activatedgeek

I tried searching for one a few months ago when my Economist subscription was expiring. Unfortunately, not even a single one comes close.

The Atlantic is often advertised as the other high-quality publication. The target audience, however, is overwhelmingly the US. Economist makes me feel like other countries exist, and there is enough going on outside the US.

On the ideological front, I fell into this trap of finding a periodical with little bias towards the political left or the political right. While Economist touts itself to be centre, it would be naive to take them at face value as they clearly are centre-left. The silver lining is that they are far better than others who make absolutely no efforts to quell their suppress their ideological biases (e.g. New York Times).

I call this a trap because I've since realized that it is virtually impossible to expect a "truly" unbiased publication. As long as I feel that I am not falling into an echo chamber, I am fine with biases. Politically charged articles generally have more than one perspectives.

One would be hard-pressed to find a single publication as diverse as the Economist. I think this is primarily a historical artifact - they just have had a significant first-mover advantage (I think) and have accumulated a solid bunch of human capital. I'm slowly inclining towards subscribing to more niche publications instead.

Financial Times is a reasonable contender, though far more costlier.

In similar spirit, Gradescope [1] is an excellent tool, that I just cannot do grading without now.

Students upload PDFs, and mark appropriate pages for each assigned question. The graders can then decide to either mark every question in a single sheet, or a single question in every sheet. I find the latter more comfortable. The key element here is that the burden of assigning pages to questions shifts from the grader to the student, which is a significantly better use of cognitive resources because it comes at little marginal cost to the student but tremendous cost for the grader.

[1]: https://www.gradescope.com

> I tried a great deal of ML approaches, only to find recently that by better accounting for the optical specificities of the problem I was trying to solve

I want to point out, that any serious machine learning researcher is not oblivious to this, despite the deep learning boom suggesting to the contrary. Modern methods have shown that we are capable of building predictors with surprisingly complex representations, that can solve large-scale downstream tasks. i.e. our models are "flexible" enough.

The next challenge is whether they favor the "right kind" of solutions. For instance, Convolutional Neural Networks (CNNs) are architecturally just sparse version of Fully-Connected Neural Networks. Why is it then that CNNs perform far better on images? A key reason is that "inductive biases" afforded by MLP aren't strongly favored towards images. Another instance of this is the covariance functions used is Gaussian Processes - the Squared Exponential Kernel is very flexible and can in principle fit anything possible. Nevertheless, if the problem has specific structures, say periodicity, one better use the Periodic Kernel because it's inductive biases rightly align with the kind of solutions we expect.

> The lesson is : physics beats AI.

As a consequence, the single biggest reason physics would beat a generic AI in the short-term is precisely due to our ability to explicitly provide inductive biases that align with our expectations from the physical system.

We haven't found the secret sauce for every possible system in the universe. I don't think we can, either. But what we can do is devise ways to "control" such inductive biases we can encode in machine learning systems, which align with our expectations of the way the system should behave.

I like the general treatment of constraints for continuous distributions.

I wrote an alternative derivation for the Gaussian as a maximum entropy distribution with finite variance here [1].

[1] https://im.perhapsbay.es/kb/the-gaussian-cheatsheet/#as-a-ma...

Wow! Black Mirror's Nosedive (Season 3 Episode 1) [1] is straight out of reality.

[1]: https://www.imdb.com/title/tt5497778/

Appreciate it!

I can confirm it works with at least my import.

Edit: Alas, some errors regarding "Book not in collection" and "No ISBN". But I think I can probably see if there is a way to contribute those books to the repository.

I don't think I am particularly worried about a proprietary solution. If I am putting my preferences in the public, pretty sure I am giving up on some dimension of privacy. With books at least, it doesn't seem so concerning as long as the recommendations don't suck too much (I don't particularly rely on recommendations that much though). But of course, you never know!

Just tried this out. Quite a large community, so I'm assuming it is a well fed machinery. The interface is outdated. Nevertheless, the website does its job quite well. I am happy I found this. Thanks!

...And they even have an API!

Since the past few months I have been searching for a Goodreads alternative. Something that only keeps my books. I don't care about the social features that much. And I think this is it. I am going to donate a tiny bit right away!

