbian fashion. The coactivated items facilitate search for the next item(s) to be updated in the next state of the FoA. WM aids in the completion of tasks through multiple partially updated iterations of the different memory stores and multiassociative search.

Generally the manuscript is well written and does a good job of clarifying the main purpose of the manuscript; explaining how WM uses iterative updating to process information. The figures were helpful (but see comment #2), the quality of the references indicate proper literature review and support the logical reasoning behind the model. As a fellow WM researcher I became excited about the prospect of novel experiments testing this model and the benchmarks of WM through iterative updating.

I was less positive about the following: I thought that the topic of AI is out of place (see comment #1). The summary and conclusion section is unstructured, introduces multiple new lines of thought that would fit better in the main text, and needs a major overhaul.

My comments below outline why I currently can not recommend publication as is.

Major comments:

1. Although I recognize the quality of the work, the AI sections feel out of place in this manuscript. First, I am not sure if Acta Psychologica is a good fit for this content (but this is up to the editor) because it contains language and methods not widely known in the psychology community (e.g., hierarchical hidden Markov models, stochastic grammars, etc.).

Second, this is a quite specific application of the model before the model has gone through any testing or benchmarking. There are known experimental effects in/of WM that models and theories of WM should be able to explain (see Oberauer et al., 2018, the benchmarks paper). N.B., I am not suggesting that for the current manuscript. But perhaps jumping ahead to applications is a bit too overzealous.

2. The spheres used throughout the manuscript are color coded in table 1: FoA=white, STM=grey, LTM=black. In figures 8, 9, 10, 11, 12, 15, white spheres are used to indicate items currently in the FoA (which fits with the color coding), but black is used to indicate “inactive” items, or items outside of the FoA. However, do the black spheres are indicate information in LTM then? Or could they technically be in the STM? Or do the black spheres purely indicate being outside of the FoA/deactivated?

In figure 11, the colors are reversed; white resembles deactivated and black resembles co-activated.

Figure 19, has different shades of grey that are not explained. I have seen that Table 1 excludes figure 19, but could nonetheless cause confusion.

Overall: I think this could be streamlined and prevent some confusion.

3. The summary and conclusion section does not mainly serve to summarize or to conclude. It is messy, unstructured, which does not fit the rest of the manuscript.

The first item is a suggestion to use single cell to find iterative updating. This is new information, not a summary and not a conclusion. Would this fit better in the main text?

The second item is regarding the unclarity whether support can be found with current neuroimaging techniques. Why is this not before the first section? State the problem, and then offer a solution (single cell recording).

Table 3 just shows up without any introduction in the text in the middle of the summary and conclusion.

The rest of the section introduces again new information in the form of “consciousness”. There are two new figures; figure 19 is complex and needs more explanation (just not in the summary and conclusions section).

Perhaps move some of these paragraphs to the main text. Then in the summary and conclusions add some summary and conclusions. What could also fit well is some suggestions on how the model could be tested in light of existing knowledge about WM/STM (e.g., convergent/divergent validity). Perhaps a dedicated section to future research; this could house a shorter section on the AI application.

Minor comments:

1. In the note of figure 1. it states that “a thing” is modified. Perhaps find a better word for “thing”.

2. All page numbers are “4”.

3. In the summary and conclusions there is the sentence: (“It is intended to inspire more detailed hypotheses that can be tested experimentally.”). It would be nice for the reader to have this intention mentioned before the model is explained.

Reviewer #2: Overall, I found this ms to be creative and extremely interesting, and possibly more indicative of classic cognitive science than anything I have seen in the past 10 years. Unfortunately, however, it is not clear how novel the approach is or what progress is made with it beyond the current set of models of LTM and STM that we have. Even worse, I find that the scope of the paper is way too panoramic for a research journal of any sort. The material really belongs in a book or chapter, summarizing a very large number of more focused and data-driven papers in which the various aspects of the theory are more rigorously developed and tested, rather than just stated.

This is not to say that I object to theoretical papers, of course. We are in dire need of them. But this one just doesn’t quite do it for the reasons I mentioned.

