thisisdave

Not a great example. Apple's graphics layer on OS X is very closely tied to PDF, and basically uses the Adobe Imaging Model that PDF is based on. It actually does include a PDF-based scripting language for drawing and updating shapes on the screen.

To some extent, this could be addressed with standard error detection/correction techniques (e.g. error-correcting codes). But that would make the strings longer and often wouldn't help until they made it to a computer anyway.

The suggestion I've heard is that the xkcd style is a good way to signal "this is just an informal whiteboard sketch, not a highly-refined analysis". I think there's a place for that.

In both models, there's an unobserved state that changes over time according to relatively simple rules, and you get indirect information about that state every so often.

In Kalman filters, you assume the unobserved state is Gaussian-ish and it moves continuously according to linear-ish dynamics (depending on which flavor of Kalman filter is being used).

In HMMs, you assume the hidden state is one of a few classes, and the movement among these states uses a discrete Markov chain.

In my experience, the algorithms are often pretty different for these two cases, but the underlying idea is very similar.

legalized theft [0]

[0]: http://www.wsj.com/articles/SB100014240531119034809045765122...

How? With a bad review, or some other method? I didn't see anything on the App Store page, but I might have been looking in the wrong places.

Not to be confused with the Eddington number in physics[0]

"I believe there are 15 747 724 136 275 002 577 605 653 961 181 555 468 044 717 914 527 116 709 366 231 425 076 185 631 031 296 protons in the universe and the same number of electrons." --Sir Arthur Eddington

[0] https://en.wikipedia.org/wiki/Eddington_number

My understanding is that this is based on the policy network output. If I understand correctly, the policy network is designed to estimate probabilities accurately, although I don't know how accurate it is for low-probability moves like this one.

Or to wipe the phone if it's updated without being unlocked.

NSFW|ADJ does work, though. Maybe NSFW is a reserved word of some kind?

And don't interfere with the "back" button, either. Scrolling back to the previous location isn't what anyone intends when they press it.

It's easy if your data is already in two columns for x and y. If it's a raster, then you have to remember the right command from reshape2 or tidyr or write one yourself.

So, it's an edge detector? What makes it better than existing methods?

>The S is for "stochastic" -- i.e. you get a different 2D projection every time you run it on the same inputs.

That's not the part that's "stochastic"; sensitivity to initial conditions is just nonconvex optimization in action. You get the same thing with most other local embeddings.

The stochastic bit is that the model is based on optimizing "the asymmetric probability, pij , that i would pick j as its neighbor"[0]. Those probabilities and the associated positions in 2D space are not estimated stochastically (e.g. with Monte Carlo sampling) or anything, though.

[0] https://www.cs.nyu.edu/~roweis/papers/sne_final.pdf

A one-watt laser is pretty powerful (Class 4). Most laser pointers in the US can't exceed 5 milliwatts. One watt is several times what's needed to light various common materials on fire.

https://www.youtube.com/watch?v=woiTedSKPrk

>the class of models Stan fits is far more general.

This is mostly true, but last time I checked, Stan still couldn't sample discrete variables like BUGS can. Stan can only fit models with discrete parameters (e.g. finite mixtures) if the programmer is smart enough to integrate them out.

"distinctive characteristics" reminds me of "The peculiar institution of the South---that, on the maintenance of which the very existence of the slaveholding States depends"

http://users.wfu.edu/zulick/340/calhoun2.html

I don't think the commenter you're responding to was accounting for the possibility that the growth occurred in discrete jumps once per year.

> I don't think that most researchers would view writing a review article for Nature as "volunteer work"

I was referring more specifically to the fact that Nature had to enlist volunteers to referee the paper he wrote.

I was curious whether he was okay with that even though he wouldn't volunteer himself, if his position on non-open journals had changed, or if there was some other explanation.

Has LeCun changed his position on open access? He'd previously "pledged to no longer do any volunteer work, including reviewing, for non-open access publications." [1] A major part of his reasoning was, "Why should people pay $31 to read individual papers when they could get them for free if they were published by JMLR?"

