murbard2

82 ms·

871.
by murbard2 12y ago
I agree that Bayesian techniques can be costly. The argument I'm making is that Machine Learning techniques are best understood as approximations to Bayesian methods. Not for some math reason but for some epistemological reason. You li
872.
by murbard2 12y ago
Interesting, I've played with the SNAP datasets and did find very high eigenvalues, but I attributed it to the graphs bring a partial snapshot. You're right, 6 degrees of separation is a statement about diameter, but a small diame
873.
by murbard2 12y ago
The decision function can be seen as a distribution too. Say you're thinking of building a classifier, you can build the distribution over the domain x {-1,1} Give me f: A -> B and I'll give you P : (A x B) -> R+ See what I
874.
by murbard2 12y ago
And yes, you can try very hard to be a Popperian and avoid the problem of induction by inventing a contrived framework like PAC. Or you can embrace description length as a universal prior.
875.
by murbard2 12y ago
SVM induce an implicit probability distribution over labelled sampled of points, whether you like or not. Yes, "parametric" is often meant to imply a finite parametrization, but that's a dumb terminology.
876.
by murbard2 12y ago
Nuh uh, anyone doing LOCV is basically using AIC. There are also other principles, such as MDL which do not rely on a test set.
877.
by murbard2 12y ago
Because the prior on your parameters smooths out the prediction. Most cookbook techniques such as ridge regressions, cross-validation, etc have a Bayesian interpretation as a prior on the parameter. Bayesian techniques allow you to use all
878.
by murbard2 12y ago
All ML is density estimation using a parametric probability distribution. Some algorithm make the estimation easy and the distribution quite implicit, others make the distribution explicit but training is harder.
879.
by murbard2 12y ago
Bayesians don't need a test set 8)
880.
by murbard2 12y ago
One often neglected idea is transduction. The idea is that you can learn to label data better by augmenting your labeled dataset with unlabeled data. The unlabeled data helps discover the structure of the data which in turns helps the super
881.
by murbard2 12y ago
Indeed. However, it's easy to miss the "random-walk" aspect of pagerank if one simply think of it as a recursive relation between inbound and outbound links. Once you think "random-walk" you can build a whole class
882.
by murbard2 12y ago
This is established by making a one-step random walk in the friendship graph. Things get more interesting when you start taking longer walks. If you take a long enough walk in a (strongly connected) graph, the probability of ending up in a
883.
by murbard2 12y ago
No, absolutely not. I'm providing a very literal summary of the technique, as I understand it from the article.
884.
by murbard2 12y ago
Thanks, I suppose I should have been a little less terse. I suspect the downvoters didn't really get what I was saying. (Or dare I say, their reaction was... pavlovian). So yes, the approach described in the article is exactly the appr
885.
by murbard2 12y ago
tl;dr: operant conditioning. The method described is exactly how you train a puppy.
886.
by murbard2 12y ago
>No, you don't have to trust the miners because there are no known collusive schemes that would make more money for that volume of miners than playing by the rules would. Yes there is, coordinated conditional double spending on gamb
887.
by murbard2 12y ago
>In Bitcoin, you trust the way the network as defined by the code has been set up, not the miners specifically. Wrong, you trust the miners not to collude and attack the network. I know you've heard the claim over and over that Bito
888.
by murbard2 12y ago
Bitcoin also requires trust, specifically the trust that miners will not collude to double spend. There is also nothing fundamental about the idea of issuing coins as a reward for mining. It's just one possibility among many, there&#x2
889.
by murbard2 12y ago
Who is running the consensus nodes? Will people choose their consensus node, and if so can you guarantee an overlap? Are the nodes maintaining a state or a full, auditable, ledger?
890.
by murbard2 12y ago
Now look up Hidden Markov models, http://en.wikipedia.org/wiki/Hidden_Markov_model How they can be calibrated in the finite case http://en.wikipedia.org/wiki/Baum%E2%80%93Welch_algorithm And how
891.
by murbard2 12y ago
You actually do both. Each measurement induces a probability distribution over the space upon which you bruteforce. Also not that timing attacks are quadratic in the noise to signal ratio and linear in key length which bruteforce is exponen
892.
by murbard2 12y ago
For N large enough, there is no noise floor. In some cases, you could realistically be attacking a system for months or years to obtain a key.
893.
by murbard2 12y ago
The whole point is that even if you prevent them from buying high risk securities, they will buy AAA rated commercial paper from ibanks that will. And surely you're not going to prevent a bank from making loans, that's their who
894.
by murbard2 12y ago
>And yet we did see commercial banks taking massive risk and going bankrupt because of it The depositors were insured, so the depositors didn't care. The bank shareholders had a free call option from the federal government in the fo
895.
by murbard2 12y ago
s/can/may/
896.
by murbard2 12y ago
>Don't you agree the strategic outlook of an investment bank is entirely different than that of a commercial bank? I do which is why banks will specialize in different niches regardless of regulation. No law prevents a cement making
897.
by murbard2 12y ago
I did mean investment bank, thanks for correcting. How are investors putting pressure on a bank? Investors are the owners of the bank, they don't "put pressure", they manage their property. Investors will want their property
898.
by murbard2 12y ago
Please, explain to me what difference the Glass Steagall act will make when commercial banks can issue paper that will be bought by deposit banks? Oh, were you just repeating an Occupy Wall Street meme?
899.
by murbard2 12y ago
They'll have no problem making it past of the commerce clause. The law doesn't matter, politics do. The rule of law is not a self-enforcing principle, it's a political equilibrium built around a consensus. Checks and balances
900.
by murbard2 12y ago
If your tower gets exponentially thin towards the top, you can keep going forever as the pressure resting on the each level is independent of the height of the level. At least when the gravity field is approximately constant.

More ›