This is part of the reasons why A/B testing might fail in reality, it is not just about collecting random number of responses about A and B.
When doing a test you should plan it in advance based on other data, you should execute the test as planned and analyze the results accordingly failing any of the tests will results in wrong to totally wrong conclusions.
If you're doing A/B testing where you continuously watch the results, and conclude your test as soon as you see the numbers significantly deviate, what you're doing is more-or-less equivalent to what's happening in the comic, only it's the number of jelly beans that you vary instead of their color. Not exactly the same, since in the numbers case your draws from the random variable aren't independent, but still.
> When doing a test you should plan it in advance based on other data
I think this is not entirely true. In areas where experimentation feedback is fast (and experiments are probably cheaper to run), the problem is much more accurately and sample efficiently solved by Thompson Sampling [1,2] which in fact dictates you to have a broad enough prior over the solution space and then let the posterior dictate your conclusions.
The link is not to the paper itself but to a response to the paper, this might be (is probably?) intentional but the title is a bit confusing.
Essentially the response, submitted by several ML / CS / math researchers, addresses a note in the paper which recommends against train/test split in model training, calling it "inefficient". The response is dedicated to explaining what generalization error is and why estimating it is important, and how that's basically impossible without any sort of train/test split or cross-validation.
The article is there, if you click on the "Article" tab.
Perhaps if any staff are paying attention on the holiday, we could get the "/rr" chopped off the end of the link, so that we get to the main article instead?
My guess is that OP intended to submit the response, I was really surprised to see the note in the article and found the response interesting (even the fact of its existence). But that's just my take :)
And if you’re a Bayesian, you can choose a stopping criterion based on your desired degree of certainty. This can have practical benefits of not having to run an experiment to its end.
A family member was part of the control group for a cancer treatment study a few years back. The study chose a stopping criterion based on Bayesian methods. Relatively early into the study they were able to determine that it made sense to move people from the control group to the treatment group.
Once you're dealing with real people, early termination is a huge deal, and it quickly goes beyond just the interpretation of statistics. You almost always want to have 3rd party referees involved anyways.
By "beyond statistics", I think they meant that trials need on-going monitoring for things beyond the outcome variable itself: safety, data quality, and even feasibility of completing the trial itself. Once you've got that infrastructure in place, interim monitoring of the outcome isn't much extra work.
Some of the frequentist approaches to early-stopping seem pretty coherent to me. Curtailiment, for example, just stops collecting data once it will no longer change the outcome of a test. The frequentist emphasis on error control for the procedure also seems like a reasonable fit for a regulatory regime where you're actually making decisions (no argument that it is...weirder for basic research)
At any rate, some approaches for early-stopping have sensible frequentist and Bayesian properties, which is nice.
Estimating a minimum required sample size is one of the most common questions asked by clinical or biomedical collaborators before embarking on a research project. This is especially true when ML is an option. This paper provides rules of thumb and a digestible amount of theory that could inform such conversations, and will surely become a popular reference.
Note intuition from traditional statistics does not universally apply to deep learning and/or extremely high-dimensional data. For example, deep neural networks with 1-4 orders of magnitude more parameters than training examples can still generalize well to unseen data.
Not sure if any of the test-driven development people have thought about this, but the same principle also applies there: if you debug and fix the code until it passes the tests, you can overfit the tests. It's no longer a good measure for code quality, once it has been explicitly optimized. You'd need new, previously unseen test cases. It would be an obvious mistake in machine learning to simply add failed test examples into the training set and rejoyce that the new model can now deal with those previously difficult cases.
It's related to Goodheart's law: "When a measure becomes a target, it ceases to be a good measure".
It also works in one's personal life. If you find that you made a mistake, don't just fix that particular thing, go back and see what other similar things may need fixing. It connects to the idea of fixing the deeper reason than fixing the symptoms. "Deeper reason" just means, something that generalizes.
By the way, you need to be careful about your data split. This always depends on how you intend to use the model. If you intend to use it on new, unseen patients, then the train and test data cannot overlap with respect to people. Another obvious case is videos: You can't just take every even frame as a training sample and every odd frame as test. Even though technically you would be testing on non-overlapping data, the performance measure would be biased compared to real world performance. Or if you want to classify burglars from security camera footage, you may need to test it on new camera setups from different houses if you intend to deploy it to new houses. If your scenario is such that you'd perform training on each new site and run a location specific model, you can test on images from the same site.
You always have to use your brain to decide what to do.
> Not sure if any of the test-driven development people have thought about this, but the same principle also applies there: if you debug and fix the code until it passes the tests, you can overfit the tests. It's no longer a good measure for code quality, once it has been explicitly optimized. You'd need new, previously unseen test cases. It would be an obvious mistake in machine learning to simply add failed test examples into the training set and rejoyce that the new model can now deal with those previously difficult cases.
This is a great point, and it articulates some challenges that I’ve seen with test-driven development.
I would say that if you can write tests that fully specify the desired functionality, rather than merely check a few possible inputs, it’s less of an issue. This is a reason to try to build things with less knobs to twiddle, so the space of inputs has lower dimensionality.
