what does it mean for a machine to “learn?” How do we quantify something like skill growth when, for humans, learning is a subjective experience?
In: 4
You feed the machine a bunch of data along with a result. The machine tries to figure out a way to turn those data points into the result you want. It does this again and again, refining the algorithm. Then you can put in a bunch of data points and ask what the result would likely be. The machine learned how to do that thanks to the training and gives an answer.
So I guess I’d say it means the refinement of an algorithm by the machine with the only human interaction being feeding data and the expected answer,v instead of manually tweaking the algorithm.
Then if it’s successful, it learned how to identify cancer cells, or risky loan applications, or whatever the data was.
A machine learning is it expanding its algorithm. So a machine that plays chess being programmed with only, let’s say, 50 possible game outcomes. Every time a human plays chess against the machine, it records the game played. It continues to expand its algorithm and it seems like it’s getting better at chess when, really, it’s just expanding its catalog of options.
Humans learning is subjective because we don’t have algorithms. One person who plays chess may come away with a better understanding of the game; another may come away with a transferable skill of analyzing and predicting outcomes before making decisions. Machines don’t transfer their skills. A chess playing machine will need new programming to do a new task.
Machine learning is a misnomer. The machine isn’t really “learning” in the same way that humans learn things.
Rather, what you do is you feed in a bunch of data and it makes correlations between the data, creating a complicated algorithm that generally gives a more or less correct answer (hopefully).
“Machine learning” is really a programming shortcut when we don’t know how to program something to do something.
There are severe limitations on this, which people try to paper over to make money.
Machine learning is a useful tool but thinking of it as an intelligent process or analogous to human learning isn’t really correct. If you teach a human a concept, they can apply it freely, but the machine isn’t being taught concepts in this way, it’s more like someone trying to observe someone else doing something and then repeating what they saw that worked, but without understanding what they were doing. Sometimes it works and sometimes you end up with cargo cults.
The threshold for what is considered “machine learning” can be absurdly low. It’s basically any statistical model that can update itself based on training data. For example, I want you to predict someone’s gender if all you know about them is their height. Without any additional information, all you can really do is pick some cutoff and classify anyone above that height as a man and anyone below that height as women. But how do you choose the exact cutoff?
Say you have training data: the heights of 100 random men and the heights of 100 random women. You want to pick the cutoff value that maximizes the number of correct guesses, from this data set where are you already know the answers is in advance. Once you’ve written a program to do that, congratulations, you’ve just achieved machine learning!
Obviously, machine learning will be much more interesting if you have more variables in your statistical model, and then neural networks will be even more complicated and interesting.
> How do we quantify something like skill growth
Split off a chunk of your dataset for validation. For a classification problem, it’s usually a simple matter of accuracy=correct classifications your program managed/total items in the validation set. If it’s a regression problem (guess a numeric value instead of trying to attach a label to something), usually something like mean squared error is used.
> for humans, learning is a subjective experience?
But I can rate your ability to recognize Monet paintings by showing you a bunch of paintings and asking you if they are from Monet or not and noting down the ratio of mistakes, yes? Or, something simpler, this is the same concept as common multiple choice tests are graded in school.
Imagine you living a housing development. Between you and the local shopping center is a dense forest.
To walk around that forest would be a nuisance. So you and your neighbors always try to cut through it, despite the thick undergrowth.
As you and your neighbors keep going shopping, you start to notice paths being formed. It’s a lot easier to walk on a path – even a barely formed one – than it is to trudge through the thickest part of the forest. So over time those paths become more well-trod as people favor them.
Eventually, you’ve got a small number of wide, smooths paths through the forest.
The forest has ‘learned’ the most efficient ways for human beings to travel through it – and that learning is codified in the paths that exist through the forest.
Both human and machine learning work much the same way. At any given time, you’ve got certain features triggering. Your brain – or your computer – learns to correlate those features. The more those features correlate with one another, the easier the paths between them become.
For most people, the difficulty lies in understanding how such a seemingly trivial process that works with a limited number of features can scale up to ‘intelligence’ when you use a massive number of features. But it does.
Learning is not a subjective experience.
When a person learns to spell, there are ways to test whether or not they have learned. For example, maybe a test will be a list of words, some of which are spelled correctly, and some of which are not. The learner must identify which is which; some answers will be correct, others not. In this way, we quantify how much the person has learned.
The same is true of AI. We can teach an AI system to identify a school bus, and then test it by showing it pictures that may or may not have a bus and counting how many times it is correct.
You want to have a computer be able to recognize speech. You could take the time to program exactly how all words sound in various voices. Or you could program some basic recognition with a learning algorithm and then feed it a bunch of voices, have it guess, and tell it whether it was right or wrong. Knowing it was right or wrong gets fed back into the recognition so it can learn. A big difference is that machine learning usually has a fudge factor so it can weigh whether a sound is one word or the other based on its learning rather than trying to do a direct match.
Learning is not a subjective experience, I’m not sure what you mean. You either can be demonstrated to know a given piece of knowledge or how to do a given process or not. Same goes for computers, only they’re a much more natural fit for knowledge than processes and we’ve only been catching them up recently.
Simply put, it means it’s better at doing the same task.
For example, if you’ve created a machine that recognises whether a photo is a picture of a horse or a picture of a house, then “learning” means it improves its rate of successfully deteremining which is which.
It’s not really referring to the subjective experience that we talk about when humans learn–we don’t expect the machine to actually understand what it’s doing. What it’s learned is that this set of calculations produces better results than what it was doing previously.
Basically, we can only teach machines to do things we can quantify. You can’t teach a machine to “make good art” unless you can define what you mean by “good art” in a way the computer can measure.
Latest Answers