> We also noticed significant improvements in performance of RND every time we discovered and fixed a bug [...]. Getting such details right was a significant part of achieving high performance even with algorithms conceptually similar to prior work.
They call bugs 'details' which, I find, is a frightening state of mind for someone publishing an algorithm.
It really depends on the algorithm. For example, a bug in the random number generator of a stochastic search algorithm that affects, say, the variance of a distribution won't have a relevant impact on the outcome.
I also had a few do-overs at the end of my thesis, but fortunately had a cluster standing by...
Well, there's also this notion of testing and regression. As I said in another comment a few days ago:
>A few weeks ago I had a conversation with a friend of mine who is wrapping up his PhD. He pointed out that not one of his colleagues is concerned whether anyone can reproduce their work. They use a home grown simulation suite which only they have access to, and is constantly being updated with the worst software practices you can think of. No one in their team believes that the tool will give the same results they did 4 years ago. The troubling part is, no one sees that as being a problem. They got their papers published, and so the SW did its job.
An independent implementation of the experiment is neccessary for a full reproduction anyways. If you just run their code again, you'll end up with all their bugs again.
(But, don't get me wrong. I like when researchers release their code. It's still very useful.)
1. People will not bother. It took a lot of minds to come up with the software used (in my friend's case, several PhD's amount of work). No one is going to invest that much effort to invent their own software libraries to get it to work.
2. Even when you do write your own version of the software, there are a lot of subtleties involved in, say, computational physics. Choices you make (inadvertently) affect the convergence and accuracy. My producing a software that gives different results could mean I had a bug. It could mean they did. It could mean we both did. Until both our codes are in the open, no one can know.
It is very unlikely that you'll have a case of one group's software giving one result and everyone else's giving another. More like everyone else giving different results.
Case in point:
https://physicstoday.scitation.org/do/10.1063/PT.6.1.2018082...
HN discussion:
If there's one thing I could convey to the world from what I learned from my time in academia, it is this: Most scientists at universities do not care about reproducibility.[1] Not only that, many people intentionally omit details from papers so that it is hard for rivals to reproduce their work - they want the edge so they can publish without competition. This isn't a shadowy conspiracy theory - this is what advisors openly tell their students. Search around on HN and reddit and you'll see people saying it.
[1] My experience is in condensed matter physics - it may not apply to all of academia.
I can believe that. I have seen code in academia and it was all one or two-letter variables without even linebreaks between statements.
Ideally the code would be part of the peer review process, but code review is really expensive, so who knows how that would play out.
Yes, but closed source helps ensure that low quality code is hidden from sight. It also means that people who distrust or doubt the conclusions have no chance to identify any bug(s) and disprove the results or conclusions.
We stop publishing in papers, and instead adopt smaller chunks of our work as the core publishing units.
Each figure should be an individually published entity which contains the entire computational pipeline.
Figures are our observations on which we apply logic/philosophy/whatyouwannacallit. Publishing them alongside their relevant code makes the process transparent, reproducible and individually reviewable, as it should be.
We can then "publish" comments, observations, conclusions etc on those Figures as a separate thing. Now the logic of the conclusions can be reviewed separately from the statistics and code of the figure.
As it is, research that yields a "failure" is buried. That means wheels are being reinvented and re-failed. That means there's no opportunity to compare similar "failures", be inspired, and come up with the magic that others overlooked.
Unfortunately, I would imagine, even if you can get researchers to agree to this the lawyers are going to have a shit fit. Imagine Google using an IBM "failure" for something truly innovative.
I agree in principle. But, for the experimental sciences, we need better publication infrastructure to make this practically possible.
For example, consider a figure that summarizes compares, between several groups, the mechanical strain of tensile test specimens for a given load. Strain is measured from digital image correlation of video of the test. Some pain points:
1. There is a few hundred GB of test video underlying the figure. Where should the author put this where it will remain publicly accessible for the useful lifetime of the paper? How long should it remain accessible, anyway? The scientific record is ostensibly permanent, but relying on authors to personally maintain cloud hosting accounts for data distribution will seldom provide more than a couple years' of data availability.
