HNHacker News
TopNewBestAskShowJobs

srean

8,900 karma · joined October 4, 2010

cat > Bourne.identity #! /bin/bash echo "$*" ^D
submissionscomments
srean··on English: A vs. An
No sarcasm intended, I was being literal. Me and my wife have different mother tongues. What's natural to her isn't to me and vice versa and we keep telling each other that it's easy.
srean··on English: A vs. An
Oh it seems you took it personally, whereas I mentioned "every native speaker". My intention was not to rule you, just to underscore the point that native speakers have it easier and may not be able to experience or understand the difficulties of a non-native speaker.
srean··on English: A vs. An
If you ever feel inclined to learn Hindi there is more to grapple with. The verb gender is sometimes set by the gender of the object, sometimes by the gender of the object. Which way this will go loosely depends on tense.
srean··on English: A vs. An
> As a native speaker of a language with genders, it's not complex

Said every native speaker I suppose.

srean··on Why do we need human mathematicians anymore?
> There are few fields as elitist and hostile as pure mathematics

Is that so ? Sounds hyperbolic.

srean··on Why do we need human mathematicians anymore?
> Why do we need any job?

If not for nothing else but to form a basis for how to distribute/share/hoard the wealth created by a society. The eternal question - who gets what.

srean··on Seeing Circles, Sines, and Signals
Understanding how complex generating functions help with combinatoric counting makes one's head spin.

(Tenuous pun intended. It keep track of how spinning arrowheads align with or cancel each other)

srean··on Weeping whales: Stillborn humpback whale grieving documented
There's Descartes dog too.
srean··on Seeing Circles, Sines, and Signals
Imagine a civilization that discovers complex numbers before naturals, with naturals showing up to count the turns.

If hexagons can claim to be bestagons complex numbers have a legit claim at best numbers.

srean··on There's no point at which turning your brain off will work
> The real question is who can sell their photos and for how much profit

Is that so ?

srean··on There's no point at which turning your brain off will work
Maybe so, but for the good shots by the thinking guy, I suspect he would derive longer term emotional gratification from how well they came out. Effort and investment adds to the experience.

Making things easy is good as long as it is not too easy, then it's Spotify muzak.

srean··on Human brain is two separate organs, research finds
That sounds about right.
srean··on Two parallel neural ectoderm progenitors contribute to the developing brain
So hindbrain is the OS kernel and forebrain the application server ?
srean··on Saving Jet Fuel
Is it worthwhile para-droping or glide-dropping excess fuel within airport open grounds reserved for that purpose.

I understand that floating into the main theatre will be a serious problem. Current weather should allow making those calls.

There's also the aspect of saving money by purchasing fuel where it's cheap and having internal fuel reserves low enough at such airports so that it can be bought in significant quantities.

As you can see from my comments, I am a complete newbie, just a curious one.

srean··on Inverse-Square Law
Ah! I was talking about the abstraction of point masses, then there is the shell theorem. But now at least I understand the discrepancy.

In any case gravitational potential at infinity is formally defined as 0 in Newtonian physics.

srean··on Inverse-Square Law
But what's the integral of 1/r^2 from 0 to \infty ...that's the part not that not squaring up.

Newtonian gravitational potential of a point mass at origin of mass M is - GM/r so at r = \infty the potential is zero. Well, partly just by definition.

I believe we are assuming different models and I want to understand yours.

srean··on Inverse-Square Law
I think he wanted to get the cuteness of the Archimedean result off his chest. Very understandable emotion.
srean··on Why don't machine learning research agents overfit?
Seems I was misremembering the dates. The Langford and Banerjee papers/turorials I had in mind were

On Bayesian Bounds https://dl.acm.org/doi/10.1145/1143844.1143855

Tutorial on Practical Prediction Theory for Classification https://jmlr.csail.mit.edu/papers/v6/langford05a.html

The first one is quite in the same spirit that you like.

srean··on Why don't machine learning research agents overfit?
Yeah, I had read one of your parallel comments and understood what you had meant. Differential privacy is a good formulation (well, the only one I know) to deal with the peeking problem in general.
srean··on Why don't machine learning research agents overfit?
You can try this one

https://books.google.co.in/books/about/User_friendly_Introdu...

Free download here

https://arxiv.org/abs/2110.11216

One of the ICMLs had a nice tutorial by Langford and Banerjee on the relationship between the different style of bounds. 2003, I think.

srean··on Inverse-Square Law
I should have said that it rises to zero, assuming standard sign conventions.

The gravitational potential infinity of a point mass at the origin is defined to be zero.

