PNAS rejects my paper - some thoughts about it

Post date: Feb 14, 2018 7:57:49 PM

73 days after submission, PNAS rejects the paper from the previous post. That paper was long in the making, and I had had basic ideas on how to provide an exact connection to optimization and spin models for a while. I disagree with Referee 1, which woke up in a bad mood I sense, but I see his points. He also missed the fact that I propose algorithms to simulate these optimization algos on a computer, which clearly is not for a small set of researchers, which is why I had thought of sending it to PNAS.

I use this opportunity to sort of say why I believe that paper is important. Although it is "not surprising" that there is a connection between Memristors and Spin Glasses (I agree, that's why I wanted to find it), I yet had to see either a proof or somebody who made a connection as close to precise as I did (although in a toy model). Toy models are good because they simplify life.

Lately, I've found that whenever one referee rejects a paper of mine, tends to cite an experimental paper published in one journal of the Major league (see below). I'm not saying the paper is perfect, because the referee is right: there are approximations and plenty. Yet, it seems clear that these days it is hard for a theorist to make a breaktrough in one big journal, in particular if you're honest about it.

Last year Kosterlitz, Thouless and Haldane won the Nobel prize. Kosterlitz and Thouless published their first article in Journal of Physics C: Solid State Physics. It is true that important paper do not necessarily go into "important journals", but technical ones. But unfortunately. I started to learn (on my skin) that researchers nowadays have to have citations, publications in good journals, and an ability to take punches (and give a few).

I'm still considering opening my climbing gym in South Italy as a valuable option for career development.

With the following motivations:

Title: "Asymptotic behavior of memristive circuits and combinatorial optimization"

Tracking #: 2017-20268

Authors: Caravelli

Dear Dr. Caravelli,

Thank you very much for submitting your manuscript to PNAS. We apologize for the delay in assessing your manuscript. I am sorry to say that the PNAS Editorial Board has declined your manuscript [MS# 2017-20268] for publication. We receive more than 16,000 submissions each year and currently accept less than 20%.

The expert who served as the editor for your paper obtained 2 reviews, included below. After careful consideration, the editor decided that we cannot accept your manuscript. Of course, this terminates your PNAS License to Publish and you are free to publish elsewhere.

I am sorry we cannot be more encouraging at this time, and I hope you will consider submitting future work to PNAS.

Sincerely yours,

Interim Editor-in-Chief

*********************

Editor Comments:

The reviews indicated that the paper does not meet PNAS's publication criteria of making exceptional scientific importance to a broad scientific audience. Specifically, (1) the primary target for this paper would be a relatively small group of researchers, (2) the paper focuses on mathematical manipulations while lacks physical insight for its connections to physical implementation, and (3) the paper is not very well written.

Therefore, I would suggest the authors to consider submitting it to a more targeted journal.

Reviewer #1:

Suitable Quality?: No

Sufficient General Interest?: No

Conclusions Justified?: Yes

Clearly Written?: No

Procedures Described?: Yes

Comments:

Review of "Asymptotic Behavior of Memristive Circuits and Combinatorial Optimization"

by Francesco Caravelli

I do not recommend this paper for publication in PNAS. The primary requirement for papers submitted to PNAS is stated in the information for authors:

"Accepted papers must be of exceptional scientific importance and intelligible to a broad scientific audience."

This paper does not satisfy either criterion. I hate it when a reviewer simply says something like this for a paper that I have submitted without saying why, so I will explain my reasons for the statement.

First, I do not find anything surprising in the paper. The fact that a collection of memristors can be configured such that some type of equation can be mapped onto it has been assumed for a long time, and many papers have been written based on this assumption, for example by the authors of Refs. 2 and 3 in the manuscript. This relatively small group of researchers are the primary target for this paper. It is also not surprising that the dynamical equations for a collection of memristors is isomorphous to a physical system such as a spin glass.

There have been several papers published that provide examples of computing with memristors, for example Ref. 21 of the manuscript and the more recent paper by Du et al., in Nature Communications 8, 2204 (2017); DOI: 10.1038/S41467-017-02337-y. The present paper looks at random topologies instead of regular crossbars, but then does not address the practical problem of how to read out the answer, for which some type of CMOS circuitry is required, as well as for physically setting up the initial conditions and boundary values for the system.

Given the lack of surprise, I don't find the paper to have 'exceptional scientific importance'.

Second, this paper is primarily a set of mathematical manipulations with very little physical insight or explanation for how the equations map onto a circuit that can be built. I am going to assume that the math is correct, because as I said I am not surprised by the result and therefor I am not motivated to grind through the equations to make sure there were no errors. Also, as stated in the manuscript the author used heuristics and asymptotic assumptions anyway, so there is no exact result to verify. Given that I consider myself to be an expert in this field, I don't see what would motivate the 'average reader' of PNAS to read this paper for understanding. I am not biased against mathematics - I have read and used applied math published in PNAS papers in my research such as represented by Ref. 17 of the manuscript. I just don't find anything compelling in this case.

Finally, the paper is not very well written. It needs a level of editing to correct various typos, grammatical errors and awkward English usage that detract from understanding the manuscript.

In conclusion, I think that this paper is likely correct (but I have not verified that) and interesting to a small group of researchers. I would recommend that the paper be submitted to a statistical physics journal in roughly its present form after editing where the intended audience would be more likely to check the math or to an IEEE or computation journal if more detail on input-output were to be added.

Reviewer #2:

Suitable Quality?: Yes

Sufficient General Interest?: Yes

Conclusions Justified?: Yes

Clearly Written?: No

Procedures Described?: Yes

Comments:

This manuscript describes a study of nonlinear dynamical memristive circuits, their asymptotic behavior and capability to perform combinatorial optimization. A simple model for the memristors was used that includes drift and decay terms. A general circuit equation was developed for networks of such memristors, and most of the paper evaluated and explored a Lyapunov function for this network.

The work is timely, as memristive circuits have received increased attention for computational applications, not just the earlier memory applications. There are interesting suggestions that such devices capture the essential features needed to replicate biological learning, and have capabilities to solve hard computational problems as well. The study performed here involves a novel examination of the complexity of random memristive circuits and how such networks can converge on locally minimal energy points similar to simulated annealing, but with higher efficiency.

The final section shows an interesting mapping between the Lyapunov function and the Markowitz function, applied to the Nikkei asset portfolio. A remarkable result is shown in the rate of return versus classical annealing algorithms (Metropolis).

Overall, the work is novel and may have important implications for computing and future hardware approaches distinct from both the current CMOS paradigm, and even approaches taken in quantum computing.