home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!munnari.oz.au!cs.mu.OZ.AU!mullian!dbg
- From: dbg@mullian.ee.mu.oz.au (David Grayden)
- Subject: TDNN training time
- Message-ID: <9224413.19523@mulga.cs.mu.OZ.AU>
- Summary: My TDNN trains too slow
- Keywords: TDNN, BP, training time
- Sender: news@cs.mu.OZ.AU
- Organization: Electrical and Electronic - University of Melbourne
- Date: Mon, 31 Aug 1992 03:10:56 GMT
- Lines: 46
-
-
- I am doing some research in phoneme recognition of continuous speech. I
- have the Timit database from which I extract my training and testing sets.
-
- At first, I have been trying to replicate somewhat the results of Alex
- Waibel and the ATR group in TDNN recognition of /b/, /d/ and /g/. However,
- the training time is almost prohibitively long. I am basing most of my
- work on the two papers:
-
- A.Waibel, T.Hanazawa, G.Hinton, K.Shikano, K.J.Lang, "Phoneme recognition
- using time-delay neural networks," IEEE Trans. on Acoustics, Speech and
- Signal Processing, vol.37, no.3, March 1989.
-
- P.Haffner, A.Waibel, H.Sawai, K.Shikano, "Fast back-propagation learning
- methods for large phonemic neural networks," ATR Technical Report.
-
-
- In these papers, it seems to be that the bdg network can be trained in as
- little as 1 minute on an Alliant super-computer. However, I can't get it
- down below 1 week for 1600 training samples. I am working on a SPARCstation
- which will explain some of the slowness, and the TIMIT English database
- may be "harder to train on" than the Japanese database used above, but the
- slowness of my training seems to be several orders of magnitude greater.
-
- For comparison, I have trained a fully-interconnected neural net in the
- same task. It trains extremely fast although each epoch takes longer due
- to the extra weights in the fully-interconnected setup.
-
- The fully-interconnected network trains in 1/10 to 1/1000 the number of
- iterations and thus takes less time by about the same orders of magnitude.
-
- The TDNN (after waiting for over a week sometimes) gives recognition rates
- between 60-80% on the TIMIT test set, while the fully-interconnected
- network recognizes between 70-90%.
-
- Has anyone else done work in this area on TIMIT and managed to get speeds
- and results anywhere close to those of the ATR group?
-
- Regards,
- David.
- -----------------------------------------------------------------
- David Grayden | Tel: +61 3 344 4974
- Dept of Electrical Engineering | Fax: +61 3 344 6678
- The University of Melbourne |
- Parkville VIC 3052 AUSTRALIA | Email:dbg@mullian.ee.mu.OZ.AU
- -----------------------------------------------------------------
-