c***z
发帖数: 6348 | 1 【 以下文字转载自 Military 讨论区 】
发信人: chaoz (没钱也任性), 信区: Military
标 题: Re: 机器学习是怎么回事
发信站: BBS 未名空间站 (Wed Jan 6 19:33:46 2016, 美东)
Do you know calculus?
Say you have a function y = f(x), where f has some parameters, and follow
some basic requirement (e.g. continuous, but this is not necessary, because
saw shaped functions work better).
Basically it is using the chain rule in calculus to move the parameters
around for minimal prediction error, which is the difference between the
predicted y and real y.
When error is minimized, we think we found a correlation between the inputs
and outputs, and try to apply that correlation to outside data.
If we are lucky, and this correlation is indeed valid in a general setting,
we push to production and party.
But alas, after a while things change and the model is not working and
nobody really knows what is going on.
And so we repeat the above process. |
|
