XKCD: It's more than a comic

I can list my more "normal" hobbies, but I'd struggle to keep a diary updated daily.

edited 19th Sep '14 8:29:52 PM by Krieger22

I was told that in a conversation it's polite to ask the other person about themself. But then they were probably told the same thing. Shit, thanks a lot mom

Well, that's one way to use your words, I guess.

I like where XKCD has been going recently. I wonder how long it will last.

Tasks

I can understand how the second thing would be incredibly hard to do, but if they're using GIS to do so, wouldn't the first part also be hard, unless you managed to get a landmark in the shot (GPS, on the other hand, would presumably by a breeze)

edited 24th Sep '14 6:04:47 AM by Thnikkafan

The first part can be 'cheated' if your app can poll the phone for location data.

The second is hard because you need OCR software that can see the bird and match the plumage with taxonomic records. Could be made easier by 'cheating' and using the mic to register birdsong, or by using the above to get the park location, then referring to a list of birds in the area and matching plumage records.

edited 24th Sep '14 7:01:03 AM by AceOfScarabs

The first part can also be a lot harder if you've turned off that geo-marking info on your phone. And I'd recommend that you do.

'Seeing the bird' also requires that the software recognize "That's a bird" even if it's partly obscured and regardless of the angle it's at relative to the camera. A bird from straight in front is nowhere near the same shape as a bird in profile, or a bird flying.

edited 24th Sep '14 8:16:21 AM by Madrugada

Yeah, OCR is really finicky like that.

The day we can scan neural activity more accurately, the first thing we'll get is better OCR.

I didn't quite get this one. ^_^;;

It's not necessarily that funny, but it's still an important point. Basically, in programming/artifical intelligence, some things are very easy to do, and some things are very hard. The problem is that a lot of the very-hard things don't feel hard to humans, because it turns out we have a lot of fixed-function brainware dedicated to those purposes.

Visual object recognition is very hard. Natural language processing is very hard. But for humans, those are both easy, and we didn't even understand just how hard they are until we started trying to teach them to computers.

On the other hand, taking image-location metadata and comparing it with known boundaries of national parks is very easy.

edited 24th Sep '14 8:39:02 AM by Shinziril

I question the usefulness of this app.

It could be a tool for the parks and scientists to catalog sitings of different types of birds that occur in National Parks. This would help in keeping track of populations and migration, and also to inform visitors what types of birds they're likely to see in which areas. If it's a publicly available app, it could also allow visitors to see and identify the birds while out in the Park, without needing to find a Interp ranger or go to the visitor center to ask what type of bird it is.

Or, for a more nefarious purpose, it could be poachers making an app available under the cover of "Identify birds on your trip!", and then collect and use that data to identify where they can find rare species.

@Thnikkafan: I thought so too at first until I realized GIS didn't mean Google Image Search this time.

For the sake of reference, probably the "GIS lookup" they speak of.

Whatever, it still requires your location data (via GPS or cell information). After that, it's a simple web service to query.

While it is indeed easier to implement the query to the web service than image recognition, it is not necessarily simpler. Just think about it. You get your location information from hundreds of satellites developed by the military over several decades (not to mention the amount of space science needed to set the satellites in orbit). Then you use your mobile network, a complex infrastructure of radio towers and computers to get your data to the internet, an even bigger and even more distributed network, so you can send your location to one server to look up what national park you are in. This server is probably designed to be able to answer thousands of queries per second (at least, Google has it in the millions). Image recognition, on the other hand, requires only a camera and one CPU (whereas previously you used hundreds of CP Us, most of them hundreds of times more powerful than that of your smart phone).

I think we'll get working identifiers and translators after the singularity.

Translation is an expert system problem; it's relatively straightforward to solve with available computing power. We can already do reasonably good real-time speech recognition on smartphones, and feeding the resulting text into a context-sensitive translation engine is just an extra step. The tricky part is training the software to recognize idioms, sarcasm, and such, but it's hardly a singularity issue.

Speech is tough, but it's crackable. Images are harder because the amount of data to crunch is orders of magnitude greater.

edited 24th Sep '14 2:12:13 PM by Fighteer

As someone who worked in speech research for a few years, I think you're underestimating the sheer amount of missing information involved in speech recognition, not to mention the amount of backtracking that humans seem to do with a sentence. When people speak, they do not speak in complete phonemes, so the system has to guess there. Once you have a likely set of phonemes, you have to guess at words. You have to guess at sentence structure. You have to realize when someone's using a proper name. And, two sentences in, you may have to realize that the conversation was not what you thought it was about and everything has to be re-evaluated. We humans do this naturally. It's a lot harder for computers.

For a known vocabulary, computers can achieve 80-95% with well-trained users and an adaptive system without background noise. Outside of that, it's pot-luck.

He's right about how difficult it is to get computers to manage image recognition. I've got thousands of pictures on iPhoto and the Faces function still has trouble recognizing myself and my family (never mind its issues with erroneously identifying paintings, statues, and even random bits of background as faces).

Which is kind of comforting. The idea that a stranger could take a photo of you and use it to find out who you were based on matching it to your Facebook profile photo or something is creepy.

edited 24th Sep '14 9:18:08 PM by WarriorEowyn

What-If? #113: Visit Every State

In this one, Randall isn't doing original work, but is rather taking someone else's work and thinking it out a few more steps.

And now I find myself curious about that XKCD strip because I don't know enough about the setup to wonder what's so interesting and complicated about it.

Taking an introductory electromagnetism class, and I couldn't tell ya.

Tricking face recognition is the best though. Especially when it then feeds into something else.

In other news, I bought the What If book and although I haven't finished it yet, it's pretty good. Kind of wish it was clearly divided into the blog ones and the new ones though.

edited 25th Sep '14 12:20:09 PM by tenebrousgaze

An alttext that is informative rather than a joke?!

There are an infinite number of finite routes the current takes, each of which has a finite resistance proportional to its length, which can be arbitrarily large, and they are all tangled together. Summing them all together is... non-trivial.