On supervised learning from sequential data with applications for speech regognition
Mike Schuster · Institutional Repositories DataBase (IRDB) · 1999
\Assumption is the mother of all screw-ups."fortune, slakware-3.6.0Congratulations.You are one of the few who started reading this thesis.If you have the patience, read it all.But who has.So if you just want to skip through it, read at least this preface to know in what order to skip.This thesis deals with supervised learning from sequential data, always having the quote above i n m i n d .Read the introduction (chapter 1), if you don't know w h a t I mean by supervised l e arning or sequential data.Chapter 2 summarizes the necessary basics to understand the underlying problem and possible approaches to solve it.Don't be afraid, the ideas (chapter 3 and 4) presented in this thesis are, compared to what you can nd elsewhere, relatively simple.In chapter 3 a recurrent neural network structure is extended to a bidirectional structure to model probabilistic expressions occuring when you treat 'learning from sequences' as a pattern recognition problem.The probably most interesting section is the one about the recurrent mixture density networks.Read chapter 4 if you want t o k n o w how I implemented a stack decoder for speech recognition, a challenging sequential-data problem.If you wonder why two so very dierent topics are addressed in one thesis, read chapter 2 again, because: To predict (recognize) a sequence you need always two parts, a generative part (chapter 3) and a search part (chapter 4), given the current state of research.