My timezone is somewhat incompatible with the live stream so I hope it’s OK to post a question in advance here.
Since we learned about tabular data last time and nlp in this lesson, I’d like to ask how to combine both. If I had a table containing continuous/categorical features as well as a text column, could I create a fastai model using both types as combined input? Ideally I’d like to use a pretrained language model for the text part.
Is there any resources here in the forum for interviews preparation on ML/AI engineer positions ? It would be great to get some insight into the expectations for this kind of roles.
Jeremy announced at the beginning of class that yes, we intend to do a part 2, most likely covering the rest of the book (and potentially other topics that become relevant in the meantime), although we don’t know any specifics of when this will be. I’m sure we will continue in our commitment to making the material freely available online.
I wanted to see if anyone had interview resources (sites, books, etc.) as well. I’m currently applying to jobs and find studying for the ML part is my bottleneck.
@jeremy, for transfer learning to be useful, how well does the pretained model need to do its task? For example, can you classify sentiments well if you started with a crappy language model? Or is time well spent to perfect the language model first?
Yes, it is important to have a very good language model first. In fact, all research seems to show that the better your language model, the better classifier you will get at the end.
(In the same way, the better model you have pretrained on Imagenet in computer vision, the better your final model on another task is).