4 ms·
I think this captures a lot of what it going on in an interesting way. In many ways 'machine learning' is yet another high level compiler to convert intent into
by ChuckMcM 10y ago
I think this captures a lot of what it going on in an interesting way. In many ways 'machine learning' is yet another high level compiler to convert intent into something executable by circuits.
Consider the evolution of computers;
'plugboards' -- where you physically rewired them to change their program.
'punch cards' -- where physical media held the list of steps to execute.
'programs' -- where a text specification are compiled into a bag of bits which can then be executed by the computer.
'scripts' -- where a textual description activates different bags of bits, depending on what the text says.
'databases' -- where a selection criteria of a which data is important to you at the moment and then fed into the selection mechanism for the bags of bits.
'machine learning' -- where the bags of bits are created by evaluating a bunch of data through a pile of data selection operators and tuning the execution based on a
data you consider 'good' and data you consider 'not good'.
In all cases the basic idea is that you have a machine that you want to do X, and it can do X through a set of steps Y. Coming up with the steps Y gets harder and harder depending on the complexity of X.
It seems like magic but really its just another form of compiler. And that relationship is made even more clear by the article when it points out that a program that can play Go is not the same one that can play Chess. What is more salient is that no one has written a program that lets a computer "play" go, instead there is a program that, after being fed data about what humans did when they were playing Go and the outcomes of what they did, tweaked a bunch of parameters in a bag of variables that when you put in Go moves to the bag it comes out with a Go move that would be a good response.
No, I'm not trying to be silly here, we have yet to create a system where you could simply explain the rules of Go and have it devise a set of steps to play Go at the master level. That conceptualization of the binding between the rules and how those rules affect play and strategy, is essentially the 'code generator' part of a compiler which takes an AST and generates executable code.
Machine learning today helps us write programs to manipulate complex data sets faster than we could before. Just as compilers let us write programs faster than doing so in assembler, and assembler was an improvement over plug boards. It does not get us any closer to having a computer that can look at a data set and tell us what is important about it. That would be a better test of 'intelligence' I think.
- danans 10y ago> we have yet to create a system where you could simply explain the rules of Go and have it devise a set of steps to play Go at the master level I don't play Go, but I don't think a human could do this, either. While we can grok a set of rules to get us started on a problem, to really master complex problems we also need examples and repetition, though far fewer examples than is needed by the current state-of-the-art in ML.
- ChuckMcM 10y agoI understand what you are stating, I see it like this; You could program a computer to 'understand' better and worse Go play, then you start it up and have it play both Go programs and human players on the Internet and it would get better over time until no one could beat it. Alternatively you write a program to predict the 'next' move in a Go game, and you process through it a million previously played games to tune its probability weights. The latter is 'machine learning' the former is 'programming' but I assert they are both programming, just using a different compilation tool chain.
- petra 10y agoIn some sense the human programmer is also "a compiler", which needs a domain expert to "program" him to do his job. So maybe one interesting division isn't between what type of technology, but about when a domain expert prefers working with a tool than a programmer. But also, maybe the game example isn't the best here - one thing i noticed in the past, reading through the academic literature - is that often - you see some researchers just plug machine learning to problems that highly skilled humans has struggled with for years, and gotten good results.
- danans 10y agoI'm with you on your description of ML as a different kind of programming tool chain, but how would a "programmed" computer the way you describe a non ML system > get better over time until no one could beat it if it weren't adjusting either weights or otherwise self-mutating it's decision path based on whether it's current strategy result in a win or loss? And if it were doing that, isn't that ML (programming by learning from examples)?