You need to think like an interviewer - what can you reasonably make someone do in half an hour (plus time for chat and questions after)? Apart from being able

You need to think like an interviewer - what can you reasonably make someone do in half an hour (plus time for chat and questions after)? Apart from being able to parrot deep learning theory, implementing things is tricky. Do you learn anything from making someone implement VGG in their pet framework? Training models also takes more time than you have to spare. Much easier to quiz the applicant how they would solve a problem, or to discuss a previous project or paper they've published (or are interested in). Some people will find that much easier than whiteboard coding, others will hate it. It really depends where you apply and if you want an applied or research role. Some places won't touch you unless you've got a publication in somewhere like CVPR. Others will go _hard_ on the stats questions. Other places want to see a strong Kaggle rank or some personal projects. It's really useful to have a portfolio here.