2. Open data hosts that aim for permanent archival of scientific data do exist (e.g., the Open Science Framework), but their infrastructure is a poor match with reproducible practices. I haven't found an open data host that both accepts uploads via git + git annex or git + git LFS and has permissive repository size limits. Often the provided file upload tool can't even handle folders, requiring all files to be uploaded individually. Publishing open data usually requires reorganizing it to according to the data host's worldview or publishing a subset of the data, which breaks the existing computational analysis pipeline.
3. Proprietary software was used in the analysis pipeline. The particular version of the software that was used is no longer sold. It's unclear how someone without the software license would reproduce the analysis.
Finally, there's the issue of computational literacy of scientists. In most cases, the "computational pipeline" is a grad student clicking through a GUI a couple hundred times, and occasionally copying the results into an MS Office document for publication. No version control. Generally, an interactive analysis session cannot be stored and reproduced later. How do we change this? Can we make version control (including of large binary files) user-friendly enough that non-programmers will use it? And make it easy to update Word / PowerPoint documents from the data analysis pipeline instead of relying on copy & paste?
If any of these pain points are in fact solved and my information is out of date, I would be thrilled to hear it.
3: analysis that uses propriatory is marked appropriately as second class
> computational literacy of scientists
Welp...
research just jumped onto jupyter notebooks, it's halfway there, someone helps the remaining step
Science should prove things...
Logical proofs will never happen for software development, but surely standards for scientific programming can be tightened up a few levels!
I think I heard of some reform proposals from the Reproducibility Crisis reformers.
There's also the fact that we have had a pretty solid grasp about the chemical reactions of greenhouse gases since long before computers, both theoretically and empirically, and we know roughly how much is put out in the atmosphere.
Where the models diverge is on far finer points than what is needed to make the basic policy changes that seems to be where we are stuck right now.
Entertaining[0] badpun’s skepticism, it is not necessary they have identical bugs, only that their bugs yield similarly biased results.
For example, if a significant number of bugs are identified by their affect on the results, then bugs contributing to “wrong” results might be more likely to be identified and fixed.
[0] In the Aristotelian sense.
If global warming is actually wrong, it's most likely because of bad/corrupt data in the datasets used.
http://www.realclimate.org/index.php/archives/2009/11/the-cr... http://www.realclimate.org/index.php/archives/2009/11/the-cr...
Since it was sound science in 2009, why would it not be now?
this will be the least of our problems. we understand very little about nature to have any reasonable prediction model. we don't know the inflection point which will cause massive collapse of major ecosystems we depend on.
all we know is that things are changing fast. faster than many non-human organisms are able to adapt.
Are you talking about predictions (and not measurements, you don't need a model to measure temperature)? Assuming you are, there's unfortunately a huge problem with modeling (and heavy math and stats-based science in general), in that researchers tend to stop looking for bugs in the model when it returns the results that they expect. In other words, if a bug in the model tells the researcher that Earth's temperature will decrease by 4 C by 2100, he will look over the model until he finds the bug, but if the model tells him that the temperature will increase by 2 C, thus confirming his inner bias, he'll declare it correct and move on to writing a paper based on the "finding".
Alternatively, as a thought experiment, imagine if math research were done in the way climate science is done. We would have proofs that are millions of pages long and were never verified by anyone. We would trust in them only because the author says that they are correct. Is this science?
>>Researchers tend to stop looking for bugs in the model when it returns the results that they expect.
Again... they are _ALL_ biaised ?
Also, how many independent, comprehensive models (with codebases) for global climate change are there in the science world?
False. Lots of past climate changes were due to changes on Earth and its atmosphere (and sometimes, specifically, life on Earth), not changes in solar output (e.g., notably, the Huronian glaciation believed to have resulted from the Great Oxygenation Event, which resulted from the exponential growth of photosynthetic life.)
That the post-industrial rise in atmospheric CO2 is of anthropogenic origin is hopefully not a point of dispute, but it is demonstrable if necessary. Thus it remains to show that this must raise the equilibrium temperature. So, as you say, CO2 selectively absorbs outgoing IR. In the lower atmosphere, this actually does not have as much of an effect as you might think. Water vapor blocks quite a bit of the absorption spectrum, and the effect of CO2 is more-or-less saturated already.
