Google research intern interview 面经

面试两轮，每轮45mins, 第一轮是国人小哥，比较nice, 会给hint，感觉还可以。 第二轮是印度小哥，感觉虽然看起来比较nice，其实比较严苛，有点笑面虎，感觉面试一般。
first interview

Self-introduction
20 mins machine learning knowledge:
1. Overfitting: what is overfitting, and how to prevent overfitting?
Extended question from interviewer: how to avoid overfitting from data-level
   2. what  is batch normalization?
Extended question from interviewer: what is the batch normalization function from the training process and testing process?
Answer: robust prediction and no parameters will be updated
    3. What is LSTM, RNN? What is transformer? Advantages and disadvantages for these models?
20 mins research paper

second interview
no self-introduction
20 mins machine learning knowledge:
1. What is graph model? What's the difference between graph models and other models? What's the difference between GCN and graphsage?
2. What if model doesn't have good performance on testing dataset? How to solve overfitting? Why regularization can solve overfitting? How to solve underfitting? Why underfitting?
Approximately 10 mins Machine learning design: give all map data (# of accident) of America, how to predict this intersection is dangerous or safe?
How to evaluate the model performance? High recall low precision or low recall high precision? What's the formula of cross entropy  loss?
How to define the ground truth?
How to predict new interactions?
How does graph models predict new intersection?
Answer: Model design, model selection, classifier, loss function,  
Research-related problem：
What is graph contrastive learning?
Tell me about one of your research.

来个dp  前几天收到updates 说进入team match 阶段

厉害，能问下楼主的bacbackground吗