compbio

> If I have a pain in my lower right abdomen and I poll the human population, I'm going to get a better answer than a doctor?

You could. It is why people prefer a second opinion even when dealing with an expert. There are programs where everyone can join and try to diagnose disease. See CrowdMed (https://crowdmed.com) -- a YC company (http://venturebeat.com/2013/04/16/yc-startup-bets-crowdsourc...)

> If I poll the human population about a physics problem, I'm going to get a better answer than a physicist?

I think it was Newton who said that anyone could be as successful as him, but they just didn't think hard enough. But 10 smart physicists could probably think "as hard" as a single Newton. This means there is no need to chase the expert. This removes inaccuracies, since groups have more variance in opinion than individuals. It is harder for groups to be all wrong. Group intelligence benefits science.

> Conspire? Undermine the wisdom? Wow...

See it as game theory with possibly adversarial agents going against the group consensus. If there is no incentive to go against the grain, or if there is damage possible to your reputation (and thus energy), people will not disagree with top management or thought leaders. All errors trickle down into the group without filters. This is dangerous as it mimics group intelligence, but is really the opinion of a single individual inside a group of yes-men. This contributes to economic crashes and hype-y bubbles.

Blood-letting is a good example of variant group behavior. We need people "insane" enough to try that. We need to be wrong many times, for someone to say: "this is not right." and help fix it. Blood-letting may have benefits though -- being a nearly 3000 year old practice it would be rare to be completely without merit and survive.

> When it's wrong, somebody conspired, when it's right, oh, well, don't mention that it's the exact same process.

You want to label a spam dataset. You can use one person for this, or a room with a hundred people. When the one person is wrong, it may be because he has a different/quirky opinion on what it means to be spam. When the whole room is wrong, you start to wonder if the e-mail was really ham to begin with. One person can only be wrong or right. A hundred people can (and preferably should) disagree, while still converging to a better solution than any individual in that group could have thought of. Groups remove arguments from authority.

Slides: https://docs.google.com/presentation/d/13gDspOfYxB3z2_SIKne0... Blog: https://www.mattcutts.com/blog/lessons-from-early-days-of-go...

I do not think logic-based learning is dead. It just smells a bit funny.

In the vein of the papers "From machine learning to machine reasoning" and "Text understanding from scratch" I expect a "First-order logic understanding from scratch" to follow naturally.

I think those quotes were a bit snotty, especially considering your stature and knowledge on this. I don't think I was trying to rebut you, as much as I was simply trying my luck at advocating that user's view. Sorry for that.

I do not think there is a one-to-one connection to be made between neural nets and the brain. That said, it is possible to fool our own visual system in a similar way: http://en.wikipedia.org/wiki/Gestalt_psychology#Properties

If you haven't seen this picture before: http://www.moillusions.com/wp-content/uploads/photos1.blogge... it should appear as meaningless noise. After you spot the Dalmation dog, you will notice it in the future.

But maybe what they did today makes that impossible. They set the whole thing on fire, and now maybe nobody is going to trust it because they’ll think there’s some big evil vulnerability in the code. But now this decision makes me feel like they’re kind of unreliable. Also, I’m a little worried that the fact that we were doing an audit of the crypto might have made them decide to call it quits. - Matthew Green

TrueCrypt's developers chose not to graciously turn their beloved creation over to a wider Internet development community, but rather, ... to attempt to kill it off by creating a dramatically neutered 7.2 version ... - Steve Gibson

I have no idea what's going on with TrueCrypt, Speculations include a massive hack of the TrueCrypt developers, some Lavabit-like forced shutdown, and an internal power struggle within TrueCrypt. I suppose we'll have to wait and see what develops. - Bruce Schneier.

As far as I can gather the Dutch authorities are not able to retrieve TrueCrypt keys. I can not say the same about BitLocker and FileVault. If the US knows how to get into TrueCrypt volumes then at least they are not sharing that knowledge with foreign allied agencies.

Original article: http://googleonlinesecurity.blogspot.nl/2015/03/out-with-unw...

Do not be silly. Turning a bug like this into a commercial venture? How do you find customers fast enough before a patch? 6337$ for a days work is very reasonable, while selling a bug like this makes you not worthy of a quarter.

