obastani

Google has also done some foundational research on learning with privacy:

https://arxiv.org/abs/1610.05755

https://research.google.com/pubs/pub45428.html

It's hard to know for sure what they (or any company) are doing in production, but I'd be surprised if they don't use these things internally.

It is need based -- it goes to the students least in need!

/s

This sounds like a really interesting paper -- do you happen to have a link to the reference?

Consider the following (traditional) program:

  x = raw_input()
  x *= 2
  assert x >= 10

Now we can ask the question: "Assuming the assertion passed, what can we say about x?" In this simple example, we know that x >= 5, but in general the possible values of x may be much more complicated.

This is the kind of question probablistic programming is designed to answer, except instead of being an arbitrary, unknown value, x is specified as some distribution (say Gaussian). Then, the question becomes, "What is the posterior distribution of x, assuming all of the assertions pass?" In other words, figure out how likely different choices of x were, taking into account both the prior (i.e., that x is Gaussian) and the new information from the assertions.

For a simple example of why this is useful, suppose we have a program that generates random images (this is our prior). We also have some real photographs. We can assert that the random image generator should have a pretty good change of generating the real photographs. Then, the probablistic program "execution" will try and compute a new generator that creates more realistic images.

I think they might be referring [1], which shows that adversarial examples transfer to models such as decision trees and SVMs.

[1] https://arxiv.org/abs/1605.07277

How do you know it's the same cat? A neural network needs to be explicitly told, "these are all the same cat".

The GP is saying that rational expectations is used not because it's simple, but because it makes the weakest assumptions about human behavior. However, I'm not sure this is entirely accurate since if some agents deviate from the equilibrium, there is no reason you couldn't end up with totally different outcomes (either better or worse).

A "gold rush" probably didn't make gold abundant -- it was just enough to make a few people rich.

The standard approach to detecting signal lights is to have a database of GPS positions of the signals, along with rough location in the camera where the signal is expected to occur [1]. Then, when the car nears the signal, it locates the signal and detects the current color.

This mechanism really shouldn't be susceptible to the same biases as humans. The described signal may legitimately be more challenging for the self driving car, but more than likely the signal was missing from Uber's database. Their lack of explanation for this failure does not inspire confidence in their approach.

[1] http://www.cs.cmu.edu/afs/cs.cmu.edu/Web/People/zkolter/pubs...

On the flip side, sometimes simple formatting in Latex can be a nightmare, for example, getting your table columns spaced correctly.

Word has its own version control built in, which works pretty well. You can see all revisions, and step through and accept/reject. You can also do diff of two documents.

I think one of the biggest challenges in moving from 1v1 to 5v5 is the substantially increased possibility of creative strategies. In a 1v1, there isn't that much room for innovation, making it a perfect target for existing AI techniques, which are good at learning what is "in the data" but have not yet been shown to be capable of thinking outside the box. In contrast, in a 5v5 matchup, a much wider range of strategies is available. I'd imagine that in this more flexible setting, AIs would be highly vulnerable to cheese strategies, or in general strategies they have never seen before. Furthermore, if there are any exploitable quirks in the AI (which seems quite likely, given the established lack of robustness of neural nets), I have no doubt that clever human players will be quick to exploit them. On the flip side, I think that makes 5v5 a far more challenging and interesting goal that would be very exciting if achieved!

As a side note: I think the most impressive aspect of this AI is that it was trained without watching any human games. In contrast, AlphaGo was bootstrapped from a bunch of existing human games, which I imagine was crucial to its training. Being able to learn how to fake out the human player without having seen this action before is quite impressive.

Seems like the foundation that published that study is anti wind and might not be the most reputable source [1].

[1] http://www.ewea.org/blog/2012/12/study-on-turbine-lifespan-j...

My understanding is that in a full match, AIs still have no hope against humans, since even though they can crush humans at micro, their macro is still abysmal [1]. I'm not aware of a match where any AI has beat a pro human player at Starcraft -- I'd be interested in learning otherwise!

[1] http://spectrum.ieee.org/automaton/robotics/artificial-intel...

Do you have a citation? As far as I know, existing AIs do not stand much of a chance again humans [1]. My understanding is that even if AIs can beat humans is small skirmishes due to superior micro, they lose so badly at macro that it simply doesn't matter.

[1] http://spectrum.ieee.org/automaton/robotics/artificial-intel...

Unless I am missing standing, the goal is only to produce a new interpretation of quantum mechanics. That is, the predicted physics do not change, only the formulation of the theory. Therefore, anything possible under this formulation is also possible under more traditional formulation of quantum mechanics (e.g., the many world's interpretation).

