Comms Review #2

[distillpub-reviewers]

The following communications-focused peer review was solicited as part of the Distill review process.

The reviewer chose to waive anonymity. Distill offers reviewers a choice between anonymous review and offering reviews under their name. Non-anonymous review allows reviewers to get credit for the service they offer to the community.

Distill is grateful to Karttikeya Mangalam for taking the time to review this article.

---

General Comments

"The article constructs and explains several early machine learning models from a physics based perspective such as energy models. It builds up gradually starting from simple Boltzmann models slowly to Restricted Boltzmann machines and to recent research works in QML and spin glasses.

From a communication perspective I found the following really shines through:

• The first interactive exhibit (effect of T on w) helps understand the underlying relationship and is well motivated,
• The gradual ram up in learning rule complexities make the reader build intuition slowly for more complex algorithms.
• The learning/unlearning visualization has clear exposition and motivates the reader to understand the content deeper.
• The QML subsection is optimistic and nudges the reader to checkout several listed works at the horizon of their understanding for cool new applications.

And the following areas can be improved further:

• Several key concepts or asserted results/facts in the ""Introduction"" are missing citations.
• Exhibit 2 showing different Boltzmann models can be made more interactive. Perhaps by adding simulation for a forward pass. Different models can also be brought under a single umbrella with interactive options for deleting/adding edges.
• The RBM sampling algorithm can be written for clearer understanding.
• More details on how the learning/unlearning visualization relates to the onging content would enhance understanding.
• In the ""Training BM"" exhibit, the 1D variables can be converted to sliders instead of circles since they only allow single axis changes.
• The corrupted data infilling visualization with RBM is cool. However, it can be better explained with the caption on top of article as well.
• The Spin glasses subsection feels very complicated and doesn't add much to overall topic understanding. It also feels out of place and more like a whirlwind literature review rather than building deeper insight for the reader. It can shifted after QML and edited for brevity with less focus on covering more paper and more focus on weaving a coherent story for the reader.

---

Distill employs a reviewer worksheet as a help for reviewers.

The first three parts of this worksheet ask reviewers to rate a submission along certain dimensions on a scale from 1 to 5. While the scale meaning is consistently "higher is better", please read the explanations for our expectations for each score—we do not expect even exceptionally good papers to receive a perfect score in every category, and expect most papers to be around a 3 in most categories.

Any concerns or conflicts of interest that you are aware of?: No known conflicts of interest

Outstanding Communication	Score
Article Structure	4/5
Writing Style	3/5
Diagram & Interface Style	4/5
Impact of diagrams / interfaces / tools for thought?	2/5
Readability	3/5

[PatrickHuembeli]

We would like to thank Karttikeya Mangalam for his comments and for reviewing our article. We address all his points below:

• We added citations to all the concepts that we introduce in the introduction.
• The aim of Figure 2 is to introduce the possible architecture of spin models and the two values (and the color scheme) of the nodes. We think that more interactivity would not really add any value and possibly even distract from the main purpose of the figure. All the figures (except the teaser) are chosen such that they do not anticipate more knowledge about the subject than we provide in the text upon arriving at the figure. The design of the figures was pretty much part of the whole writing process and in our opinion, it makes the most sense to leave it like this. I hope the reviewer agrees with us after knowing more about our process of thought.
• We added the RBM sampling algorithm in the form of pseudocode. We hope this is more or less what the reviewer imagined.
• We realized that the figure for learning and unlearning was not ideally positioned in the manuscript. We moved it further down so it is embedded where the learning and unlearning is described in the context of the positive and negative phase of the contrastive divergence. We added a few more explanations to the article and the caption and hope it is clear now.
• We agree that it is not necessary to use circles instead of sliders, because there is only a 1D movement possible. Nevertheless, we were thinking that it might be nicer to use a better way of visualizing the external field and this is why we used the background colors. Furthermore, these colors visualize the preferred spin direction when the node is placed in it, e.g. the whiter the background, the more the white spin direction is preferred. We are happy to change to sliders if the reviewer insists on changing it. In our opinion, we can convey more information with this approach than with a simple slider. In the end, we think it is a matter of taste.
• We understand that the figure needs more explanation. Since this figure at the top is just a teaser for what is coming in the article, we added a very generic explanation, avoiding technical terms. We are not sure if it is at all necessary to add a caption to the teaser. For example, this and this article don’t give any explanation about what is in the teaser. We hope this is sufficient. If not we can add more information.
• We agree that the Spin-Glass section was too detailed and complicated. As suggested, we moved it to the end of the article and shortened it. We did not delete any references, because we think it might be of interest for some readers to look up these concepts in more detail. The main purpose of this section was to show that there is still a lot of research going on in this direction and that the physics and ML communities can actually learn a lot from each other.