Could you expand on why you think it approaches a constant, presumably a non-zero constant. In any case absolute values do not matter, only the potential difference matters, but am still interested in your thought.

srean··on Inverse-Square Law
In a 3d world one can escape from the Earth's gravitational field. On the flat-world equivalent there is no escape. You can checkout but never leave. There ought to be a science fiction story based on this premise.

Things get interesting in 2d flatland.

There the field or force has to decay as 1/r because the circumference of the boundary scales as O(r). But that's the field, to get to the potential you need to integrate and then you get a function that is logarithmic with distance.

A consequence of that is the potential does not drop to zero as you go further away. Unlike what is the case for the 3d case.

If you have heard that about a random-walking bird that returns infinitely often but a random-flying bird one that doesn't, that's sorta related.

Totally unrelated, was quite thrilled to see this map of the sphere

https://substackcdn.com/image/fetch/$s_!8JEP!,f_auto,q_auto:...

because I had been entertaining myself by making toy globes out of paper and it seems this polyconic map was the one I had used (an interrupted version of this).

The more conventional way is to use gores using interrupted sinusoidal that look like a string of lobes connected at their common equatorial hip.

https://www.wolframcloud.com/obj/resourcesystem/published/De...

What I was working with were more like flowers, one for each hemisphere, with the pole at the center.

srean··on Why don't machine learning research agents overfit?
You raise a good point.

I had not considered the pedagogical value of starting with PAC Bayesian bounds. It will be interesting to see ML courses that take that route.

Similar issues show up in traditional statics when considering multi-hypothesis testing. I wonder if betting/gambling over hypotheses might be a more accessible way to convey these ideas to new students.

srean··on Why don't machine learning research agents overfit?
What threw me off was the (possibly misunderstood) suggestion for minimizing the generalization bound over the prior after the data has been incorporated.
srean··on Why don't machine learning research agents overfit?
"Can be misleading" is an accurate characterization.

It is an uniform bound and will have problems with very large hypothesis classes but its statement isnt wrong (just that the tool is a little heavy handed). For simpler models they are adequately useful. So I would suggest new users to learn if they want to use simpler models rather than deep learning methods.

Real world isn't the adversarial/ worst-case that these models of generalization assume. So the generalization performance you experience is usually a better than what PAC indicates, but estimating that experience was not PAC's charter. It's charter was to quantify a adversarial/legal guarantee, the minimum (infimum) guaranteed program.

I too like PAC-Bayes a lot (see my other comments) but it does have a cop out in that by definition you do not know how good your prior is, yet the quantitative estimate of generalization error depends on how good it was.

Nevertheless, PAC-Bayesian and statistical physics based bounds are closer to what we experience, but you can't guarantee that the worst case will not bite you.

srean··on Why don't machine learning research agents overfit?
> Then reduce the variance on a Gaussian prior.

Are you sure that doing this after seeing the data is valid and does not suffer from the equivalent of peeking-into-the-test-set problem ? There are ways to address the peeking problem but that requires additional machinery.

I don't dispute your broad claim but the first counterexample you quote seems problematic.

srean··on Why don't machine learning research agents overfit?
PAC isn't asymptotic in general. It gives finite estimates for finite sizes of training data.

If you say, and I think you are indeed saying so, that PAC is ridiculously pessimistic, I would be in violently agreement with you. That's one reason why for practical training data sizes and practical (infinite version space) hypothesis classes PAC gives bound such as -- probability of error is less than 41K. This isn't exactly incorrect but not very useful or informative.

A far more useful formulation is PAC-Bayesian where you get error bounds less than 1 guaranteed and usually less than 0.5 on reasonable sized training data sets.

srean··on Why don't machine learning research agents overfit?
> I can literally specify a huge mega-algorithm that just searches over all possible Turing machines and evaluates them

Then why don't you ? and did you mean all or those that halt ? I presume you have a way of separating those.

srean··on Why don't machine learning research agents overfit?
True.

It so happens that one gets the best generalization error bounds when one combines PAC with Bayesian ideas -- the PAC-Bayesian bounds.

srean··on Why don't machine learning research agents overfit?
Except for the fact that eventually we are all dead. So it is kind of important to get there faster.

For complicated hypotheses, where complicated is defined appropriately, it takes many many examples to realize that it was a wrong hypothesis all along. There lies the rub.

For a particular instance of a learning problem we can't tell much, however using a Occams razor over many instances, one would be correct more often than not. Provided, of course, the PAC assumptions are true or they are not very far from being true.

How far is not very far ? That gets very hairy to quantify.

← PreviousPage 3 of 34Next →