于机器学习来说,为什么读计算机比读统计更好?

先问是不是，再问为什么

读CS吧。统计的东西可以蹭课或者自学。反正越往上走，学科交叉的现象越突出。

machine learning的算法对数学的要求不高。上手的话看懂算法就行了
大数据的时代时间空间利用很重要，所以很多计算机的更擅长
不过这些算法都是搞数学的人研究出来的

lukeutd 发表于 2016-6-7 12:04
很多机器学习算法都是按照mr形式写的？我表示质疑。。 我感觉很多算法用mr效率反而提不上= =

It does not mean anything unless you provide with evidence.

dennis_szsy 发表于 2016-6-7 14:37
sparse encoding

你没必要存整个矩阵

this is for sparse matrix and the memory for single node is always limited. There is always a limit.

But for learning ML, you don't really need to implement such a demanding memory allocation.

To whoisit: please, learn "statistic learning". You just need to know how to use package, that's enough for a data scientist, which is the most common ML job. However, if you want to go to big company as Google to develope ML method, you really need a PhD in Machine learning, whatever it is in CS or STAT.   

dennis_szsy 发表于 2016-6-7 12:46
不仅是效率
原算法改成mr形式之后往往会从exact estimation变成approximate estimation，estimate结果往 ...

can you explain why the method changes from exact estimation to approximate estimation?

wwk55551111 发表于 2016-6-5 23:09
机器学习不仅仅是要考虑一个问题的解决，还在于解决问题方法的优化。这样的优化往往和算法和计算机的内部构 ...

This is unfair.

For most machine learning problems, find the right feature, feature selection, dimension reduction is more important than improving the infra-structure for your ML platform. Moreover, what you said, can be done by a very few number of super expert by developing certain packages for you to use. For example, I BET 100$ you cannot write a better code than the ML package already in spark. Do not make wheels, unless you are doing a phd.

For most ML job in the market, no one really needs to develop new ML method. Unless you are in google or some big company where they can afford but you have to know most of the task in daily life is really not that complex to develop new wheel to solve it. What you need is really about just like taking a kaggle competition. Know which package which method to use, and what feature to choose and what should be expected from implementing them.

大数据的东西，确实Map Reduce的框架很多都不合适，最好是一个算法一个计算框架，这就需要深入学计算机了。但是我觉得建议学计算机的理由是，即使研究不出个啥，学CS，工作还是有保障的。

leth 发表于 2016-6-13 09:30
大数据的东西，确实Map Reduce的框架很多都不合适，最好是一个算法一个计算框架，这就需要深入学计算机了。 ...

this is true.

For example, usually there is no one can find a data scientist job very soon after they graduated with a MS or lower degree, most of them find an BI or DA first, then after five years, they can be eligible for DS.

WHile on the other side of computer science, software developer fresh out-of-the-school can get the salary more than the above positions except for DS. Also, in bay area, still, SE is the largest market much much more than big data.