arg min

I could rant about this for hours. I actually just went to a defense for a deep learning paper that had a ton of abstract algebra. I am honestly not really a fan of deep learning and algebra because all the papers to me like- seem to stop at describing some really basic feedforward network as some really specific mathematical structure but these theories a) provide very little explanation of empirical phenomena b) provide no new directions of research in terms of like useful network architectures. I haven't really come across algebra in machine learning other than people applying it to deep learning. i.e. https://arxiv.org/pdf/1802.03690.pdf ie. https://icml.cc/Conferences/2018/Schedule?showEvent=2048 I don't personally find papers like this valuable but idk I have never really enjoyed abstract algebra. For areas of mathematics to do theory in ML (and to do ML more generally!) -probability/concentration/hoeffding bounds [the PAC model] [Key] -linear algebra [key] -optimization [key] for books -understanding machine learning by shai ben david This book is nice since it really balances theory with a more practical understanding. -An Introduction to Computational Learning Theory by kearns is a classic [low priority]. this is fun since the proofs are simple and deep but is very very far away from practical algorithms. -convex optimization by boyd Course Notes: [I think a good alternative to blogs is stalking course notes for other schools-they are very…

If anyone enjoyed this article, then i'd recommend reading this one as well[1], it's an interesting article with a focus on the relationship between radio and probabilistic reasoning in the early 1900s. https://www.argmin.net/p/the-spirit-of-radio

The Perceptron is also interesting because its classic mistake bound analysis is one of the earliest instances of modern learning theory. The mistake bound states that if the data is linearly separable (i.e. a perfect linear classifier exists), then Perceptron makes no more than (r / m)² misclassifications, where r is the radius of a ball containing the data set and m is the margin between the two classes. The analysis is modern because 1) unlike traditional statistics, it does not make any assumptions about the distributions of each class, and 2) is a non-asymptotic guarantee. It is also an online analysis, in the sense that it only compares performance against the optimal fixed classifier for the sequence that was actually observed, rather than any notion of an underlying distribution. More details: http://www.argmin.net/2021/11/04/perceptron/

The Bangladesh study failed to produce a significant result when simpler nonparametric statistics were used to analyze the data. There were also significant discrepancies in the follow up rate between the control and treatment groups. And even if you ignore those issues, the absolute effect size (20 cases) was tiny. The effect size was expressed in terms of the relative rate of symptomatic seropositives, i.e., normalized by population denominators, which makes it appear larger and is sensitive to the follow up rate discrepancies (see above). Ben Recht at UC Berkeley has several good blog posts on this issue: https://www.argmin.net/2021/12/01/unblinding/

The blog post "You Cannot Serve Two Masters: The Harms of Dual Affiliation"[1] from prominent AI researchers Ben Recht, David A. Forsyth, and Alexei Efros also highlights the challenges of co-employment by big tech companies. Well worth a read. [1] http://www.argmin.net/2018/08/09/co-employment/

this is a followup post i wrote to answer exactly your question: http://www.argmin.net/2018/01/25/optics/

Not a comment on that paper, but for those interested I would highly recommend Ben Recht's (ont of the authors of that paper) blog series at http://www.argmin.net/2018/06/25/outsider-rl/.

Ben Recht also has an excellent series of blog posts (very related to this survey on arXiv, but broader) on the intersection between reinforcement learning and optimal control. An index is available here: http://www.argmin.net/2018/06/25/outsider-rl/

ali rahimi alludes to the problem of google engineers simply needing to tweak models that were previously tuned by google researchers who do have well-developed intuition [0]. because the intuitions in explicit form are at best heuristic and not necessarily even consistent, signing up to improve a model without them might result in spending indefinite time and compute resources without guarantee of positive results. which is a terrible perf-theoretic strategy... [0] http://www.argmin.net/2018/01/25/optics/

That makes sense. I think the author would argue that the model is never needed, just go back to the actual population, and make inferences based on random sampling. From another prior blog post[1], he states: "So what is the remedy here? The thing is, we already know the answer: if we randomized the assignment ... it is critical to decouple the randomness used to probe a system from the randomness inherent in its system itself. Statistical models are not necessary for statistical inference, but randomness itself is amazingly… let’s say… useful for understanding natural phenomena." This makes sense to me if there's always a population that you can random sample. And yes, you'd have to sample a lot for the sample and population summaries to converge, but this seems fine in certain contexts (i.e. when you have access to computer simulations). Your counterpoint seems correct when you aren't able to randomly sample the population in this manner. Would you agree with the author based on the specific framing I'm making (randomly sampling population beats building a probabalistic model if you are able to do the former). Again, I'm a stats novice so apologies if I'm making an obvious point. 1. http://www.argmin.net/2021/09/21/models-are-wrong/

