Computing desk
< May 22	<< Apr \| May \| Jun >>	May 24 >

Welcome to the Wikipedia Computing Reference Desk Archives
The page you are currently viewing is an archive page. While you can leave answers for any questions shown below, please ask new questions on one of the current reference desk pages.

May 23

Non-applications of Machine learning

Reading through the Machine_learning#Applications list of applications of ML, I wonder if there are problems where you can't apply it. --Llaanngg (talk) 02:07, 23 May 2016 (UTC)[reply]

CAPTCHAs might be an example. StuRat (talk) 03:10, 23 May 2016 (UTC)[reply]

What an inexplicable reply. Using machine learning to solve CAPTCHAs is an extremely common task. DTLHS (talk) 04:24, 23 May 2016 (UTC)[reply]

The whole idea is to make them so they can't be solved by machines, so if they are easily solved by machines, that means they are no good. StuRat (talk) 04:28, 23 May 2016 (UTC)[reply]

StuRat, see [1] 2001:630:12:2428:7545:3F04:2666:1332 (talk) 05:57, 23 May 2016 (UTC)[reply]

That gives us a cue in the right direction. Somehow there are patterns computers cannot process easily. But the case of Captchas is artificial. Are there non-artificial problems, not created intentionally by humans, that are resistant to ML treatment? --Llaanngg (talk) 12:34, 23 May 2016 (UTC)[reply]

Wouldn't the entire class of NP problems be resistant to machine learning? Humans cannot find a simple solution to those problems. Humans developed machine learning. So, machine learning shouldn't be able to find a simple solution to those problems. 199.15.144.250 (talk) 13:38, 23 May 2016 (UTC)[reply]

Re: "Humans developed machine learning. So, machine learning shouldn't be able to find a simple solution to those problems." I disagree. Computers can solve problems by brute force that humans can't, because they can process records so much faster. It's problems that can't be solved by brute force alone, where you need human intelligence, perhaps in concert with computers, to solve the problem. StuRat (talk) 15:05, 23 May 2016 (UTC)[reply]

The idea that computers can solve problems that humans cannot is a bit silly. A computer is a tool used by humans. Do you claim that humans cannot drive nails, but a hammer can? I believe the point is that an unsolvable problem cannot be solved by a computer because the computer is just a tool. If the computer solved the problem, the human that used the computer also solved it.

The human could only solve it if he had a billion years, because computers, at their best, do the same things as people, but much, much faster. For an example, computers can calculate prime numbers way larger than a person alone could. StuRat (talk) 04:07, 24 May 2016 (UTC)[reply]

You are missing the point. The computer is a tool created by humans to do work. In the realm of "is this problem solvable or not", it doesn't matter if the work is done slowly by a human or quickly by a computer. Nobody is arguing about the ability of a computer to do many repetitive calculations very quickly. The argument is that if a human uses a tool to solve a problem, then the human solved the problem. Using the example above, would you refuse to pay a carpenter with the claim that his hammer and saw did all the work? Apparently you wish to tell a statistician that he doesn't deserve credit for his work because the computer did all the work. 209.149.114.20 (talk) 12:22, 26 May 2016 (UTC)[reply]

The question is of course far too broad, and I think you might be misunderstanding what ML is. It's just a class of tools that are very powerful in very restricted contexts. They also happen to be very hip right now. Then again, as you can see from the topic list, sometimes the term is interpreted very broadly. One may wonder what is the "learning" that is relevant to both genetic algorithms and support vector machines - and one may wonder what the utility is gained by grouping these unrelated algorithms together, but I digress.

ML and its kin are great for certain tasks that involve making certain types of decisions and classifications. Lots of types of predictions cannot be addressed through ML. Consider proving a theorem. Now, we can and do use computers to help with that (Automated_theorem_proving), but it usually uses methods that are rather distinct from what we call ML. Logical inference has basically nothing to do with splitting data clouds with the best hyperplanes.

Even when deciding how to classify things, ML is often not best. Evolutionary algorithms use evolutionary ideas to solve problems about other things. But sometimes we want to use algorithms to solve problems associated with evolution and the resultant phylogeny. That leads us to Computational_phylogenetics, which is all about taxonomy, yet does not usually use any techniques from ML.

Then of course there's almost anything related to mathematical models or simulations. So e.g. none of the top climate models have anything to do with ML., and whole branches of physics,chemistry,biology that have to do with quantitative simulation have nothing to do with ML, and frankly very little, if anything, to gain from it.

In summary I stress that ML is ultimately a rather small class of techniques. While very powerful in the proper domain, there's zillions of things that are "resistant" to ML treatment. So I have given you some very broad and very active areas of research that have nothing to do with ML, and whose problems are not well treated by ML, but I am limited more by time spent than number of things that ML can't do. Finally, I'll add that even when ML "works", it seldom helps with understanding anything - like if I sorted out your bug collection for you that wouldn't help you understand what diptera is.

If you ever get a chance to see someone talk about their research that deploys ML in an applied context, ask them something related to how this method helps our understanding of the problem or solutions, and 9/10 times they either dismiss the question or uncomfortably squirm ;) SemanticMantis (talk) 14:00, 23 May 2016 (UTC)[reply]

Love. The Quixotic Potato (talk) 20:54, 27 May 2016 (UTC)[reply]

Open source software(s) sought

I sought a software(s) similar to "Windows 7 Logon Editor v2" (which consists of a default button), in order to change "mouse pointer" and "Files and Folders icons" Can someone help me with this please? -- Apostle (talk) 18:34, 23 May 2016 (UTC)[reply]

Is this on Windows 7, or somewhere else ? StuRat (talk) 22:01, 23 May 2016 (UTC)[reply]

Yes, my OS is Windows 7 Ultimate Service Pack 1 so any Windows 7 software will work... A software(s) that helps me change all the mouse pointers and Files and Folders icons with ease... Default option is desirable. -- Apostle (talk) 04:50, 24 May 2016 (UTC)[reply]

I believe native software on that O/S allows you to change them. Do you mean you want a larger library of choices ? StuRat (talk) 06:09, 24 May 2016 (UTC)[reply]

I reviewed. Let me know if there's a larger library available in the PC...

Any idea how I could change all the folder icons in one go? Restore to default when desired in one go again?

Also, there is one Folder icon I worry i.e. e.g., when you insert an image/doc/any file into the normal folder, the folder automatically displays what's inside the folder (a few files or so) - what is this particular folder known as? -- Apostle (talk) 18:14, 24 May 2016 (UTC)[reply]

Always make backups of all important files. You can use something like ResHack and look inside shell32.dll. [2] The Quixotic Potato (talk) 20:50, 27 May 2016 (UTC)[reply]

Noted! -- Apostle (talk) 09:21, 28 May 2016 (UTC)[reply]