jaschasd
The effect strength on humans ranges from a few percent deviation of human judgements from chance for subtle adversarial perturbations (epsilon=2), to ~15% deviations of human judgement from chance for large magnitude perturbations in the largest magnitude experimental condition.
For experiments 1 through 4, N was 38, 389, 396, and 389. The subjects were not undergrad psych students.
The article linked in the parent comment does not correspond to any experiment in the blog post or the Nature Comms paper.
Anyway -- I wanted to apologize for misreading -- I missed the parenthetical "interpolation" in your comment. I think we are both interpreting the plot the same way.
In terms of your comment about anecdotal evidence -- are you talking about the case where data and model size are increased jointly? If so, I agree, though I don't think that is any longer cleanly to do with double descent/overparameterization.
I believe the figure labeled "Figure 1" illustrates what your are suggesting (despite being labeled Figure 1, it is actually at the bottom of the blog post, so maybe easy to miss).
A brief note that I do discuss the deep double descent phenomenon in the blog. See the section starting with "One of the best understood causes of extreme overfitting is that the expressivity of the model being trained too closely matches the complexity of the proxy task."
I avoided using the actual term double descent, since I thought it would add unnecessary complexity. Lesson learned for next time -- I should have at least had an endnote using that terminology!
I'm using the minima Jekyll theme. I also used Markdeep (https://casual-effects.com/markdeep/) rather than vanilla markdown to write the post. (Markdeep is awesome. It took me a full day to figure out how to get it to play nice with Jekyll, but in the end of course it turned out to be simple. You can see how I did it by looking at the blog page source -- note the <pre class="markdeep"> block, and the "mode: 'html'" option.)
Similarly, concept drift can be a source of overfitting -- the objective you care about is the one after the concept drift occurred, but the objective you trained on is the one from before the concept drift. (Here's a scholar search for papers where the two concepts co-occur: https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&as_vis... )
Also -- as you say, feedback loops and non-stationarity make everything more complex, and are ubiquitous in the real world! But in machine learning we also see overfitting phenomena in systems with feedback loops -- e.g. in reinforcement learning or robotics, where the system changes depending on the agent's behavior.
(blog author here)
Yes! The post actually talks about that a bit: how overfitting can result from treating "leaders that have the most support in the population" as a proxy for "leaders that act in the best interests of the population"; and some ideas for improving that with noise regularization or capacity restrictions.
Overfitting can happen in many ways -- your training objective can be different at train and test time, or as you suggest the datapoints you use can be different at train and test time.
For overfitting induced by datapoints: If you include the datapoints in your problem specification, then you can say they induce bias at test time. If you treat the choice of training datapoints as a random variable, separate from the problem specification, then you can say they induce variance at test time. The difference is essentially semantic though. In general, you can freely move contributions to the error between bias and variance terms by changing which aspects of the modeling framework you define as fixed by the problem definition, and which you take to be stochastic.