Here are a few specifics:

How does this model differ from other iterative models, such as Howard and Kahana’s (2002) TCM model and its various expansions and instantiations?

What experimental data is the model describing and can we see some figures showing human data and model data for comparison?

What specific predictions does the model make about unseen data that other models do NOT make? I.e. what does this framework do that isn’t already being done?

Assumptions, such as all-or-none slots, should be defended rather than just picking a particular representation out of convenience.

Table 1: Some of the entries don’t seem correct (Repetition, Mnemonics is about LTM encoding, not maintenance; Irretrievability as the main departure is controversial). The author writes under the Table these are “active areas of debate,” of course, but that seems to raise questions about the whole enterprise if even the basic questions on which the model is built are, again, just kind of chosen for convenience.

Fig. 10, others: What is the role of similarity in all of this? For instance inter-item similarity (“homogeneity”) is an important factor in the formation of prototypes for at least some elementary visual STM stimuli (see, e.g., Dubé, 2019), and there is evidence these integrated representations are formed using vSTM contents (Zepp, Dubé, and Melcher, 2021).

The author likens the different updating percentages in Fig. 10 to System 1 and System 2 processes, however it is clear from the history of the field, going back to Von Neuman’s book “The Computer and the Brain” and through Chomsky, up to the present, that the cleavage is between an algorithmic rule-based system that cannot be well described using mathematics, but rather computer simulation and logic, and a “lower” system that is within the domain of mathematics and follows a psychophysics or physics of the nervous system reasonably closely. In other words, the Systems are exactly NOT different degrees of the same fundamental operation, but are fundamentally different forms of calculus, one logical one mathematical. These of course cannot be reduced to one another, as we know from Goedel’s Incompleteness and all the fallout from Russell and Whitehead’s attempt to demonstrate strict isomorphism between the systems.

What directs the search described in Fig. 11 and the text?

How are expectations modeled in this system? Stephen Grossberg spent an entire career mathematically modeling and explaining the central role of prediction and expectation in STM and attentional systems, from a neural network formalism. But this work isn’t even mentioned.

Cognitive Systems Research

Editor and Reviewer comments:

Although your manuscript falls within the aim and scope of this journal, it is being declined due to lack of sufficient novelty. We receive a much larger number of papers than we are able to accept.

Reviewer #1: This article proposed a new theory about an iterative working memory conceptual model. Problem is that this theory is based on (partially) outdated litterature and over simplifications. There are a lot of ill-defined concepts and some problematic statements as well. Furthermore, the proposed theory is extremely similar to the Adaptive resonance theory (Grossberg, 1987) where the vigilance parameter would correspond to the FoA proposed by the author. I’m not saying models are identical, but this ART model would need at least to be discussed in light of author’s proposal. For the actual implementation, I woudl advise author to read “Sepp Hochreiter; Jürgen Schmidhuber (1997). “Long short-term memory”. Neural Computation. 9 (8): 1735-1780.”

Furthermore, the linear structure of the paper makes it difficult to read: each section introduces a new idea without clear relation to others and/or the litterature. I would advise the author to deeply restructure the paper and to have a thorough review of the litterature in both Cognitive Science, Computational Neurosciences and Machine Learning.

Finally, starting from section 9, the reader is invited to imagine how this model could work but is not offered any actual implemntation. Unfortunately, it is all too easy to imagine how a model could work, not taking into account the myriad of unforeseen side effects.

I do not recommend this paper for publication as it is highly speculative without strong evidences and not supported by a model.

** Introduction

Comparison between iterative working memory and iterative design is dubious at least. I think the adaptive resonance theory would be more relevant in the context of memory. The “search by spreading activation” corresponds actually to a dynamical system that settle onto some patterns (e.g. attractors, cycles) after having interacted neighbouring structures. This is the case for dynamical neural networks excahnge informations.

** Interactions Between Sensory Memory, Working Memory, and Long-term Memory

I think author need to review recent litterature in neuroscience and computational neuroscience about working memory models. For example, the update mechanism corresponds to the gated working memory idead where information can be selectively updated based on a specific signal (the gate).

