ludwigschubert

Gedy, thank you, I love the result! And that SunDial.svelte is <100 lines formatted, too! \o/

Man, those two comments sound uncharitable. In the first paragraph Andy states:

> This guide assumes basic familiarity with spaced repetition systems.

The words "prompt"/"spaced repetition" are basic jargon for such systems.

But to make this comment at least a bit helpful, too: An example of a "prompt" in a spaced repetition system is the front of a flashcard. It's "spaced repetition" because you look at them multiple times, spaced out over time.

Afaik they're unique—and not monotonously increasing in difficulty either. Here's Stanford's numbering system for reference: https://cs.stanford.edu/academics/courses

Great to see actual testing on a representative population sample, though I’m not sure what the consequences ought to be:

> These prevalence estimates represent a range between 48,000 and 81,000 people infected in Santa Clara County by early April, 50-85-fold more than the number of confirmed cases.

Looks like it was just published today! https://medrxiv.org/cgi/content/short/2020.04.14.20062463v1

Here’s the HN submission for discussion: https://news.ycombinator.com/item?id=22900730

How would one extract an algebraic function from the network’s weights? Why would the function the network approximates necessarily be algebraic?

Have you seen the YouTube channel Ctrl Shift Face?

https://m.youtube.com/channel/UCKpH0CKltc73e4wh0_pgL3g/video...

Here’s a pretty well done one in which Bill Hader does impressions, during which his face is replaced with the one of the people he’s doing an impression of: https://m.youtube.com/watch?v=kjI-JaRWG7s

If you're interested in Data Structures classes, I also strongly recommend Keith Schwartz's CS166. Unfortunately there are no public video recordings afaik, but slides and handouts are available, and also of exceptional quality: http://web.stanford.edu/class/archive/cs/cs166/cs166.1196/

Can anyone see a reason why they wouldn’t just allow you to provision (and pay for) a persistent Google Cloud VM instead? (I currently do that manually and need port forwarding to a machine that runs Jupyter.)

It’s hard for me to understand why Colab would build such a vague pro tier instead of the simplest possible solution: let me pay for my compute.

There’s so much more potential, too; they could offer whole clusters on demand, with really simple Python integrations say using dask, or ray.

I think that’s a _really_ fair question! xD

I’ve heard “mixed reality” being used to refer to the continuum between reality - augmented reality - virtual reality. In the above question I’d read it as asking about the state of problems particularly pertaining to the transitions along this continuum. (Made up example: of you solve locomotion in VR by putting people on moving belts, a problem in MR would be how to transition someone from walking around in unobstructed reality to walking in VR. I don’t know if that’s a real problem—please take it just to illustrate what I mean.)

Edit: Wikipedia’s article on MR includes a whole section on “differences in terminology”! \o/ https://en.wikipedia.org/wiki/Mixed_reality They also use the word “continuum”, but don’t seem to directly back up my further claims about using the word to refer to problems arising specifically at the interface of AR/VR.

The UC budget 2019–20 suggests:

68% staff + faculty salaries and benefits 15% financial aid 17% equipment

From page 7, https://www.ucop.edu/operating-budget/_files/rbudget/2019-20...

If someone knows a more detailed breakdown of the budget, I’d be curious! :)

In Germany they certainly self-identify with that name. Unsure why the website is down these days, but here's an archive link: https://web.archive.org/web/20160304102539/http://www.antifa...

...and many would say that’s because us humans are bad at imagining optimizing agents without anthropomorphizing them. This is a reasonable, even typical suspicion that many people share! The best explanation I know of why it’s unfortunately wrong is by Robert Miles in a video, but if you prefer a more thorough treatment, you could also read about “instrumental convergence” directly. If you find a flaw in this idea, I’d be interested to hear about it! :)

Robert Miles’ video: https://youtu.be/ZeecOKBus3Q

Instrumental Convergence: https://arbital.com/p/instrumental_convergence/

Now afaik nothing in this argument says that we can’t find a way to control this in a more complex formalism-but we clearly haven’t done so yet.

Hi! I help maintain Distill's infrastructure. We'd love to find a robust solution for archival, too, but we haven't been able to think of a satisfying automatic way of providing static fallbacks to interactive diagrams. (That's why index.archive.html isn't a public feature ^_^)

If you have an idea of what we could do on our side with a fixed amount of work, please let me know! Unfortunately we can't require our third-party authors to be experts in graceful degradation. I'm sure better tooling could help, but currently have no specific inspiration.

As a fallback for your specific use case—we do actually provide a tiny bit of print styles in our CSS. Try your browser's PDF conversion/print feature; I find it often produces acceptable results. (Minus proper alignment to pages, of course. :/)

I’m aware this has become a cliche on HN, but I’d be happy to pay that—if it helped align incentives better. Maybe if a company as big and ad supported as Google would make that step, it could become more accepted to pay for software again?

