TTPrograms

Seriously, as soon as he said "flat prior on theta" I had huge alarm bells go off. Garbage in garbage out.

The main issue there is just that if you see something in operation sufficiently different from training you may violate those penalties. Eg. you can train an arm not to hit a person in simulation by penalizing it, but that doesn't guarantee there's not an input that would still cause the safety violation. Generalization in these regards can be still be shockingly bad for modern approaches.

It's a specification of essentially a complex graph of mathematical operations. If there's a function called

  def mult(a,b): return a*b

it's not much more informative to write:

  def mult(activation_a, activation_b): return activation_a*activation_b

Many of these functions are not much more complex than that, and the names along with their comments are more than sufficient given familiarity with the literature. If you think familiarity with the literature is unreasonable, it's still not clear what could improve code like this in reasonable space. "This is a linear function, which means that it satisfies f(x+a)=f(x)+f(a)"? "This is the attention head, it acts as a mask on the sequence input"? It would be like complaining that someone made a tree class and didn't put a comment explaining what a leaf node is. Code readability always assumes some reader context and minimum pre-existing knowledge (as do all forms of technical communication).

Ez: https://news.mit.edu/2011/trillion-fps-camera-1213

Or most of the rest of the stuff from that group: http://cameraculture.media.mit.edu/research/

I don't think this was one of the most well-known or highly-praised projects before this scandal. Certainly I can think of at least 10-20 other projects from the same time that were more well-known to me.

lol this is around 1% of the projects going on at any given time in the Media Lab.

The absorber/emitter assembly doesn't really resemble a solar cell. Demonstrated tungsten emitters have exceeded 1500K - it's not really that wild, it's what's in incandescent light bulbs. You concentrate sunlight typically.

One sketch of such a system is to physically mate an effective high T solar absorber to an effective narrow spectrum photon emitter (as described in this work). Then that can be coupled with a typical solar cell with bandgap matched precisely to the emitter wavelength. So your solar cell is near ideally efficient for the photons it receives. As I understand the emissivity of these materials can exceed blackbody radiation in the near-field but not the far-field (or something like that? Search "Superplanckian emission").

The devices as I am familiar are often called thermophotovoltaic cells: https://en.wikipedia.org/wiki/Thermophotovoltaic

See eg http://xlab.me.berkeley.edu/pdf/259.pdf for a great overview, esp section 3.3.

I saw this video recently that I found accessible from an undergrad physics background and got me interested: https://www.youtube.com/watch?v=XnVVyTD7CzM

They must be restricted to linear models, right? So this is basically an implementation of logistic regression? Doing nonlinear operations optically is much trickier, and almost definitely active instead of passive.

EDIT: Actually they describe some nonlinear material inclusions in the paper, interesting.

If you had two job offers, one salaried at $100,000 and one at $100,001, I assume you would pick $100,001 with zero consideration for any other factors?

Attractiveness is not a one-dimensional scale, it is a two-party phenomena.

None of those markets are really zero-sum. People value different things in all of those markets (jobs etc). Zero-sum is an edge case.

That's not really a distinction. Kin selection is a phenomena by which genes can propagate despite them not benefiting (or even harming) the individual that possesses them if it increases the probability of reproduction for other individuals possessing the gene (eg. if they are likely to be related).

It's evolutionarily plausible and explains a massive range of traits and behaviors up through full altruism. See eg. "gay uncle" hypothesis: https://en.wikipedia.org/wiki/Biology_and_sexual_orientation...

I would describe kin selection as a phenomena that is caused by natural selection because natural selection is a phenomena describing propagation of genes and not propagation of individuals. This is a subtle point and often not taught unless taking specific relevant coursework (I learned about it in an animal behavior class).

I am likely using slightly incorrect phrasing and evolutionary science experts may phrase things differently, though I am confident in the essence of this argument.

That sort of behavior would be expected for even trivial organisms according to kin selection, though.

Watch out with your eigenvalue analysis - your covariance matrix is limited to rank 500 because you're only using 500 images. Furthermore, the smaller eigenvalues will have eigenvectors with high variance due to the random image sample (ie. "noisy"). If the singular values start to fall off within ~1-3x of your sample size then that's indication you need larger sample - eg. random matrix spectra will still exhibit rolloff in spectra.

A good engineer hits the "physics limit" brick wall pretty early in the design process...

I would think it more likely that they thought they could change FCC regulation via lobbying efforts in time for release.

The article explains that this is about better forward models for the measurement process rather than priors on body structure. This makes sense for systems with more complex measurement physics that common reconstructions approximate (eg. CT, PET).

Wtf is this "zoominfo" site that, upon clicking "Read More", tries to drive-by download an exe onto my computer? What is this, 2002?

How do you einsum convolution? Arguably the single most important linear operation in deep learning?

"Is advanced machine learning a very complicated way of reproducing the results of trivial forecasting methods? In many cases it appears this is the case."

If by "many cases" you mean "the one example I looked at which was carried out by an undergrad" than sure. That sort of phrasing really gives away the author's bias.

Similarly I think the theory developed in the Three Factors paper predicts this scaling law, might be worth citing: https://arxiv.org/abs/1711.04623

For those wondering, I just clicked through a few randomly and I only saw ~1/10 from China, and it was fairly applied (i.e. not especially novel research). I'd be interested in a more thorough analysis.

To be clear, white bread is thousands if not tens of thousands of years old. I don't see how the 300 year old baguette could be considered "not real".

When talking about "convex optimization" one nearly always means that both the function and the domain are convex.

1) Just because you can prove convergence for convex functions does not mean you can't prove it for non-convex functions.

2) This is "a way" not "The only way".

(If A then B) does not imply (if not A then not B)

If AMD developed CUDA transpilers / DL backends / CUDNN equivalents people would switch over in a heartbeat. It's fully their fault for not investing in the software.

If they did that and had a card that got 2x performance / $ or more I would switch in a heartbeat.

This is really fantastic. I just managed to find a 3x speedup on a compute heavy job I run using Py-spy - there was an unneeded hotspot in a library I'm using that I didn't previously suspect. It would have taken a long time using kernprof to dig in through the call stack to find the issue.

There was a Chinese decentralized exchange that (foolishly) decided they would include ERC20 coins for trading according to which had the most transfers from unique accounts to the exchange (I believe?). This incentivised individuals who wanted their coins included to distribute them among as many sock puppet accounts as possible and transfer them separately to the exchange. The resulting transaction traffic had a non-negligible impact on network capacity. In cases of load like that (or many individuals with strong incentives for their transactions to get in) people pay higher fees to increase the probability that their transaction is included in the next block.

For practical interaction with the network this can be drastically reduced by using light sync, of course.

My statement is not that people can't benefit from a PhD, but that it is not a strict requirement. You are presenting a strawman.