** The Focus of Attention is Embedded within the Short-term Memory Store

The “Focus of Attetion” is only referred without a tentative explanation of its origin and implementation. This would be required or else, it appears at some kind of homunculus

** Sustained Firing Maintains Information in the Focus of Attention

There are now several evidences that working memory does not correspond to sustained activity but is encoded at the population level using complex dynamics (see for example “Reconciling persistent and dynamic hypotheses of working memory coding in prefrontal cortex”, 2018; “Neuronal population coding of parametric working memory”, 2010).

** Synaptic Potentiation Maintains Information in the Short-term Store

The “Synaptic Theory of Working Memory” (2008) is only a theory without yet strogn evidences. The cited paper (Silvanto, 2017) does not provide strong evidence as claims in the article since it is mostly a review of relevant works.

The table page 8 neds to be clarified and justified. How this table was built ? Does the provided reference in last column give the all the information on the same line ? What does a “extremly large” long term memory capacity means exactly ? What author designate by “cerebral cortex” as opposed to sensory cortex or association cortex ? How do we compare “numerous features” to “1 to 9 items”? What is the difference between “pre-attentive and “recent attention”? Also, why is the “focus of attention” considered to be a form of memory?

** Persistent Activity Causes Successive States to Overlap

“quite simple reasoning strongly suggests” is not an argument. Either you have experimental evidences showing that a given sustained firing pattern has formed gradually, or you need to express this statement is only an educated guess not backed up by experimental data. Furthemore, as mentioned earlier, the theory of sustained activity in the PFC is now controversial (Rigotti, M., Barak, O., Warden, M. R., Wang, X.-J., Daw, N. D., Miller, E. K., & Fusi, S. (2013). The importance of mixed selectivity in complex cognitive tasks. Nature, 497 (7451), 585-590; Machens, C. K., Romo, R., & Brody, C. D. (2010). Functional, but not anatomical, separation of “what” and “when” in prefrontal cortex. Journal of Neuroscience, 30 (1), 350-360. doi:10.1523/jneurosci.3276-09.2010)

Figure 6 needs to be reworked. If horizontal axis represents time, what is the semantic of vertical axis? Also, this is not a Venn Diagramm unless you author explains the semantic of overlapping region. Overall, I don’t understand the usage of discs to represent FOA and STM and relation with time.

** Iteration Causes Consecutive States of Working Memory to Be Interrelated

“If updates to the working memory store involve partial rather than complete replacement, then these dynamics indicate an ongoing pattern of recursion and iteration.” This is a logical fallacy. Why would a partial updating involves a recursion? Furthermore, author seems to diregard incremental updating (as opposed to partial or complete). If on is given a list of number to memorize such as a phone number, we can hypothesize the corresponding working memory to “grow” in order to memory numbers when they are processed (heard).

** Iterative Updating of Items Creates Narrative Continuity

What is a “psychological item”?

** Remaining secions

I stopped my review here beause remainign sections are an invitation to the reader to imagine how this could work and it is too speculative.

** Artificial Intelligence Should Employ Iterative Updating

Long Short-Term Memories (LSTM) do exactly that.

Dear Dr. Reser:

I write to you regarding manuscript # BRB3-2022-03-0296 entitled “The Iterative Updating of Persistent Activity May Provide Functional Structure to Working Memory, Thought, and Consciousness” which you submitted to Brain and Behavior.

In view of the criticisms of the reviewer(s) found at the bottom of this letter, your manuscript has been denied publication in Brain and Behavior.

Wiley Editing Services Available to All Authors

Should you be interested, Wiley Editing Services offers expert help with manuscript, language, and format editing, along with other article preparation services. You can learn more about this service option at http://www.wileyauthors.com/eeo/preparation. You can also check out Wiley’s collection of free article preparation resources for general guidance about writing and preparing your manuscript at http://www.wileyauthors.com/eeo/prepresources.

Thank you for considering Brain and Behavior for the publication of your research. I hope the outcome of this specific submission will not discourage you from the submission of future manuscripts.