The mean free path of an outgoing IR photon in the lower atmosphere is quite short. Absorbed photons are re-emitted in a random direction, but take an overall upward course, the mean free path rising with altitude. At the (radiative) top-of-atmosphere, the mean free path is infinite: the photon is more likely to leave Earth. At the edge of space, there is essentially no water vapor, so the action of CO2 is greater. The effect of increasing the amount of CO2 in the atmosphere is to push the CO2-dense region of the atmosphere further out into space. Photons must take a longer path out of the atmosphere, and this must raise the overall temperature of the Earth proportionally, specifically by 3.7 W/m^2 per doubling of CO2, which is commonly held to be equivalent to 1 degree C of global temperature. This must be the case unless our understanding of thermodynamics is very wrong (and if you have an issue with thermodynamics then you have some pretty serious issues).
So, one degree C ain't so bad, right? Well, it wouldn't be if that were it. However, there are several problematic feedbacks. One is that melting a lot of ice lowers the Earth's albedo, which causes it to absorb more heat. Another issue is that there is a lot of this "water" stuff around, which is very readily absorbed by the atmosphere, in a manner that increases very sharply with temperature. Water vapor is a much better greenhouse gas than CO2 by all accounts.
Climate science is not an extrapolation from the temperature record. There is a solid minimum bound on the temperature effects of doubling atmospheric CO2, and a variety of amplifying positive feedback effects. So far, in the last twelve decades, we have not managed to find anything which would reduce those effects to something manageable. At this point, the effect would need to be both very large, in order to offset the strong H2O feedbacks, and very small, to not have been noticed. The most plausible option would be "something poorly understood about the H2O feedbacks". I believe the most successful of such theories would be Dr. Richard Lindzen's Iris Hypothesis, which has generally failed to find support. At this point, there are no particularly plausible mechanisms which would transfer this extra energy to space, and if those did exist, then they would not necessarily be a non-issue: even if thermodynamics and optics are entirely wrong, the planet is warming, and we will have to deal with that even if it can't be prevented.
If you have any other questions, or would like citations for any of the above, do feel free to ask.
- https://en.wikipedia.org/wiki/Milankovitch_cycles
- http://www.indiana.edu/~geol105/images/gaia_chapter_4/milankovitch.htm
Interestingly, the original paper proposing AGW (in 1896) was actually intended to explain Ice Ages:
- http://www.rsc.org/images/Arrhenius1896_tcm18-173546.pdfThere could be lots of reasons for that. Anyway, the temperatures were not always rising, clear temperatures rise was observed in 1930-40s and in 1980-90s. Cooling in 1960-1970s. And yes, prediction has to be precise. If a model predicts rise of 3K in 100years, and you measure 1.3K - your model is wrong. It's even more wrong if you don't take into account any of the natural cycles, even if prediction is accidentally correct.
I think at this point the group think is so strong that if anyone came out with evidence against global warming they would commit career suicide by publishing it
Oh good. So you'd rather see millions of people be displaced. We're fucked as a species when "I'm a contrarian" is a valid reason to disagree.
> they would commit career suicide by publishing it
Meanwhile some researches create their career by publishing this sort of stuff, even when it isn't well researched
The bugs, if serious enough, can make it not scientific, in the same way that a paper that has grave errors in it is wrong and thus obviously not scientfic. So, before we do a thorough review of the code, we should treat it with due scepticism.
EDIT: this can also be expressed in terms of risk analysis. For typical software, the consequence of a bug is low - most software is commercial, so the impact of the bug will be limited to that company's bottom line (with exceptions of software that can kill, but these people already are serious about bugs). Also, most bugs in most software are either highly evident (the button does not work, you get random segfaults etc.) or have limited impact.
On the other hand, the bugs in climate change models, given their "pipeline" nature (wrong result from one module is propagated downstream all the way to the final prediction of expected temperature change), can quite often have severe impact. They can also be not evident at all - they can for example change the final outcome predicition by 1 C.
Compound that with the fact these predictions are used to make trillion dollar decisions on global policies, and you can see that the actual damage done by bugs is not unlikely to be in the trillions. And that's why I say it's probably wise to subject the models to extreme scrutiny.
"Already 18824 SIGNATURES – sign the open letter now!" https://publiccode.eu
I wouldn’t be surprised if a significant portion of published computational research has bugs that totally invalidate the conclusions. I think we need to push hard to require all taxpayer-funded research to make any code that results in a journal article publicly available.