A sibling post talked about special relativity, but you don't need special relativity to answer this question, just Galilean relativity. The answer is that you can measure velocity in any inertial reference frame you want. If you are in an inertial reference frame, you can say your own speed is 0 km/hr, or you can consider an inertial reference frame of the galaxy, which is some ungodly number, or you can measure your velocity relative to the cosmic microwave background. It's all the same -- only the relative velocity between you and another object has a real physical meaning.

In this situation, wouldn't the person sitting in the middle be willing to trade for either the aisle or the window seat?

I don't think your comments really answer the questions that are being asked. The question, as far as I understand it, is why humans can't operate exactly as they do now, but without the sensation of consciousness.

I don't think the question is "hard", I think it simply does not have a good solution. Science can explain how things work, but not why. One day, we may even be able to have an empirical theory of consciousness, e.g., how manipulating various parts of the brain affect how we are conscious, what physical structures are needed to generate consciousness (or at least, beings that act as if they are conscious), etc., but we can never explain why we are conscious, i.e., why we have experience.

The attack you suggest is ruled out by differential privacy. The precise guarantee is a bit complicated. The first thing to note is that the output of a differentially private mechanism must be random. Then, the guarantee is that Pr[output] does not change by very much whether or not you are included in the dataset. In other words, even if you were omitted from the dataset, there the chance that the algorithm produced the same result is very similar.

This definition rules out the attack you suggest. In particular, if you are removed from the dataset, then the probability of the output (i.e., a ride starts in the region) goes from very large to very small. Therefore, the algorithm you describe (i.e., adding noise to the start location) is not actually differentially private.

The confusion arises because oftentimes adding noise is sufficient. For example, the average of n real numbers in [0,1] is affected by at most 1/(n-1) if you delete one point from the dataset. Therefore, you can just add a little bit of noise and the dataset becomes differentially private.

For the dataset you describe, a sibling comment proposed the correct mechanism -- you have to add noise to the count returned by the query, not the start location. (Technically I think you could just add noise to the start location like you propose, but the amount of noise would have to be large enough that all the start locations overlap by a sufficient amount.)

I believe most consumer industries have consumer surplus, since different customers are willing to pay different prices but the companies can only charge one price. This includes the taxi companies that Uber displaced. From reading the abstract, this paper does not calculate the additional consumer surplus generated by Uber beyond taxis.

Note that it doesn't follow that Uber would make more money simply by raising prices -- consumer surplus measures the amount Uber could make by charging each user their "true" value, but if Uber simply raised prices, then they would lose demand.

However, Uber is actually starting to use price discrimination (i.e., charge different users different prices based on their characteristics), which would eliminate consumer surplus [1].

[1] http://www.business-standard.com/article/companies/uber-s-ne...

Unlike adding RAM, adding each extra qubit is also exponentially harder, since maintaining coherence of all the qubits becomes more and more difficult. That's why scaling from the tiny quantum computers we have today (which are not useful) to a useful quantum computer remains a decades long research agenda.

I don't understand your comment. Given the same rate of return, investors prefer lower risk. So what you seem to be saying is conglomerates produce lower returns.

But I don't think that's what the conglomerate discount is, which I believe says that the components of a conglomerate are worth more than the conglomerate itself. More precisely, "Conglomerate discount is calculated by adding an estimation of the intrinsic value of each of the subsidiary companies in a conglomerate and subtracting the conglomerate's market capitalization from that value." [1]

[1] http://www.investopedia.com/terms/c/conglomeratediscount.asp...

I don't think what you are saying contradicts the gp. All the expertise that went into developing tools to host their own website probably went into developing AWS. Just because they maintained two separate infrastructures doesn't mean that it's not an example of turning an internal tool into a product.

Humans know when they can't see properly. The problem with adversarial examples is that the neural net thinks it's doing great. You have to manually specify every possible failure mode for the neural net before hand -- if you miss one, you might have an accident.

Are you sure you'd still feel that once you're actually there? What if many people might believe what you say, but then quickly change their minds as soon as they find out how difficult and uncomfortable it is to live on Mars?

Just because people collaborate from all over the planet to make the pencil, doesn't mean that's the way it has to be. This kind of collaboration simply enables economics of scale to kick in (and possibly improve quality as well).

From the context, I think it means that the 51 volumes (the ocean) contains very little content (the cup). I also don't think the grandparent comment reflects the intent of the original post -- prolificness is the number of attempted works, not the length of any single work.

I think this article is being somewhat selective in it's evidence. Several great mathematicians were highly selective in what they published, e.g., Gauss and Riemann. I don't think Einstein published that many papers. Though it might be the case that gauging quality is easier in math and physics, so being prolific is less important.

For example, because they believe in a cause -- this might still be pursuit of happiness in an abstract sense, but it's not pursuit of happiness in the immediate "I want fame/money to be happy" sense.