This article is a little too glib in my opinion, preferring citations and statements to substance and explanations. For a more cutting and insightful critique, watch Ali Rahimi's short talk at NIPS 2017 (where he was presenting for a paper that won the "Test of time" award, for standing out in value a decade after publication). The standing ovation he received at the end indicate that his comments resonated with a significant fraction of the attendees. https://www.youtube.com/watch?v=Qi1Yry33TQE&feature=youtu.be&t… Here's a teaser from the talk: "How many of you have devised a deep neural net from scratch, architecture and all, and trained it from the ground up, and when it didn't work, felt bad about yourself, like you did something wrong? This happens to me about every three months, and let me tell you, I don't think it's you [...] I think it's gradient descent's fault. I'll illustrate..." [Addendum] Ben Recht and Ali Rahimi published an adendum to the talk, elaborating on the direction they envision -- http://www.argmin.net/2017/12/11/alchemy-addendum/ Ali also has a post taking a stab at organizing some puzzling basic observations about deep learning, and motivating that with analogous historical progress in optics -- http://www.argmin.net/2018/01/25/optics/ ----- PS: The first 11 minutes, on the idea of using random features (the main idea in the research he presented) are also interesting.

FWIW, you may enjoy https://www.argmin.net/p/lore-laundering-machines

> It's a really fabulous study... Ehhhh... not so much. It had serious design flaws in both the protocol and the analysis. This blog post is a fairly approachable explanation of what's wrong with it: https://www.argmin.net/p/are-developers-finally-out-of-a-job

Here is Ben Recht’s response: https://www.argmin.net/p/toward-a-transformative-hermeneutics

The article here is responding to an original blog post [1] that is not really saying the Higgs was not discovered (despite its trolling title), but raising questions about the meaning of "discovery" in systems that are so complicated as those in modern particle physics. I think the author is using the original motivation of musing on null hypotheses to derive the title "The Higgs Discovery Did Not Take Place", and he has successfully triggered the controversy the subtitle ironically denies and the inevitable surface reading condemnations that we see in some of the comments here. [1] https://www.argmin.net/p/the-higgs-discovery-did-not-take

Be sure to check the comments at the bottom. The authors of the papers in question respond to these criticisms. https://www.argmin.net/p/is-the-reproducibility-crisis-reproducible/co… https://www.argmin.net/p/is-the-reproducibility-crisis-reproducible/co… https://www.argmin.net/p/is-the-reproducibility-crisis-reproducible/co…

The next post in the author's blog addresses the comments by the authors of the article [0]. Also, I don't think the arguments written in the post are "unserious", you might not like them but the author of the blog makes some interesting points against the original paper. [0]: https://www.argmin.net/p/arbitrage-in-data-exchange-rates

Alas even today there is no RCT on masks. Why? Countries and states with strict mask policies did not do better than those without these policies (see US States for an example). So do mask work? It is like a religious belief system, you would get the same reaction when questioning if God existed and asked parishioners and atheists. Everything is evidence for those that want to believe. Studies that claim to "prove" that masks work turn out to be shambolic messes - yet are published in the most selective of journals, see for example: Effect size is significantly more important than statistical significance. http://www.argmin.net/2021/09/13/effect-size/

Interesting, I didn't know about cybernetics. For the ones interested there is a book that discusses both: 'Reinforcement Learning and Optimal Control', by Dimitri P. Bertsekas. It covers exact and approximate Dynamic Programming, finite and infinite horizon problems, deterministic and stochastic models, model-based and model-free optimization. Aside from this book, Ben Recht has some interesting blog about Optimal Control and Reinforcement learning: http://www.argmin.net/2018/06/25/outsider-rl

Make that zero. To their credit, the authors of the Bangladesh mask study released the raw data. "The difference between the two groups was small: only 20 cases out of over 340,000 individuals over a span of 8 weeks." Drawing any conclusions from such small amount of data is deeply unserious. http://www.argmin.net/2021/11/23/mask-rct-revisited

Ben Recht [0] has worked on control theory and reinforcement learning. See his blog post "What we've Learned to control" [1] and the survey paper "A Tour of Reinforcement Learning: The View from Continuous Control" [2] may be of interest. [0]: https://people.eecs.berkeley.edu/~brecht/ [1]: https://www.argmin.net/2020/06/29/tour-revisited/ [2]: https://arxiv.org/abs/1806.09460