I now think Facebook is going to give us Artificial Facebook friends. Modeled on your interactions with your Facebook contacts, combined with the data of a million other connections, it will be more attentive, always available, knows just how to cheer up your mood, or find you things that interest you when they are going viral. This may be many years out, but Zuckerberg needs to think about this future too.

Something fell into place for me when I read this interview. Like when Google bought YouTube, it took me a while to realize that Google would own internet video. Facebook is sitting on a huge pile of social data. With social graphs and tagged faces the more obvious ones, they also have a huge corpus of natural language and personal likes. With this one could create AI that is much more down to earth and friendly to communicate with. It may posses a form of emotional intelligence. It may have knowledge on social interactions, status, moods, human relationships etc. it knows what makes us tick.

Contrast with Google which basically indexed the entire internet. I expect an advanced future interaction with Google AI to be more factual. The kid you want to do your homework for you, not the kid you hang out with later, who suggests new Jazz songs for you to listen to when it starts to rain outside.

Very interesting. Thank you for correcting me.

a relatively large fraction of examples will be misclassified by networks trained from scratch with different hyper-parameters (number of layers, regularization or initial weights). The above observations suggest that adversarial examples are somewhat universal...

I was not talking about overfitting. I've seen that paper.

The original paper asked if images that could fool DBN.a could fool DBN.b. The answer was: certainly not all the time. They used the exact same train set and architecture for DBN.a and DBN.b, just randomly varied initial weights. I think this is too favorable for a comparison with a voting ensemble made with nets with a different architecture, train set and tuning. Can they also find images that can fool DBN.a-z?

Also, to test if a net can learn to recognize these fooling images, they simply add them to the train sets. Those noisy images would be far simpler to detect: They have a much greater complexity than natural images. To detect the artsy images, a quick knearest-neighbors run should show that they do not look much like anything it has seen before, so it may be an adversarial image.

Deep nets are only loosely inspired by neurobiology. That's why LeCun calls them "convolutional nets" and not "convolutional neural nets" and prefers "nodes" over "neurons".

It is, however, possible to have a deep net produce 3D models/images: https://www.youtube.com/watch?v=QCSW4isBDL0 "Learning to Generate Chairs with Convolutional Neural Networks".

I also suspect a different part of cognition is used when humans are asked to recreate a "fire truck" than when humans are asked to classify a "fire truck" from a "car". The former seems closer to using memory ("what did the last five fire trucks I saw look like?"). A fairly recent addition to deep nets is making use of memory: http://arxiv.org/pdf/1410.5401.pdf "Neural Turing Machines". So the difference may quickly become less significant.

Now I have to try this out. My intuition tells me it becomes increasingly hard to create a fooling image, which looks alien, and is able to fool all the nets in the ensemble, even though they have different settings and params. I think they can only fool one net at a time, and have to get very lucky to be able to evolve the image for the other nets, while keeping the same classification. You can't "train" these images on all nets at once, by simply treating the ensemble output as a single net.

If your intuition is right though, then the ensemble may be able to counter with a random selection of nets for its vote: You'd need to evolve images for every possible combination and/or account for nets added in the future.

If we have a singularity AI then we don't need CAPTCHAs: For all intents and purposes an AI is equal or superior to humans, there is no distinction to be made. The Turing Test is a form of CAPTCHA.

Also, one could make CAPTCHA's incredibly hard. Humans and dumb computers won't be able to solve it, and singular AI's get a pass. So give access to anyone who isn't able to solve the CAPTCHA and redirect the singularity AIs to google.com :).

Finally, the singular AI could create a CAPTCHA which separates humans from AI. The problem of separating humans from AI will quickly become too difficult for humans to solve.

It is good to know that they need access to a lot of predictions from a net, before they can create an image that will "fool" the net, but look alien to humans. Secondly, this doesn't account for ensembling: "fool me once, shame on you. Fool me twice...". Since the images are crafted for a single net, a majority vote should not be fooled by these images. I suspect this effect rapidly goes away when adding more nets (which is basically industry-standard practice to increase accuracy).

