The current issue of Computing in Science and Engineering (CiSE) is a special issue on reproducible research, edited by two pioneers in the field: Jon Claerbout and Sergey Fomel. They have assembled a great set of articles from experts with a lot of first-hand, personal reproducible research experience, so I would highly recommend this to my colleague researchers!
I got a pointer earlier this week to a New York Times article about R. A very interesting article about the use of R in scientific communities and industrial research, mainly for statistical analysis. R is open source software, so it is free and has already taken advantage from contributions made by various authors. And (although I haven’t used it myself yet), it is a great tool for reproducible research. Using the package Sweave, authors can write a single document containing their article and the R code to reproduce the results and put them in place. This ensures that all the material is in a single place.
It also shows something about the amazing power of open source software developed by a community of authors (and typically users at the same time).
I seem to be dwelling quite some time on the web lately… After my post about the lifetime of URLs, here’s one about domain names and reproducibility. I recently noticed when looking around that there are quite some websites and domain names related to reproducible research.
reproducibleresearch.org is an overview website by John D. Cook containing links to reproducible research projects, articles about the topics, and relevant tools. It also contains a blog about reproducible ideas.
reproducibleresearch.com is owned by the people at Blue Reference, who created Inference for Office, a commercial tool to perform reproducible research from within Microsoft Office.
reproducibility.org is used by Sergey Fomel and his colleagues as home for their Madagascar open source package for reproducible research experiments.
reproducible.org is a reproducible research archive maintained by R. Peng at Johns Hopkins School, where the goal is to host a place for reproducible research packages.
Quite a range of domain names containing the word “reproducible” (or a derivative), if you ask me! And then I didn’t even start about the Open Research or Research 2.0 sites. Let’s hope this also means that research itself will soon see a big boost in reproducibility!
I am getting worried these days about the volatility of URLs and web pages. I guess you all know the problem: it is very easy to create a web page, and hence many people do so. Great! However, after some years, only few of those web pages are still available. Common reasons include people retiring, or moving to other places, and therefore their web pages at their employer’s site disappear. Similarly, registering a domain name at some point in time does not mean you will keep on paying the yearly fees forever. Or also, web sites getting an entire re-design often result in broken URLs.
Why does this worry me so much?
Last week, I attended the Berlin 6 Open Access Conference in Düsseldorf (Germany). It was an interesting conference, on different aspects of Open Access: making publications freely available online. There was a wide variety of talks, from publishers’ perspectives over financial models for Open Access and open standards, to benefits of Open Access for developed and developing countries.
One of the sessions was organized by Mark Liberman around the topic of reproducible research. I gave a talk there about my experiences with reproducible research, but that’s not what I want to talk about here. I found it very interesting to see the wide range of subjects and perspectives that Mark gathered in that session. Slides of the entire session are available here for those who are interested.
Reproducible research, literate programming, open science, and science 2.0. All different namings, and (in my opinion) all covering largely the same topic: sharing code and/or data complementing a publication as a presentation of your research work. While literate programming is more focused on adding documentation to code, and science 2.0 seems to include the assumption that you put work in progress online, there really seems to be a very large intersection between these topics.
This clearly shows that from various sides of the scientific community, in very different fields of science, the same ideas pop up. That is a really exciting thing! And at the same time it also shows that there is a clear need for such open publication of a piece of research. And I think everyone will agree that there would be nothing nicer than being able to really start from the current state-of-the-art when starting to do research in a certain field?
Should all these efforts be merged under a single “label”? It would definitely be exciting. And it would create a huge impact, as a joint effort for “open science”, “reproducible research”, or whatever the name may be, would receive a lot of attention, and cannot be overlooked by anyone anymore. At the same time, every research domain needs other specifics or finetuning, and it is not clear to me now what the “best” setup would be for the type of work I am doing now. So maybe we should let these variations co-exist for some more time, and see later which ones survive, are the simplest to use, and which tools can be combined to create an optimal method for research.
But of course (if anyone is reading these posts), I would be very happy to hear your own opinion on this!
A few months ago, I read in a Belgian newspaper that 9% of the participants in a study among 2.000 American scientists said they had witnessed scientific fraud within the past three years. And it seems they were not talking about those cases where people use Photoshop to crop an image or so, but rather inventing fake results or falsifying articles.
Although I wasn’t able to find this back on the web with Google, I am quite sure the original authors checked the number. Wikipedia reports on another study, where the actual number was 3%. Anyhow, whether it is 3 or 9 percent, this number is much too high. Let us hope it can be taken down by requiring higher reproducibility of our research work. I do realize that there will always be people cheating, and falsifying results (Wikipedia even keeps a list of the most famous cases). But I also strongly believe that in the end, most researchers just want to do good work. And many of them perform non-reproducible work, just because they don’t feel the need for making it reproducible (yet). Or are too busy with their next piece of work to properly finish off the current one…
One of the reproducibility problems with many current papers is that everyone applies his new algorithm to his own set of data. So did I in my super-resolution work, too. A problem with that is that it is very difficult to assess whether the data set is used (a) because that was the one the author had at hand, (b) because it was the most representative one, or (c) because the algorithm performed best on that data set.
To allow more fair comparisons, competitions are being set up in various fields. Often in the period before a conference, a competition is set up, where everyone can try his algorithm on a common dataset given by the organizers.
To my knowledge, the reproducible research efforts in computational sciences were started by Jon Claerbout (who retired earlier this year) in the early 90s. In his Stanford Exploration Lab at Stanford University, Claerbout and his colleagues (working in seismic imaging) developed a system using Makefiles that allows to remove all figures, and reproduce them using a single Unix command. This allows any person (with a Unix/Linux system) to reproduce all the results in their work. I think it is about as close to “one-click reproducibility” as one can get! Claerbout and his lab performed a lot of the pioneering work in promoting reproducible research, which has spread later to various disciplines. A history by Claerbout himself is available here.
In their work, Claerbout and his colleagues make a distinction between three types of figures/results. First of all, and most common, there are easily reproducible results, which can be reproduced by a reader using the code and data contained in the electronic document. Secondly, conditionally reproducible results are results for which the commands and data are given, provided that certain resources are available (such as Matlab or Mathematica), or for which it requires more than 20 minutes to reproduce the results. And finally, non reproducible results, a label used for results that cannot be reproduced, such as hand-drawn figures, scans, or images taken from other documents for comparison.
Their Makefile setup was recently developed further by Fomel et al. in the Madagascar project, using SCons, a similar language to Makefiles, but which should make reproducibility even more simple, and cross-platform! See their project page for more details.
An article close to my current work on 3D now:
D. Scharstein and R. Szeliski, A taxonomy and evaluation of dense two-frame stereo correspondence algorithms, International Journal of Computer Vision, 47(1/2/3), pp. 7-42, April-June 2002.
In their article, Scharstein and Szeliski make a comparison of stereo estimation algorithms. But they do not just offer this overview of algorithms. On their webpage, they also provide the source code, and a widely used dataset of stereo images. They also invite other researchers to try their own algorithm on this dataset, and upload the results. This has resulted over the years in a performance comparison of almost 50 stereo algorithms, nicely listed on their webpage.
A nice example of what reproducible research can do! I think we need a lot more of these comparisons on common (representative) datasets.