Although, I just tried importing my Goodreads export into Open Library and I get the following "Internal Error":

> Hmm... > Sorry. There seems to be a problem with what you were just looking at. > We've noted the error xxxx-xx-xx/yyyyyy and will look into it as soon as possible. Head for home?

Anyone else facing this issue?

Eleven Years A-Go?

I think if in any context one uses the usual matrix algebra, it is the same linear map, just in a more abstract setting.

If you tell me storing a lookup table is matrix, I think that is sort of out of context. But then again, I could be labelled a purist here, and I'm trying really hard to not fall into the trap.

For the more machine learning-minded folks, this happens almost always in doing inference with Exact Gaussian Processes (GP), where because of the non-parametric nature of a GP model, the covariance matrix grows with the size of data points. The inference routine is cubic in the number of data points. Hence, sparsity in the posited covariance matrix is _extremely_ important for fast inference.

To be sure there isn't a perspective here that I may be unaware of, is this a reference to circulant matrices, which bear the Toeplitz sparsity structure?

I'd prefer stating that matrices actually don't mean anything in isolation. They are simply representations of linear maps. What happens to the elements of a matrix before the elements are formed is irrelevant to the matrix itself.

..And to be clear, I am not a purist.

Did not see this mentioned in any of the comments, but Metaculus [1] has been around for a while.

Do you directly compete with them? Nothing wrong in competing, I have a feeling that people are probably more _aggressive_ (in a positive sense) in their outlook when "Facebook profiles" are not on the line. Nevertheless, I think it is a great experiment.

I do imagine that at some point, if this product goes mainstream, journalists would start referring to these statistics for topics which really require correcting for exposure bias. For instance, exit polls. In places where Facebook penetration (or even internet penetration) is not-so-great, this will blindside decisions. Of course, this by no means will be Facebook's fault but I am curious to see how Facebook is proactively adjusting for this adverse outcome.

[1]: https://www.metaculus.com

Never write someone or something off without solid evidence.

> Check out this "map" of mathematics: ... Every subject needs a similar map.

I often find this sort of approach to be rather counter-productive. If I am learning something, I'd be happy with my local part of the graph, the things that I specifically care about. It is much easier to make multi-hop connections in knowledge when one is really familiar with their local knowledge cluster. I, for one, find an absolute knowledge graph of everything quite intimidating and feel lost in the sea.

LearnAwesome looks interesting. Good luck!

I see that this is the author of the excellent "Quantum Country" so perhaps the case being made here is obviously biased, for a good reason.

I get a sense that the author seems to argue that most book authors are not conscious of the working memory of the reader. I don't think that is true. Many "good" non-fiction books I've read actually rely on space-repetitions throughout the text quite heavily. The author, to their credit, does note this for textbooks. Often concepts in long texts are referenced in a non-linear manner, either in expectation of ideas to come or in retrospect to draw parallels & new connections. Unfortunately, the onus here is on the reader to realize and extract the maximum juice out of this non-linear flow.

The key idea of this post is to shift this burden to the "medium" and not the reader. Books can't handle this responsibility in their current form. I certainly agree with this. Distill [1] is pioneering this from the perspective of medium-reader engagement. Unsurprisingly, Michael Nielsen, a collaborator of the author of the post, is a member of the steering committee at Distill. :-)

I love these new directions in dissemination of ideas!

[1]: https://distill.pub

Interactive articles is one of those things that everyone wants to do but very few have the time or the incentive. Nevertheless, I think

a. publications like Distill have a solid place in the future of research dissemination.

b. this will not remain a niche but be more widely adopted (further in my concluding paragraph).

c. interactively will allow the reviewers and readers connect better to the key ideas in a submission and feel more confident in accepting good work, however crazy it may sound.

d. in the same way a certain language style hinders our thought, our limitations in expressivity via PDFs can be overcome.

e. transparency + interactivity of results should notably improve the arguable "reproducibility crisis"

My apprehensions are however with regards to the tooling. Research visualization use-cases will almost always tend to be ad-hoc (well, because it is research). It is extremely hard for me to imagine a set of tools that will be general as well as easily usable. There has to be a trade-off. I always think about TikZ [1] or even matplotlib [2] in this regard. Both of them are extremely powerful systems but often find themselves struggling in the trade-off between ease of use and flexibility. I don't think it is the fault of either systems themselves, it is just the nature of this kind of work. Although, I'd be remiss if I don't mention the advances Vega [3] has brought. In general, I think we do have reasonably expressive systems available today for interactive content, only that custom use-cases require a non-trivial effort to get meaningful results, as the article notes. As a result, I do not think better tooling is what is going to shift the needle here, unlike what the article notes in the challenges section.