Furthermore, I am seeing the security concerns, but I figure this is far from a practical attack. Deep Learning Classifiers do not act as gatekeepers: You have not much to gain from a single faulty classification. You won't be granted access to secret information if you happen to look like the CEO.

http://www.cs.cmu.edu/~clegoues/docs/legoues-icse09.pdf "Automatically Finding Patches Using Genetic Programming ... We introduce a fully automated method for locating and repairing bugs in software" http://dijkstra.cs.virginia.edu/genprog/

Though I think the author was talking more about coding up a neural net, where it can be said that the net learns the features itself, and optimizes a loss function, without much reliance on a superviser.

I saw Wikipedia grow from a playful listing of facts and "did you know?"'s to what it is today. I came to rely on the Wikipedia facts and trust was very high.

It took joining and partaking in the creation of a "politically contentious" article when I realized just how organized, blatant and scary these propaganda efforts are. You have it easier when you are a crackpot, than when you play by the rules and source all facts, because then you force them to earn their pay using more drastic measures. Yet, I can recommend this revelation to everyone.

See also another scuffle which relates to these current wiki-astroturfing campaigns: http://en.wikipedia.org/wiki/Church_of_Scientology_editing_o...

Also remember that Wikipedia is in a lot of countries. Some of those countries have governments and companies which are more blatant or forceful dealing with unpleasant information in the open than say, an American company. It is also easier for those editors to form cliques.

It is useful to look at countries which require that you identify yourself if you want to sell stuff online.

For commercial sites in Germany a "Web Imprint" (Impressum) is required by law. This "Web Imprint" has to be on a prominent, easy to reach position on the site. It lists the contact data for the owner(s) of the website.

Making anonymously-run webshops against the law shields the consumer against fraud and always gives non-law enforcement a place of contact to file their complaints.

Check for yourself the correlation between WHOIS protected/anonimized webshops and shady business.

"Earlier this week, we reported a glitch in Google Search, where clicking on a link after a search for Gmail listed someone’s email address, complete in a window ready to send them email. With today’s Gmail outage, that error — never fixed — has resulted in at least one person getting thousands of unwanted emails." http://searchengineland.com/google-search-error-gmail-mistak...

In 2004 user "Lonely" on the Movie Codec Forums posted a message with the title: "I am lonely will anyone speak to me.". Within days this thread became the number one result in Google for "I am lonely". 10 years later this thread has amassed over 20.000 replies. All from people who felt lonely and typed their mental state into Google. http://lounge.moviecodec.com/on-topic/i-am-lonely-will-anyon...

A Dutch politician responsible for investigating failed ICT government projects took to Twitter to deride Microsoft's support desk for calling him 10+ times a day from India. "Dear user, we will never call you and ask you for passwords. This was a criminal phishing attempt." https://twitter.com/Juliedewidt/status/574887734608805888

Wolfenstein 3D had an easter egg in a secret room which said: "Call Apogee. Say Aardwolf". Hackers found these rooms through reverse-engineering so the entire easter egg phone campaign was scratched. This one Easter Egg somehow remained in the final product and for many years Apogee got these "Aardwolf"-calls. http://static.rookiemag.com/2013/03/1362676256wolfensteinpic...

HN does not allow to give hat-tips or other Twitter-like credits. Where should they even put that? In the title?

It is: first come-first serve and I don't think you can even mention the OP of this post without scrolling back and looking it up. It doesn't matter much to you and me.

I was hoping the challenges would include machine learning. There is so much exciting overlap between ML and security. From detecting intrusion attempts to scraping and processing pastebins.

I think it is amazing what you guys did here: you got more points on a give me your e-mail, we are building something post, than Stripe got when announcing their company. Really speaks volume on Patio's writing.

Will there also be marketing challenges? Like: Every day patio11 clicks on one email. Your task is to write an engaging headline. Or: Here is some Google Analytics code. Make it show 50.000 visits any way you can. Very hard to cheat. (But so is getting Google security to file a bug report).

I really enjoyed the comments. Especially the ipod-hasnt-got-wifi criticism. You guys must know that what you are doing is terrible and you should feel bad.

Who knew decoding morse in some wave extracted from a corrupted image file in Cicada 3301 could also lead to a job. Will there be an ARG element to the CTF's? Like you pose as an agent from the NSA or something?

