Hacker Read top | best | new | newcomments | leaders | about | bookmarklet login

s/almost at random/via machine learning/


sort by: page size:

ML can be used as an unbiased random number generator, unlike humans.

With random sampling?

If you can build a simpler implementation just using random data can help a lot.

I I try to type randomly without checking the prediction, I get 67-70%, if I check it, I can easily get it to 46-50%.

It is really interesting.


This is possible, and in fact probably implemented in some probabilistic programming languages, but I think you are looking at the wrong direction.

The point is that even for fairly simple real use cases, the computation complexity is so huge, that all computers in the world couldn't compute it in your lifetime if you don't employ some approximation or optimization and stick to naive algorithms.

So, that is what the whole field of machine learning is about: finding some clever ways to deal with random variables in a computationally feasible way...


Interesting. I don't think people realize just how slow rand() can be if it is called frequently in your c/c++ program. Marsaglia's xorshf is the fastest algorithm that I know of that also give a ok statistical quality.

That's true if the sample is random.

It's fisher-yates with multiple iterations, so should be pretty random

But the values are generally generated pseudo randomly by machine. This seems similar to the birthday problem, where the odds of encountering a value in a given range is higher than you'd expect.

that if you take enough small pseudo random samples, you'll get one far outside the norm eventually, by pure chance.

That's random, but you still need to analyze the output for a while to normalize it.

Sound a lot like Scott Aaronson's free will challenge.

A user types a 'random' sequence of Ts and Fs.

A computer can predict about 70% correctly though, by just counting 5-grams.

Here the taks is the opposite, make the prediction even easier.


A random selection from a set of in-series numbers is still random, it's just got a potentially known range.

Uneven distribution seems like a sign of a good random number generator.

If you're not interested in the tails, there's also always

(random()+random()+random()+random()-2.0)*sqrt(3.)

as a cheap (in terms of brain power) Gaussian (sigma=1, mean=0) rough approximation :-)


... not with 100% accuracy, but it's totally plausible that you can do substantially better than random (or a simple regex). So there is incremental value here.

I read a random sampling paper recently which took a simple problem and approached it in ways that were much more elegant than I did.

Its almost a toy problem but I found the paper really interesting:

http://citeseer.ist.psu.edu/viewdoc/summary?doi=10.1.1.138.7...


1 5 3 8 5 3 Is that a random number sequence? It depends where the data came from. Same goes for AI algorithms. Yes theres a risk of the data being biased, but the key is what goes in, not what comes out.

That sounds like the gambler’s fallacy. Less runs than what? Most truly random input haa far more runs than what people “think” is random, and in fact that’s one of the statistical tests for whether a data set was random.

You’re essentially saying that a good neural network can predict the next value of a good random number generator. Good luck with that one!

Maybe while you’re at it, have neural networks invert cryptographically secure hash functions :)

next

Legal | privacy