The change has to come from the way incentives and incentives alone are structured. I care more about investigating deeper into a line of research rather than disseminating research in the most accessible manner at every tiny milestone (a published PDF paper). While I care about disseminating research, I care more about getting all the "insider knowledge" about a method before I go around acting as the evangelist. And this takes time, as any researcher would acknowledge. This is what motivates researchers in the first place. The dissemination is a less systematic part of science and involves people problems beyond any single researcher or organization. Consequently, it is also the part where researchers are least invested in.

In this sense, I think interactive publications like Distill will evolve to be more like Foundations & Trends (FnT) [4] than your usual fast-paced conference with bleeding edge ideas. FnT is extremely well-written, broad scoped and I've enjoyed every issue I've read. There is a market for something between a broad survey and a deep reference textbook. Like both, it takes years to distill knowledge to be accessible. Distill-like publications are poised to fill this gap. This is certain to be much more than just a niche art. But hoping distribution as widespread as cheap (in terms of the writing effort) PDFs seems like wishful thinking. Of course, this in no way means the pursuit of such idealized systems is in vain as history reminds us every time.

[1]: https://en.wikibooks.org/wiki/LaTeX/PGF/TikZ [2]: https://matplotlib.org [3]: https://vega.github.io/vega/ [4]: https://www.nowpublishers.com/FnTs

Please change the title. It is misleading. This particular link is for cases where "I can’t access my account and want to delete it, what can I do?".

1. Anything and everything. It is often about my current research or an interesting tid-bit from math, philosophy, social sciences etc. It is more of a fluid knowledge base than a blog.

2. I commit to the source code a couple times a week, but keep updating posts whenever I find time.

3. I host the source code on Github and deploy on Netlify.

4. https://im.perhapsbay.es

> 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.

[1]: https://www.microsoft.com/en-us/research/wp-content/uploads/...

[2]: http://www.economics.uci.edu/~ivan/asmb.874.pdf

On the note of Parkinson's Law, here's some fun extensions: https://www.economist.com/business/2020/07/11/slackers-and-s...

An accessible version - https://outline.com/CbyPxD.

Sure thing! I only meant to imply the relative ordering of considerations.

I think the argument here is about pedagogy not performance.

Yes, I did mean dimension squashing in either direction - left or right.

Those are very important points. I've edited the parent to point out this clarifying comment.

> Rotation matrices just describe the unit vectors of a new coordinate system

More generally, _every_ matrix describes how a change of coordinate system should happen. For any m x n (m rows, n columns) matrix, each of the the n column vectors represent how the current coordinate system (which may or may not be unit vectors) should be represented in the new coordinate system (for a left multiplied matrix).

Whenever a matrix's determinant is zero, it means that you squashed some dimensions. As you can imagine, when m is not equal to n, there will always be dimension squashing. Even when they are equal, that can happen. If you take a 3 x 3 matrix and it transforms all the 3-D vectors into only planes (which are 2-D objects), the determinant will be zero. This would be stated as having a rank of 2. More simply, you'd say the "volume" of the transform to be 0 (because planes have zero volume).

EDIT: Make sure to read important clarifications by JadeNB below.

David MacKay was an absolute rockstar and this book is grossly underrated among beginners in machine learning. This should be THE complementary reference for anyone who uses the Bishop book. Both these books follow a philosophy which some people may not completely agree with, that from the "church of Bayes".

My guess is that information theory went through its hype phase and much of the ideas are so pervasive across real systems that people forget how important those connections are.

To tell you its importance, skim this work on information-theoretic probing [1]. I find this so satisfying. Its most famous alternative, linear probing, always felt inelegant. This paper experimentally shows how terribly linear probing fails.

[1] https://arxiv.org/abs/2003.12298

Funny, the website aims at "Design Thinking" and fails to understand some basic needs right on the front page. The collapsible "Goals" section has arrows flop-sided. They point upwards when the goal is collapsed and downwards when the goal is expanded. Seems like a 101 error to me.

And before someone points out, I understand design thinking goes beyond pixel design.