Then he can wait for the RNG to produce this same random data. Eventually it will produce a file which matches. A dynamic solution of sorts, because he would have to be quick to diff, before the file starts changing again.

I feel that the compressor for a true random stream is a true random generator. If I quickly show you a screen of black-and-white unpredictable noise, and ask you what it was, you'd compress/understand/recall that as "generate_noise()". I do not feel that this is lossy compression, for what did you lose? The ordering of a random file? Random files have no order to lose.

Yes, if you want to be sure that your solution is correct, you must run the compressor yourself. Then you count the number of collisions it takes to happen upon the correct solution and feed this counter to your decompressor. But then you place the burden of solving the halting problem on yourself and then you got more serious problems than compressing random data.

Blast! That would require our solution to break all known encryption. Then it would be easier to just target Goldman's setup.

A file with output from AES-256("sekrit") would not be random to Goldman. It would appear random to us, until we crack it with a plausible password. Key to randomness is unpredictability. The moment Goldman saves some atmospheric noise on his computer and makes a short pointer to it on his filesystem, is the moment this file loses its claimed unpredictability. Goldman knows this file deeply: He himself has compressed meaningless random data into predictable information that has meaning and purpose, a feat he set out to prove impossible. By simply entering the challenge with real meaningless randomness you make Goldman solve it for you.

Chances of getting a collision is higher than getting a good solution, but having multiple hashes does help in increasing the chance at a good solution. With a hash 1 bit less than the length of the file, we put two pigeons inside one hole, and have a 50% chance at picking the right pigeon. The fewer/smaller hashes, the more we get "sorry, but no".

A trick similar to the recursive Barf compressor (add information to the filename). http://mattmahoney.net/dc/barf.html

A longer running challenge is http://www.drdobbs.com/architecture-and-design/the-enduring-... No entry fee, $100 prize, and just as unfair.

A completely serious compression challenge with serious consequences for AI and NLP: http://prize.hutter1.net/ up to 50.000$ prize money, but severe restrictions on memory and running time.

You can not beat Goldman's troll-ish challenge (certainly if the rules are retro-actively clarified in favor of the organizer). You could however try to put the challenge in limbo by creating a decompressor which bruteforces a solution, 'till some hashes match or the final heat death of the universe or the halting problem is solved, whichever comes first. Goldman will not be able to ever verify your solution, and when he does (theoretically it is not impossible), it means you win.

Or, instead of above Schrödinger's Compressor you can send a good random number generator back as your solution. If Goldman wants his file to be random, any random file should do. Why does he want exactly his own random file? Why does he want to do a diff between two random files, is he perhaps looking for order where there is none? But that's the same foolishness he accuses his participants of.

Could be part of it.

1. Add a link to the uninstall instructions on my download page. I told him that my uninstall was completely standard for the platforms I support (Windows and Mac). But he insisted.

I looked at point two: 2. Remove some links that were ‘redirections’ from my reviews page.

and was wondering what that was originally all about. Could be linking to malware, but I also do know that Google has become stricter in attacking linking practices where money could have exchanged hands. When you run an affiliate program with Amazon for instance, it is a good idea to "nofollow" these links, to make sure you don't run afoul of these rules.

Then again, it seems a lot of things just went wrong in this case. Having just received a warning that a 5-year-old 960.gs fixed width site of mine was delisted from Mobile search for being "not friendly to mobile users" I can now better sympathize.

  These [no malware] requirements apply to software hosted 
  on your site or linked to from your site.

Perhaps it started with this? You linked (or let a reviewing user link) to a hacked site. I can understand Adwords taking action in that case (even when it is a simple hyperlink, it is dangerous to users of the ad network).

The support was just bad. I can not defend that. Good on you for keeping your cool.

AI at Google knows which document to show for your query and which advertisement you are likely to click on. Goldman Sachs uses machine learned algorithms already. Running a company requires vision, creativity, leadership, human interaction and usability.

If we allow for a conscious AI to house in non-biological neurons, then we should also allow Turing-complete companies to have a consciousness. Visualize every Google employee with a two-way radio transmitting signals just like neurons would do. Would such a Google-brain be conscious and capable of new and unique mental states, or more familiar states like stress?