The Machine Learning community is currently experiencing a reproducibility crisis and a reviewing crisis [Littman, 2021]. Because of the highly competitive and noisy reviewing process of ML conferences [Tran et al., 2020], researchers have an incentive to oversell their results, slowing down the progress and diminishing the integrity of the scientific community. Moreover with the growing number of papers published and submitted at the main ML conferences [Lin et al., 2020], it has become more challenging to keep track of the latest advances in the field.
Blog posts are becoming an increasingly popular and useful way to talk about science [Brown and Woolston, 2018]. They offer substantial value to the scientific community by providing a flexible platform to foster open, human, and transparent discussions about new insights or limitations of a scientific publication. However, because they are not as recognized as standard scientific publications, only a minority of researchers manage to maintain an active blog and get visibility for their efforts. Many are well-established researchers (Francis Bach, Ben Recht, Ferenc Huszár, Lilian Weng) or big corporations that leverage entire teams of graphic designers designer and writers to polish their blogs (Facebook AI, Google AI, DeepMind, OpenAI). As a result, the incentives for writing scientific blog posts are largely personal; it is unreasonable to expect a significant portion of the machine learning community to contribute to such an initiative when everyone is trying to establish themselves through publications.
Our goal is to create a formal call for blog posts at ICLR to incentivize and reward researchers to review past work and summarize the outcomes, develop new intuitions, or highlight some shortcomings. A very influential initiative of this kind happened after the second world war in France. Because of the lack of up-to-date textbooks, a collective of mathematicians under the pseudonym Nicolas Bourbaki [Halmos 1957], decided to start a series of textbooks about the foundations of mathematics [Bourbaki, 1939]. In the same vein, we aim at providing a new way to summarize scientific knowledge in the ML community.
Due to the large diversity of topics that can be discussed in a blog post, we decided to restrict the range of topics for this call for blog posts. We identified that the blog posts that would bring to most value to the community and the conference would be posts that distill and discuss previously published papers.
The format and process for this blog post track is:
Write a post about a paper previously published at ICLR, with the constraint that one cannot write a blog post on work that they have a conflict of interest with. This implies that one cannot review their own work, or work originating from their institution or company. We want to foster productive discussion about ideas, and prevent posts that intentionally aim to help or hurt individuals or institutions.
Blogs will be peer-reviewed (double-blind, see Section 2.5) for quality and novelty of the content: clarity and pedagogy of the exposition, new theoretical or practical insights, reproduction/extension of experiments, etc.
We believe that a formal blog post conference track would increase the posts’ visibility, impact, and credibility, while simultaneously providing benefits to the conference.
Adoption: we think that a conference’s “stamp” will give more credibility to blog posts, making them more broadly recognized and adopted by the community.
Accessibility: maintaining a blog is time consuming , and requires many blog posts to gain visibility and recognition. By allowing researchers to publish a single post, we will enable occasional blog writers to publish their ideas to a much broader audience - something that is relatively impossible right now. Moreover, it will make this format accessible to independent/junior blog writers that do not have a company or a research lab to support them.
Synchronization: the fast evolving field of ML advances at the paces of its conferences. Following the same pace the blog posts will add value and momentum to the conference. It will benefit from the same advantages of conferences with respect to scientific journals: faster publication process and cross-fertilization of ideas.
We develop the potential positive impact of a blog post track for the conference itself:
Increases the value of the papers submitted to ICLR: blog posts will discuss previously published papers, thus increasing their visibility and quality.
Incentivizes researchers to submit their best research to ICLR: high quality work will likely get highlighted in future years in a blog post.
Improves reproducibility and transparency: the blog post track will identify and publicly document pitfalls and “tricks” that were not clearly communicated in the original publication.
Provides a scientific value by itself: such blog posts will reproduce and extend results of previously published papers. They will distill important theoretical and practical ideas improving their adoption and impact.
Tests of time: this track will provide a sort of crowd-sourced test of time at a shorter timescale than the current test of times awards.
Promotes accessibility: because many of this track’s blog posts will vulgarize past content, this track will make the conference broadly more accessible (to students, non-natives, and, more generally, non-experts in the field).
In this section we identify potential issues arising with such a track and explain how to mitigate them:
Adversarial Blog Posts: Since the guidelines are to write a blog post on a previously published paper, one may expect some researcher to try to use bad faith arguments to criticize a concurrent paper through one of these blog post. We do not think this will happen, because these blog posts will be public and thus researchers would discredit themselves by using bad faith arguments.
