cosmic_ape

The context is that various authors comment on Atwood:

>> The office of Jonathan Franzen: While he was pleased to receive your invitation, he must decline due to the fact that, although he is aware of Ms. Atwood’s love of birds, and he loves her for it, he’s never birded with her personally and can’t speak to any specifics regarding her conservation work.

Apple Health seems to allow exporting all the data as a file, including individual measurements from the watch (those come every second some times). Haven't tried it yet though.

I'd add V. Arnold's books on Differential Equations and Geometry to that list.

Wikipedia summarizes it nicely: "His views on education were particularly anti-Bourbaki." :)

I like the "Discipline is Freedom" on that guy's t-shirt. Its arguable, but its deep.

Containers were not supported on Android last time I checked, a few month ago. No containers on iOS also.

One more problem is that containers are an addon and do not sync between Firefoxes on different desktops. Still extremely useful though.

re:conferences: That's not quite right. In conferences you also meet people, some of whom you knew before and some that you didn't. You get to talk, and ask questions. The ROI of this may be questionable, but it does have value.

The course you refer to and the paper are somewhat different. The course simply says that for a given problem, there are good datasets and bad datasets, and the bad data may be bad no matter how big it is. Bad means biased here.

The paper considers a somewhat different situation. There you have a single dataset, for which you know in advance[1] that some points in it, and you know exactly which, are more noisy than others. The question is then whether to use these points, and how.

[1] But you know this because you assume a model. A particular timeseries model in this case. Its not necessarily in the data itself.

Staying in a hotel room and not going outside... But also telling other people that you are someone else, as he mentions doing. I can understand the phase, a kind of complete isolation. Not sure it should be instead, rather that in addition to, other kinds of activities.

I can relate to room service being annoying. Though leaving "do not disturb" on the door usually works, even for 5-6 days in a row. They do not knock when they see that. Just pester you at the corridors if you pass by.

What I find interesting is that gestures seem to be a part of the thought process itself. Its always fun watching people gesturing while speaking on the phone.

Return times of random walks are power law distributed in some graphs. For example return to an interval on the line.

If that's true, isn't there some organization in the government that can sue them for this?

>>UMAP has already demonstrated its efficacy as a tool in any data scientist’s belt.

Any chance you have some references? All their examples in the UMAP paper and in this talk look very toy-like.

There probably is a similar effect when people look at their own photos versus other people looking at the same photos. Wonder if anyone experimentally verified that.

Well, lets try. For the embeddings, you'll have to search the internet for details, and this may be not a "proper" ML. But the general idea is that if you have some nice euclidean algorithm, like k-means, and you want to cluster some data w.r.t a different metric, you could try putting the data in the euclidean space without distorting the distances too much and then apply the algo. there. For k-means in particular the nearest neighbor search in Euclidean space can be made efficient, I think. This didn't take off much because most of the results were unfortunately negative. Although there were some recent positive results, even with discussion here [1]. Bourgain had some early results on the embeddability of finite metric spaces into classical spaces.

The concentration of covariance matrices is fundamental for most "spectral" methods, in particular spectral clustering. See for instance [2]. People proved and reproved these concentration results a lot of times, but Bourgain's was the earliest with N\log N sample complexity for log concave measures, improving on a result of Lovazs et al. (!!!) More importantly, the method with which he proved this, controlling the concentration by decomposing into bounded and "rarely non-zero" parts is now all around in statistical learning theory. So each time people want to prove uniform law of large numbers and nothing standard works, people try to find a good decomposition.

[1] https://news.ycombinator.com/item?id=17752836 [2] Clustering with Spectral Norm and the k-means Algorithm, Amit Kumar and Ravindran Kannan (Section 6...)

I learned so much from Bourgain's work. While most of his work was in Fourier analysis, and later in number theory, he really was a very versatile analyst. Just two of his early papers, on embeddings of metric spaces, and on concentration for covariance matrices started whole fields later. They are very useful now in theoretical machine learning.

What I said is that "very few people cared" is a better description than a "25 year old problem". It is better because it better aligns with the facts presented in the Wired article above and other similar sources, due to reasons above. Nothing here is a guess.

I read the article. I see how much published work on the problem is mentioned, and where is it published. In this case, not much. If a mathematical problem is really hard, and really important, of the type to reasonably qualify as a "25 year old problem", you will see a lot of partial results, many different approaches, connections to other ares of mathematics, a lot of work by many people. Nothing like that happened here.

Describing this as 25 year old math problem is a bit misleading. There might have been two or three otherwise not noted mathematicians that have looked at the problem a couple of times after dinner during the last 25 years -- is a better description.

Well, p definitely is the solution in the case I mentioned. It is indeed the maximum likelihood solution. You could incorporate prior info about theta via a regularization term, if so inclined. What does not make sense in this?

Not sure what the OP meant, but I though it might be useful to mention how estimators may be interpreted as anything probabilistic at all. Often, arbitrary numbers between 0 and 1 are termed "probabilities", but in this case there actually is some proportion or probability to which f(x) should ideally correspond.

>>How do neural nets approximate a Bayesian posterior?

Not sure what GP had in mind, but if a feature x appears in a dataset n times, with pn times with positive label, and (1-p)n times with negative, and your classifier is f(x) which is trained with the "cross-entropy" cost, then the ideal value, that minimizes the cost should be f(x) = p. In this sense, f(x) is the probability of positive given feature.

Whether neural nets really realize this and how reliable that is, is another question. But that's the intention of the cross entropy cost.

Its quite important that its not so much about "non-convexity", its more about "random initialization". All these results, and there are others, more sophisticated, basically say that after random init everything is already good, and the gradient descent just has to train the highest layer, without screwing up the previous ones. They are all pretty naive that way.

you're right, fixed.

This is not necessarily a trivial fact, but I wouldn't call it incredible. It says a net trained with gradient descent can fit the input perfectly if its much larger than the input.

But this says nothing about generalization, i.e. performance on test set -- which is what we really want.

Yeah. And they require that the number of parameters grows at least polynomially with the size of input. In this regime, we also have another algorithm capable of zero loss -- the linear regression!

Exciting. And the cache size is likely the main reason. But that can not be concluded from the analysis.

>>VTune agrees, and says that Z3 spends a lot of time waiting on memory while iterating through watched literals.

But how much is "a lot" is not specified. And the propagation also allocates memory sometimes, to keep the learned clauses. Not sure how Z3 manages this, but couldn't it be that the mallocs are just slower on that desktop, so an OS issue?

This mentions ICA and NMF, but in contrast to those the proposed method is supervised learning, not unsupervised.

I'd suggest the authors try something like an autoencoder in the waveform domain. That would be a more close analog to the ICA methods.

previous discussion here:

https://news.ycombinator.com/item?id=16243067

>> The answer is that these three models are special cases of more general exponential Lévy models. Options cannot be priced with general exponential Lévy models using the traditional approach of the use of the risk-neutral density of the terminal stock price because it is not available.

Does this mean there is no hedging strategy in these general exponential models? My understanding is the Black-Scholes gives the price, such that if the price was different, there would be an arbitrage strategy (under some assumptions on the variance). And this arbitrage strategy is used for hedging.

In linguistics there is also work that attempts to understand the formation of abstractions from sensor-motoric concepts. Here's one example:

http://fas-philosophy.rutgers.edu/goldman/Spring%202008%20Se...

More like how to do a cheap philosophy PR stunt.