linkedIn猎头主动勾搭的成功率是极高的,基本上一定能保证OA或者phone screen。缺点是主动来找你的一般都是小公司,并且对时间的要求一般比较specific,基本上超过3个月之后入职都是直接pass,以及可能不sponsor H1B。因此后来我也养成了习惯,猎头给打电话第一句话就说"I am graduating in XXX, 2021, is that OK",第二句话是"do you sponsor H1B?" 如果两句有一句是No就挂电话了。
. 机器学习,和统计其实是有一定交叉的。面试一般分两类问题。第一类是一些经典的机器学习的方法,例如Linear/Logistic Regression, Tree models, Ensemble models, Clustering, Optimization, Metric, Overfitting, 等等。另一类则是挑一个你简历上ML的经历,让你介绍并且回答关于你选用model的specific的问题,经常和他们做的东西有关。比如NLP就基本上一定会问transformer,LSTM。CV就一定会问CNN,等等。我在准备前一类问题时主要是回归教科书。我看过西瓜书,但觉得写的东西已经有点老了。在实战中我给我帮助最大的书是这本:A high-bias, low-variance introduction to Machine Learning for physicists,这是我校物理系一个老师写的ML方法的综述,也是我们上课时用的教材,重点它是免费的,arxiv上可以直接下载得到,还配套jupyter notebook的例子,包含了你面试时可能遇到的几乎所有经典ML方法。还有一个我觉得帮助我理解一些ML算法帮助比较大的是youtube的 Statquest,讲的东西虽然简单但是可以当作一个查询工具,比wik更直接。
3.2 面试中技巧
这方面地里总结得已经足够多了。这里我就写几条worked for me的。
1. 抱着和陌生人交朋友的心态去面试。保持genuine比较重要。. 1point3acres.com
2. 多和面试官沟通。大声说出自己觉得可能的思路,时不时向面试官要confirmation,比如"Am I on the right track?",没思路大胆问"Can I get a hint?"。 沉默才是最要命的。
3. 适当示弱,有的时候卡壳了,可以说"sorry I'm a bit nervous",或者"I'm blanking out a bit"
4. 不要不懂装懂。面试官问知不知道XXX,不知道就乖乖说不知道,一般也不会继续问下去。如果听过名字就强行说知道,那一般都死得很惨。
5. BQ:小公司很少问BQ。我一般都说自己真实的经历。如果强行假装很容易follow up翻车。但是如果你编故事能力很强,那尽管编,亚麻尤其适合这样的。面亚麻时BQ真是搞得我心力交瘁。四轮下来十几个bq。
6. 问你对公司有什么问题。我也不知道什么样的问题会加分。我自己一般只问"do you like your experience in the company" 或者"how do you think about the WLB here"之类这种面试官本人体验的问题,不能加分至少不至于减分。
statistical modeling:
What's the problem if logistic/linear regression if the features are correlated? How would you solve this problem?.
If you are designing a model for a multi-class classification for XXX data points with YYY features each, how will you do it?
What metrics do you use to evaluate a model? In XXX situation, which one is the best and why?
How to compare models and determine which one is better?-baidu 1point3acres
If you have a model that has a 90% accuracy, but we know that some data points are mislabeled unfortunately you only can test a small subset of your entire dataset, which datapoints would you choose to do the test and why?
If we try to create a linear model for a dataset that is generated by y = 0.5x + 0.2t + e1, where t = 0.3x + e2, e1/e2 are gaussian noise, what do you think will be the coefficients of x and t respectively? what if e1/e2 are varying according to x and t?
What is the meaning of the p-value for a weight in the model?
If you want to perfectly fit a function with X datapoints in Y dimensions, how many learnable parameters do you need to have at least? ..
What can you do if the dataset is very small?
How to deal with data imbalance problem?
We have millions of unlabeled data points from our customers coming in every day, how would you use these data into training your model?
If we have a well trained language model that has a exceeding size, what do you think we can do if we need a model that's 100 times lighter but preserve the performance as much as possible?
General machine learning:
Explain what is bias and what is variance, and how to reduce bias/variance respectively?
How to reduce overfitting (in XXX circumstances)
Explain gradient descend? Difference between different gradient descend methods?
How do you do hyper-parameter tuning (and what will you do if you only have limited time?). 1point 3 acres
Explain what is good about Adam optimizer, what are the meaning of betas in Adam?.--
Methods of dimensional reduction? Choose one that you are the most comfortable and explain. . 1point3acres.com Methods of clustering? Choose one that you are the most comfortable and explain.
What's different between bagging and boosting? what are the pros and cons? Do you use bagging or boosting in XXX situation?
Explain idea of XGboost?
How do you design a recommendation system?
Why is activation important in Neural Network?
Why does L1-regularization prefers sparse weights and L2-regularization prefers small weight?
Name methods for feature selection and explain how do they help reduce redundant features
What's different between batch norm and layer norm?
What do you do if test data has less features than training data?
NLP:. .и
What's the problem with RNN? (vanishing gradient) why is there such a problem? how to solve such a problem?
What is attention? how do you use it and why is it important?
Methods of pre-trained word embedding?
Describe self-attention and transformer. why is transformer faster than RNN-type model?
How do BERT train word embeddings?.google и
What is positional encoding in Transformer and why use it?
How would you design a ASR system?
Are you familiar with HMMs? how are HMMs used in speech recognition?
关于第一点,我们要意识到,找工作的过程中,你99%的时间都是在接受别人对你的judge和负面评价,你要意识到这些负面评价和你的能力背景没有任何关系,所有的拒信只是表明你们不合适。你不会因为路边随便一个人跟你说“我们不合适”而伤心吧?找工作也是一样的道理。尤其是,你申的公司大部分你连名字都没有听过,谁还离不了谁呢。到后来我拿到offer前夕,去面了一个startup,讲到我不认同某些大厂pip文化时,被面试官挑刺说那说明你对公司没有ownership。我回头直接和recruiter讲对不起,I decide to move on to a different direction。搞笑的是他们居然还是给我发了onsite面试邀请。.google и