I wonder if these guys are a bit quick to give themselves credit. There's a lengthy (and depressing) series of posts by Ben Recht[1] that basically outlines all the reasons that model-free RL is effectively bunk if you have access to any model of how the environment evolves. These guys (as far as I can tell), haven't made any attempt to disambiguate how much of the improvement they see is down to having a model of the environment _at all_, as against the specific model they propose. I think there's probably more these guys could do to prove that they're actually onto something and aren't just confusing throwing compute at the problem with a genuinely better solution. [1]: http://www.argmin.net/outsider-rl.html

Additional context from Ali Rahimi and Ben Recht: http://www.argmin.net/2017/12/11/alchemy-addendum/

To quote Ben Recht "When you end up with a bunch of papers showing that genetic algorithms are competitive with your methods, this does not mean that we’ve made an advance in genetic algorithms. It is far more likely that this means that your method is a lousy implementation of random search." http://www.argmin.net/2018/02/20/reinforce/

In case you'd rather read text than watch a video, I believe this is the same? http://www.argmin.net/2017/12/05/kitchen-sinks/

Regarding the presented studies that assessed mask wearing: Bundgaard (not Bundagaard) 2021 [1]: −0.3 percentage point (95% CI, −1.2 to 0.4 percentage point; P = 0.38) difference; statistically not significant. Limitations: “Inconclusive results, missing data, variable adherence, patient-reported findings on home tests, no blinding, and no assessment of whether masks could decrease disease transmission from mask wearers to others.” Doung-ngern 2020 [2]: 11% of control not tested; 59.2% of cases linked to boxing stadium; Persons who wore masks were also more likely to practice other measures [e.g. washing hands often (79% vs. 26% for those who did not wear a mask)] Krishnamachari 2021 [3]: Heavily confounded. States with no mask mandate had a 2.16x higher case rate, yet “stay at home orders and school closures had no significant influence”? No analysis of “restrictions on gatherings, closing of nonessential business and restaurant closures”? Lio 2021 [4]: Retrospective, internet questionnaire with 24 cases. Wang 2020 [5]: Telephone interview; Other than that, no issues imo. Xu 2020 [6]: “Despite the relatively large sample size, the total cases of COVID-19 infections were still small so that the relationship between NPIs and a COVID-19 infection should be confirmed by other larger epidemiological studies. Fifth, the potential risk compensating effects of wearing a mask against other NPIs should be considered as being of a hypothesis-generating nature given the…

I consider GPs in two ways: hyperparameter optimization (where GPs are a specific instantiation of bayesian optimization) and a powerful non-linear regression method that supports uncertainty via the rich posterior distributions you get with GPs. However, on the side of hyperparameter optimization, an experiment comparing bayesian optimization to random search with only double the computational resources [0] suggests that BO is not at least 2x better than random search which, to me, means that in practice BO isn't my first go-to for hyperparameter optimization. On the side of regression, its poor time complexity (you need to invert the data matrix which is O(n^3) and slow in practice with maybe >1000 points) makes it less of a first go-to for me than neural networks, which maybe irrationally benefits from its current "hotness", but also for which there are well-explored strategies for generating predictions with uncertainty [1]. Of course, GPs compare significantly more favorably on the axis of interpretability, since you can get a closed-form posterior distribution for any point of interest, and the manner in which your dataset influences this posterior is transparent. There have been, however, interesting developments in improving the efficiency of GP regression for large datasets, such as estimating the GP with a reduced and weighted set of data points [2]. The recent Deepmind paper on neural processes is also interesting, as a combination of GPs and neural networks…

Yes, I am aware, I did not mean random search as in random actions, but random search with improved heuristics to find a policy. The point being that that the bells and whistles of PPO and other relatively complaticated algorithms (e.g. Q-PROP), namely the specific clipped objective, subsampling, and a (in my experience) very difficult to tune baseline using the same objective, do not significantly improve over gradient descent. And I think Ben Recht's arguments [0] expands on that a bit in terms of what we are actually doing with policy gradient (not using a likelihood ratio model like in PPO) but still conceptually similar enough for the argument to hold. So I think it comes down to two questions: How much do 'modern' policy gradient models improve on REINFORCE, and how much better is REINFORCE really than random search? The answer thus far seemed to be: not that much better, and I am trying to get a sense of if this was a wrong intuition. [0] http://www.argmin.net/2018/02/20/reinforce/