You will be asked a wide range of ML-related questions (ML theory, PyTorch, CNNs, etc.). You will also be asked to code towards the end of the 1 hour session (Leetcode medium). Most of these questions have well-defined answers (e.g., how do you disable gradient computation in PyTorch) while others are more open-ended (e.g., how would you use unlabeled data to boost the performance of your supervised tasks). My major complaints are with these open-ended questions. The interviewer had specific answers in mind and would not understand/accept alternative approaches. The depth of the interviewer's ML knowledge is also questionable as the interviewer did not understand how pretrained networks can be used as feature extractors. The interviewer also asked about variational auto-encoder without knowing the underlying probabilistic formulation. Overall, a negative experience.