Yes but then the tests would formally describe the requirements - i.e.they would be a de-facto implementation. Congratulations, you just coded an (indirect) solution to your problem (a solution that you postulate is correct/bug-free)
My beef is with the assumption that tests can fully encode the requirements - to the point where any bug would trigger a test failure. Having such a test suite is no simpler than having a perfect implementation of the requirements - i.e. it's probably only feasible, at all, in the simplest/ "didactic" cases.
Having a perfect implementation is not bad, of course. It's just not realistic.
I was super into TDD and never thought that there was such a thing as "TDD going too far" until I read that post by an ex-Oracle employee: https://news.ycombinator.com/item?id=18442941
This is a big reason why I prefer property testing to unit testing. One can think of focusing on the abstract behavior and invariants of the code, rather than just the expected outputs for specific inputs, as being somewhat akin to taking the time to plan out your experiment and the statistical tests you want to perform ahead of time, do your power analysis, etc. It's a fair bit more up-front work, but helps to ensure that you have a crystal clear idea of what you're going to do before you start doing it.
21 comments
[ 3.3 ms ] story [ 65.2 ms ] threadWhen doing a test you should plan it in advance based on other data, you should execute the test as planned and analyze the results accordingly failing any of the tests will results in wrong to totally wrong conclusions.
For starters, consider this xkcd comic: https://xkcd.com/882/
If you're doing A/B testing where you continuously watch the results, and conclude your test as soon as you see the numbers significantly deviate, what you're doing is more-or-less equivalent to what's happening in the comic, only it's the number of jelly beans that you vary instead of their color. Not exactly the same, since in the numbers case your draws from the random variable aren't independent, but still.
I think this is not entirely true. In areas where experimentation feedback is fast (and experiments are probably cheaper to run), the problem is much more accurately and sample efficiently solved by Thompson Sampling [1,2] which in fact dictates you to have a broad enough prior over the solution space and then let the posterior dictate your conclusions.
[1]: https://www.microsoft.com/en-us/research/wp-content/uploads/...
[2]: http://www.economics.uci.edu/~ivan/asmb.874.pdf
Essentially the response, submitted by several ML / CS / math researchers, addresses a note in the paper which recommends against train/test split in model training, calling it "inefficient". The response is dedicated to explaining what generalization error is and why estimating it is important, and how that's basically impossible without any sort of train/test split or cross-validation.
Perhaps if any staff are paying attention on the holiday, we could get the "/rr" chopped off the end of the link, so that we get to the main article instead?
A family member was part of the control group for a cancer treatment study a few years back. The study chose a stopping criterion based on Bayesian methods. Relatively early into the study they were able to determine that it made sense to move people from the control group to the treatment group.
Once you're dealing with real people, early termination is a huge deal, and it quickly goes beyond just the interpretation of statistics. You almost always want to have 3rd party referees involved anyways.
The reasoning for doing so is much more coherent (imo) in a Bayesian framework where you don't have to say that you are "going beyond" statistics.
Some of the frequentist approaches to early-stopping seem pretty coherent to me. Curtailiment, for example, just stops collecting data once it will no longer change the outcome of a test. The frequentist emphasis on error control for the procedure also seems like a reasonable fit for a regulatory regime where you're actually making decisions (no argument that it is...weirder for basic research)
At any rate, some approaches for early-stopping have sensible frequentist and Bayesian properties, which is nice.
Agreed that things are weirder for basic research.
Note intuition from traditional statistics does not universally apply to deep learning and/or extremely high-dimensional data. For example, deep neural networks with 1-4 orders of magnitude more parameters than training examples can still generalize well to unseen data.
It's related to Goodheart's law: "When a measure becomes a target, it ceases to be a good measure".
It also works in one's personal life. If you find that you made a mistake, don't just fix that particular thing, go back and see what other similar things may need fixing. It connects to the idea of fixing the deeper reason than fixing the symptoms. "Deeper reason" just means, something that generalizes.
By the way, you need to be careful about your data split. This always depends on how you intend to use the model. If you intend to use it on new, unseen patients, then the train and test data cannot overlap with respect to people. Another obvious case is videos: You can't just take every even frame as a training sample and every odd frame as test. Even though technically you would be testing on non-overlapping data, the performance measure would be biased compared to real world performance. Or if you want to classify burglars from security camera footage, you may need to test it on new camera setups from different houses if you intend to deploy it to new houses. If your scenario is such that you'd perform training on each new site and run a location specific model, you can test on images from the same site.
You always have to use your brain to decide what to do.
This is a great point, and it articulates some challenges that I’ve seen with test-driven development.
I would say that if you can write tests that fully specify the desired functionality, rather than merely check a few possible inputs, it’s less of an issue. This is a reason to try to build things with less knobs to twiddle, so the space of inputs has lower dimensionality.
Tests describing requirements acts as documentation, validation, and regression prevention with future refactors.
Having a perfect implementation is not bad, of course. It's just not realistic.