管理员
- 积分
- 77022
- 大米
- 颗
- 鳄梨
- 个
- 水井
- 尺
- 蓝莓
- 颗
- 萝卜
- 根
- 小米
- 粒
- 学分
- 个
- 注册时间
- 2009-5-28
- 最后登录
- 1970-1-1
|
这些问题提的很好,题主看起来思考已经挺深的了
ML DL 构架基本不是DS做,每家公司情况很不同
以产品为导向的公司里面,ML/DL infra 好到能让随便一个DS就push model to production的,目前不知道有谁家(我个人看过的几家都远远做不到)。
. 1point 3 acres高大上的建模只是DS日常非常小的部分 --- 而且随便上个公开课就要求非建模工作不做的小同学们也是够了,大把的DL/ML出身PhD在那边呢。。。大部分时间,需要理解应用场景,需要协调人与人的合作,组与组的合作,pull and clean data, 很多试错,调参,推到生产后也很多杂事。构架做的好的公司,这些“杂事”就少些。构架差的一般是相对新的公司,小的公司,或者非技术导向的公司, then be prepared to do the dirty work, in order to get to the interesting work. At the end of the day, companies pay for DS team, for "business impact", not for fancy models.
以下纯属个人ramble,不一定很准:. 1point3acres
In the better setups, DS should have good enough coding skills to "some" production code, with guidance; Eng should know enough about data/ML to collaborate with DS; ML eng + ML scientist can make a powerful team that produce great work.
In less than ideal situations, DS do most of ML but can't effectively productionize, and their partner eng teams don't understand enough about ML to push stuff to production. Or DS are excluded from ML effort, and eng own ML 100%, but often failing to capitalize on huge opportunities. .google и
In very small teams, ML effort can be very ad hoc, and the data product is hard to maintain, is not monitored (or even validated) for model performance, let alone checking for perf degradation over time.
In very elite places, things may be different.
下一段证据就更少了,纯个人主观意见:
Just from the outside, it seems Amazon is doing well -- not sure if they used extremely fancy models, but their choice of the application scenario is fantastic. New products powered by ML/DL from amazon almost always amazes me. 他们帮人评估穿衣服的应用,看样子将来会秒stitchfix.
Linkedin and Didi seem to be working on solid applied DL, and seeing good results.
Not familiar with FB on the applied side (there's newsfeed, ads and search ranking teams using ML) FAIR is obviously great, not super familiar with their applied ML stuff. But DS (non core) at FB don't get to do model AT ALL. I'd be shocked if any product analytics DS got asked ML during their interview.
G is obviously far superior in this aspect. But most of the fun stuff is owned by eng.
Uber has good applications. Airbnb has some applications. Excluding the self driving car stuff, I would be suspicious if either claim to lead any real bleeding edge stuff in this area. . Χ
Snap obviously has good computer vision dept - so does FB. But not for DS..
Lots of Chinese companies are making solid, rapid advancements.
MSR has good research, not sure how they got applied.
. 1point3acres.com
So, where can DS work on ML?
G? NO
FB? NO
Linkedin? Yes
MS? Yes, some teams
U/Lyft? Yes, some teams
Snap? NO. 1point 3acres
Airbnb? Yes, some
Amazon? Yes.--
Netflix? Yes. .и
Stitchfix? Yes
startups? For sure |
|