Anonymous at Mon, 27 May 2024 13:46:38 UTC No. 16195788

MLP was a mistake.

Anonymous at Mon, 27 May 2024 15:16:38 UTC No. 16195954

>>16195716
I think it just depends on what part of the field you envision yourself in.

NLP has basically three subsets with a little bit of overlap.

There's the mathematical linguistics people who are primarily concerned with information encoding and information theoretic learning with large heterogenous alphabets. This subset has both the "soft-science" side of NLP and the very mathematically rigorous information theorists.

There's also the computer science/algorithms people who worry about things like efficient dictionary searching and making sure that the learning process is scalable (meaning we want the NLP system to be able to behave as if it were performing tabular learning without actually needing massive tables of learned words and huge encoded joint distributions).

Finally there's the Machine Learning folks who are primarily concerned with the statistical learning and statistical optimization part of the problem. These are the ones who do the work on things like new reinforcement learning paradigms, performance analysis to determine appropriate metrics for comparison of different approaches, as well as new network architectures.

If really any of those sound interesting to you, then you could probably find enough meat on the bone to build a thesis.