Examiner Comments

Thesis Planning Report

Hi John,

Looks great and I don't see any warning signs at this point, so I'm happy to let the project proceed. The text is fine, and the project and the time plan look feasible.

A few comments. These are just some things that I thought of when reading the text, and I don't expect you to submit a new version that addresses these things, but you might keep them in mind when writing the final report.

• I was a bit confused when reading section 3 because you introduce a few ideas (MWFD, residual learning, identity mappings) without saying what they are and what problem they are supposed to solve. This was clarified when I continued reading, except that the identity mappings aren't mentioned after section 3. Would still have liked to see a stronger explanation for why you believe this techniques would be useful for your task.

• sec 3.2: if I understand you correctly, in the beginning you will use a single-label classifier even though the problem is multi-label. Is it trivial to train a single-label classifier if your annotated data is multi-label? Not sure I understand what you mean by "We will assume closest".

• The research question "How do we tune the hyper parameters" is not mentioned later in the text and there are no methods described for how to address it.

• If you need to use the Amazon service, will you pay for this on your own?

Great work anyway!

Regards, Richard