The usual argument goes: consistent user experience means adhering to your users’ expectations, which are shaped by their platforms.

...we all said, and then ML became popular. :D

(Lowercase ‘θ’ is quite common to indicate a vector of parameters. I’ll admit, though, that few people are ~courageous~ adventurous enough to use the actual symbol in their variable names rather than the string ‘theta’.)

3rd party vouch for shameless plug: I was a RA in a minor role at this team and these are some of the friendliest and most approachable people I've had the pleasure to work with in academia. (Hi Elliot! \o/)

Overlaps on widths >1500px.

(My recommendation would be to not make this a background image, but simply an image in the markup after the headline.)

Grid is not meant to replace Flexbox. I like to think of Grid as CSS’s 2D layout system, while Flexbox is its 1D layout system.

I do agree that it seems pointless to adopt a Flexbox based framework to do what is more clearly expressed in Grid. (People often like using frameworks for the pre-styled components, so that may be one factor...)

We need to go from it being special to open source code for a paper to making it almost seamless to interact with a presented idea, to try it in new contexts, to build upon it, and, yes, to falsify it.

When we introduce examples into an article we do so to make it easier to understand our message, but this opens us up to criticism of cherry picking our examples. We take a first step by allowing readers to choose from a set of examples, but real progress is open sourcing the code behind each diagram, and making it seamless for readers to run.

If you primarily care about finding truth, then you don't only need to make reproduction and falsification possible, you need to make it easy.

You're right: we are taking the mean of the activations of a given channel `z` over all its `x,y` coordinates. (We could sum, but we use mean so that step sizes are comparable between channel and neuron objectives.) Thanks for the feedback that this notation is not super clear, we will consider rewriting those expressions.

When we do feature visualization we do start from a random point/noise. For the diagram showing steepest descent directions, however, the gradient is evaluated on an input image from the dataset, shown as the leftmost image. There's no real step size either as we're showing the direction. You can think of the scale as arbitrary and chosen for appearance.

Section numbers are on their way—and figure numbers also sound helpful! I've added a ticket. (https://github.com/distillpub/template/issues/63) For now you can already link to figures like this: https://distill.pub/2017/feature-visualization/#steepest-des...

Thanks for the heads-up! We are not sure there is a standard for this, but we now include the DOI in our bibtex citation.

In "Examining the Internals of the Model" it could be interesting to show an overview activation map that sums over the absolute activation of all the examples—to highlight which cells activate on all examples—and one that is colored by variance of activation across examples—to show which cells influence the character/differences in writing.

This is incredible important research. There are fields in which adoption of ML techniques is dependent on it being explainable and inspectable to stakeholders, e.g. in certain health care, policy or finance applications.

There is work being done on explaining non-linear models, such as LIME [1], but much more is needed. At a recent d3 meetup the topic was discussed but we primarily noted its importance. Notes at [2].

[1] https://github.com/marcotcr/lime [2] https://docs.google.com/document/d/1o-DO65PEZShLuadyRd35PsTJ...

Oh this is a fun rabbit hole to go down:

Gauss indeed wrote "methodisches tatonniren"[1], the latter of which is not a common German word, nor does it suggest grouping. I believe it to be an expression used at the time which was borrowed from French "tâtonner", which means "fumble", to describe a step-by-step process, "approaching" a solution. The closest I get in modern German is "herantasten", which in turn does not have an elegant English translation; you will have to take its elements "heran" (onto) and "tasten" (touch/feel/grope) individually to judge how close it is to fumble.

So, methodical groping is not as far off as you might have thought!

[1] https://books.google.com/books?id=3jEDAAAAQAAJ&pg=RA1-PA106&... This may be a typo—the regular verb form would be "tatonnieren"—then again a lot of German spelling has changed since the time of those letters.

...and on the wrong side. ;-)

That's a good point.

I guess I am trying to say that a map that accurately represents the area of seas is potentially not what everyday map users want from a world map.

But I see that they optimized for that instead of my vague sense of usability. :-)

I get what you're saying, and you are right, of course.

I'm trying to think of this not from a prescriptive position, though, but thinking of "why do people like Mercator". And I believe that people like the perceived efficiency of Mercator, in the sense that you can make a world map that fills more of its area with land mass.

Or, addressing your comment more pointedly: yes, yes you can. You can ignore the Pacific Ocean on a world map most of the time. :D

While interesting, I'm not sure the actual map as pictured is a "good design":

One of the strengths and success criteria of the Mercator projection is that it allows us to largely ignore the vastness of the Pacific Ocean on our maps. Notice how much of this map is taken up by it. While interesting from a planetary perspective, it simply isn't that important to everyday life of most map users.

The proposed design thus has too low a land area to map area ratio to be very useful.

The projection itself would be useful as a digital map (with zooming and panning) or as an icon, though.