Subject: Re: Convolutive feature extraction (Searching for papers) From: Eckard Blumschein <Eckard.Blumschein@xxxxxxxx> Date: Mon, 16 Oct 2006 08:37:46 +0200Paris Smaragdis <paris@xxxxxxxx> wrote > http://www.merl.com/reports/docs/TR2004-094.pdf Dear Paris Smaragdis, You wrote in last sentence of discussion in this paper: Unfortunately, what makes non-negativity constraints so successfully, in many domains, is still a subject under debate and investigation. Unfotunately. Really unfortunately? I would like to say fortunately, instead. Perhaps you meanwhile got aware of my paper http://iesk.et.uni-magdeburg.de/~blumsche/M277 I will put a newer one at M283. Natural frequency analysis in cochlea has no possibility to obey the manmade arbitrary choice of a reference. So it cannot exhibit linear phase at all. Elapsed time is always non-negative. Nobody is using complex Fourier transform within his organ of Corti. Let's learn non-negative sparse coding from at least 100,000,000 years of natural evolution. The older generation of mathematicians was less mislead than the present one. Duhamel's integral did not range from minus infinity to plus infinity, but it started convolution at zero. So we have to reinvent IR+. Kind regards, Eckard Blumschein