Sincerely,

xxxx

Reviewer(s)’ Comments to Author:

Reviewer: 1

Comments to the Author

The manuscript proposes a theoretical framework for the incremental updating of working memory contents, suggests possible neurophysiological substrates and recommends an Artificial Neural Network architecture implementing the concept of iterative updating. Overall, the article is clear, well written and develops interesting ideas. However, the article is not suitable for publication due to several significant concerns:

1) As mentioned by the author, it has long been assumed that working memory is supported by the sustained activity of single neurons. However, in light of recent studies this assumption has come into question (e.g. Lundqvist et al, J. Neurosci., 2018, Miller et al, Neuron, 2018). Even the cited review by D’Esposito & Postle, 2015 criticize the idea that persistent neural activity is necessary to maintain representations in working memory. Other studies have emphasized the idea that it is not necessarily the sustained activity of single neurons that accommodate working memory contents but instead dynamic activity states across the neuronal population (Stokes, Trends Cogn Sci., 2015). Such state-related changes need not be restricted locally but can involve anatomically distributed patterns of interaction (e.g. Antzoulatos and Miller, eLife, 2016; Jacob et al, Neuron, 2018). Although the author acknowledges that this is an area of active debate and research, he does not discuss recent developments in the field and heavily bases his framework on sustained activity (and short-term potentiation) without discussing alternative possibilities and their implications in the proposed framework.

2) The author bases his framework on the link between the focus of attention (FoA) and working memory. Although there is a broad consensus about the relationship between attention and working memory, it is not entirely understood how these processes are linked (Oberauer, Journal of Cognition, 2019) or what are the benefits of employing attention during working memory (Heuer & Schubö, PLoS ONE 2016). For example, previous studies either conclude that attention serves to reduce memory load or that attention is employed to protect focused items. These concerns lie at the core of the ideas explored in the submitted manuscript and should have been discussed.

3) In Figs. 12, 20 the author proposes that the representation of an active item in working memory can be traced back to the activation of a distinct subset of neurons. Although this is a compelling idea, recent work has suggesting more complex processes and interactions including dynamic population codes and low-dimensional subspace representations (e.g. Panichello and Buschman, Nature, 2021; Parthasarathy et al, Nature Communications, 2018). These studies are relevant to the ideas presented in the manuscript and should have been at least discussed.

4) In pg. 15 the author states: “In animals, a lower percentage of iterative updating might be correlated with greater working memory capacity as well as higher fluid and general intelligence”. Are there any studies in support of this claim? For example, there is work suggesting a link between the updating of working memory and fluid intelligence, however the issue is far from settled (Ecker et al 2010).

5) The ideas summarized in Figs. 10, 11 are highly speculative, without any reference to experimental or previous theoretical work.

6) Finally, the author attempts a link between working memory and consciousness. His approach, however, does take into account recent experimental and theoretical work (e.g. Trubutschek et al, eLife, 2018).

Dear Dr. Reser,

Thank you for the opportunity to read your manuscript on the nature of thought, and how this might arise from a process of iterative updating in the focus of attention. The ideas in this manuscript are interesting, and I enjoyed reading it. However, I am rejecting it for publication in Computational Brain and Beahvior. The theoretical framework developed in the manuscript shows potential, but it is does not yet make close contact with empirical data. It also does not yet quite address the “computational” side of the quantitative theory-building effort, which makes it difficult for me to see ahead and figure how the model will be rigorously evaluated against data in the future. Without these elements, the scope of this journal does not align well with the manuscript.

I would like to thank you very much for forwarding your manuscript to us for consideration and wish you every success in finding an alternative place of publication.

With kind regards,

xxx

Computational Brain & Behavior

xxx

The Article Was Rejected By:

Trends in Cognitive Science (21)

Cognition (3.6)

Neuroscience and Biobehavioral Reviews (8.0)

Neural Networks (5)

Neurobiology of Learning and Memory (3.6)

Behavioral Brain Research (3.0) (peer review)

Brain and Cognition (2.6) (peer review)

Neuropsychologia (2.6)

Cognitive Psychology (3.0)

Consciousness and Cognition (2.0)

Physiology and Behavior (3.2) (peer review)

Acta Psychologica (1.4)

Cognitive Systems Research (3.5)

Brain and Behavior (2.7)

Computational Brain and Behavior