This is a page for Joint work of Marie Roch, Jonathan Walton, John Hildebrand, Yoav Freund, Sanjoy Dasgupta, Matus Telgarsky and Cory Rivera
Pieces of the puzzle
(appearing in chronological order)
- Dolphin Classification Using Echolocation Clicks
- Using JBoost
- Comparing LogLossBoost and GMM
Thoughts on click features
Yoav had suggested trying the time signal of a click for a feature vector. I am somewhat concerned that alignment based only on the start of the click may be insufficient to do this. If the productions are not uniform (as in the case of speech), you may see different parts of the same production arriving on different samples which would translate into different dimensions if the time-domain click is treated as a feature vector. Dynamic time-warping could be a possibility, but given the variety in the clicks due to axis issues and possible variation both within and across species, I think that finding a suitable reference template would be difficult at best.
This does not mean that we should not try this, but I would stack a number of clicks vertically to see whether or not this seems appropriate before we actually try it out. If there is little variation within the click (e.g. uniform slow down/speed up), we might be able to get away with a simple linear normalization. - Marie
--Yoavfreund 17:02, 17 December 2008 (UTC) Another approach might be to use the envelope of the wave-form, instead of the waveform itself.