I had assumed that meant he wouldn't write for them either (and thus wouldn't enlist other people to volunteer as reviewers when the final product would cost $32 to read).

[1] https://plus.google.com/+YannLeCunPhD/posts/WStiQ38Hioy

>I'm still pretty sure a mathematical mind on the order of Newton's could find a way to state evolution by variation and natural selection as a theorem rather than a theory, that follows necessarily from the laws of probability and the facts of chemistry

Apart from the chemistry part (where I don't understand what you mean), biology has had this for 85 years, with numerous extensions and generalizations [1].

[1] http://www.ias.ac.in/jgenet/Vol74Nos1&2/19.pdf

My understanding is that early humans were substantially closer to the equator than South Africa is, but you could be correct.

Perhaps they should have specified that they were referring to specific populations of humans that evolved in the last few hundred thousand years, then. Seven million years ago, our lineage hadn't even split off from gorillas.

>Seven million years of human evolution were dominated by two challenges: food scarcity and cold.

Almost all of that time was spent in Africa. Even during glacial periods, was it ever icy? I don't know much about the paleoclimate, so I'm not sure.

I'm having fun with this, thanks for posting it.

I'd love to see reports of basic genetic genetic properties of the population. What's the variance in fitness within each generation? This turns out to be the key factor in the rate of evolution [1], assuming that one defines fitness the way that geneticists do.

Also, estimates of the relative strengths of mutation, selection, and drift would be nice, as would options for distinguishing between hard and soft selection. These sorts of diagnostics and options could really improve the method's effectiveness.

[1] https://en.wikipedia.org/wiki/Fisher%27s_fundamental_theorem...

The argument is that single-layer networks would need to be impractically wide to have the same representational power as a narrower network with multiple layers.

I haven't looked into the math myself, but apparently there are situations where the increase in representational power for networks of a given size can be increased exponentially by moving nodes into higher layers.

This could be an important result, since there's a lot of theory and expertise regarding renormalization group methods, but I think it's too early to tell.

The write-up is a bit misleading, though: the model in their preprint[1] is about stacks of restricted Boltzmann machines (RBMs), and are very different from the other examples of deep learning mentioned. The Google Cat Detector model, for example, didn't describe a probability distribution over images, which is the kind of task that the preprint is about. And in almost all of the recent cases where deep neural networks have made substantial progress over the previous state of the art, the models have not been probabilistic, or trained layer-by-layer, or unsupervised, like the RBM-based approach in the preprint.

I don't speak physics, but my reading of the paper says that could be summarized pretty accurately as follows:

1. Hinton et al. (2006) showed that each layer in a stack of RBMs improves a variational lower bound on p(x).

2. Variational methods for RG also iteratively improve a variational lower bound on p(x).

3. The two methods would thus be equivalent, if we could fit them without error (which we can't).

4. Here’s a figure from a stack of RBMs that vaguely looks like RG results (not shown)

I don't see any comparison between the approximations that physicists normally use versus the contrastive divergence for training RBM-based networks, or evidence that their results are more similar in practice than any other technique.

Am I missing something?

[1] http://arxiv.org/abs/1410.3831 [2] https://www.cs.toronto.edu/~hinton/absps/fastnc.pdf

It was mostly true in 2007 [1]:

>Detroit just watched its last mainstream grocer, Farmer Jack, close its doors for good.

[1] http://www.npr.org/templates/story/story.php?storyId=1247787...

Right. I think that there's a big difference (legally and morally) between the beams of light passing between my cell phone an the cell tower, versus something that's being broadcast specifically so that everyone with an antenna can hear it.

An overlapping group of authors (compared with the ones that wrote the "intriguing properties" paper) recently published a neat preprint on using adversarial networks as part of the training process. The goals of the paper are different from the networks discussed here, but I thought it was really interesting.

http://arxiv.org/abs/1406.2661