..
1)面试官准时上线,没有自我介绍,让我介绍了一下自己。没有follow-up直接开始问答。面试官是国人,大概35-40岁左右。
2)整个面试过程没有任何跟工作相关的问题,标准的MLE八股题,非常细致,从传统模型问到深度学习。面试官应该是有标准答案的,如果我回答出了关键词,那他就不再听我说了;如果没有,他会一直问我“还有呢”。整个回答过程没有任何有用的交流,他也没有给出任何引导性信息。
3)面试官口语表达能力非常差,而且口音非常重,作为中国人我只能连蒙带猜,让他复述了几次问题之后明显看出他不耐烦了。.google и
4)35分钟ML,10分钟coding,最后留了5分钟让我问问题。
面试(部分)问题
因为没有做笔记,被问到的问题也超级多,记不全所有的问题。我在这里说一下我还能记得或者答得不好的。
1) What are the basic assumptions of Linear Regression? How does violating one affect the inference?
2) What is variance versus bias (need to give precise definition). Can you give examples?
3) What are ensemble models (need to give precise definition). Can you give examples? How does Tree model works? How about regression tree? How about boosting model? I gave multiple examples on each and walked through the algorithms.
4) Is Linear Regression affected by normalization? Why? How about Tree models? Why? Give examples of regularization techniques.
5) How to process words? I mentioned TFIDF, and more recent methods such as word2vec. I gave a detailed example of word2vec approach, but I didn't remember the name of it. He asked me if I know continuous bag of words, and i said I cannot remember. I later found out that the example I gave IS cbow. He then asked me about transformer, and then I walked him through transformer such as QKV.
6) What are the metrics of ranking? In RAG, what are the metrics? What are the metrics for retrieval? How about precision? How to balance the two? What is pair-wise ranking vs list-wise? Do you know DCN (I only read about it never used), do you know DeepFM (again I read about it never used).
7) Coding - https://leetcode.com/problems/best-meeting-point/description/. 这个题我两三年前看过,还有印象所以给了verbal solution,但是也跟面试官说具体写不出来需要hints。面试官说他听不懂我的思路(尽管我的思路dryrun没问题),但是要求我按照他的思路来做,结果我就更不会了。Coupang家没有什么面经,所以我在准备coding的时候只能做到复习常见的algo和问题。