Too many/few submissions: As this is a new track, it may be difficult to predict the volume of submissions. The fact that there are currently many independent blog posts on the web is a good indicator that there will be positive interest. To get a better estimate of the volume of potential submissions, we intend to leverage social media to gauge the interest of the ML community in such a track; this will allow us to gather a large enough reviewing committee.
Reviewing: Once again as this is a new track, it may be unclear how to judge blog posts during a review process. We will recruit a large reviewing committee and define clear guidelines for the reviewing process. Our primary focus will be on the originality of the perspective and the novelty of the ideas, insights, and experiments. For instance, posts that reuse less content from the original paper (results, direct quotes) will be scored more favourably than those that use more.
Too many posts on the same paper: We may mitigate this by only selecting a small numbers of blog posts on the same paper. This could actually be a strength since this can encourage discussion and highlight different perspectives on the same work. Moreover, we could explicitly state that we will have this hard limit (e.g., accepting a maximum of 3 blog posts on the same paper) to entice researchers to submit blog posts on papers that have less visibility.
Created in 2016, Distill is an online scientific journal based on blog post publications. We address our differences with respect to Distill:
Visualizations: Blog posts should take advantage of the fact that they’re not paperbound, and use innovative visualisations. But the process of creating the intricate, dynamic visualisations associated with Distill posts is a daunting for most authors. Creating blog posts should be more easily accessible to newer authors and researchers. Sometimes, being able to embed videos and gifs is enough.
Content: Distill does not target the same type of content as our track. Distill aims at presenting new research, and at making this research more accessible. We want our blog post track to incentivize researchers to revisit and discuss on other researcher’s works, in a more natural way than scientific papers allow. Such a practice would undoubtedly be useful for the community, both as a short-term “test of time”, and also as a way to extract the key ideas from lengthy articles.
Limited adoption by the community: we believe that since Distill is not associated with a big conference track, its widespread adoption is hindered. This lack of association confines it to a small subset of the community that is already familiar with blog posts.
Leveraging the momentum of the conference: Distill describes itself as a scientific journal. A large amount of the publications in the ML community are conference papers. A blog post track that follows conferences would be better suited to follow the pace of the community.
A recurrent workshop in the ML community is the ML Retrospectives Workshop (NeurIPS 2019, 2020 and ICML 2020). This workshop is a venue for researchers to talk about their previous work in a more open and transparent way. More precisely, emphasis has recently been put on addressing:
Flaws or mistakes in the paper’s methodology
Limitations in the applicability of the work
Changes in understanding or intuition
We share the ultimate goal of “making research more human”, but with a completely different format. We believe that the constraint to write about someone else’s work using natural language will channel fruitful discussions and provide more visibility to previously published papers.
We believe that our proposed blog post track differentiates itself from a tutorial track because tutorials operate at different scales. On the one hand, a tutorial regarding a whole topic (e.g. GANs, adversarial examples, Random matrix theory in ML) contains a long talk, slides, and potentially exercises to get familiar with the topics. It is usually made by a team of expert researchers on the topic. On the other hand, the call for blog posts we propose focuses on a single publication. It regards a single paper that can concern a more precise and recent topic (e.g., a specific paper that addresses mode collapse on GANs, a novel technique to perform adversarial training, etc.) and could be written by a single researcher (once again making it more accessible to junior researchers).
Recently, the Rethinking ML Papers Workshop at ICLR 2021 fuelled the discussion (see references therein for related past workshops). The presenters discussed the importance of accessibility, web demonstrations, visualization and blog posts (among others). One particularly related discussion was the talk by Lilian Weng (time=4h25mins) on the usefulness of blog posts to get up-to-date with the field of ML.
In alignment with these initiatives, this new track is another step in the direction of making research more human.
David Tran, Alex Valtchanov, Keshav Ganapathy, Raymond Feng, Eric Slud, Micah Goldblum, and Tom Goldstein. An open review of openreview: A critical analysis of the machine learning conference review process. arXiv, 2020.
Hsuan-Tien Lin, Maria-Florina Balcan, Raia Hadsell, and Marc’Aurelio Ranzato. What we learned from neurips2020 reviewing process. Medium https://medium.com/@NeurIPSConf/what-we-learned-from-neurips-2020-reviewing-process-e24549eea38f, 2020.