[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Information loss (Re: Analytical approach to temporal coding...)

For me it's not so much about what is lost by a transform, it's what is easy
to lose.  There are many useful and perceptually relevant transforms that
are reversible:  consider short-time fourier transforms for audition and
wavelet transforms for vision.   One of the things that makes these
transforms useful for natural signals is their information preserving
properties under different kinds of noise, such as additive noise or
quantization noise.  Some transforms make it easy to separate an additive
mixture of sounds, and some make it easy to ignore certain information, such
as phase,  when it doesn't suit a particular recognition task.

What I would like to see when a new representation is proposed is an
analysis of what is lost when different types of noise are introduced, how
pattern recognition fares for natural signals, how the subjective quality of
a natural signal degrades under lossy encoding, and how overlapping natural
sounds interact under the representation.

It is especially important, if the representation is proposed as a model of
neural processes, that the kinds of noise considered are neurally relevant.
For instance, perhaps we can expect that a temporal code will be susceptible
to temporal dispersion further down the line in a neural system: what
information is lost, and how quickly, when the code is perturbed by temporal

----- Original Message -----
From: "Dan Ellis" <dpwe@EE.COLUMBIA.EDU>
Sent: Saturday, February 24, 2001 2:43 AM
Subject: Information loss (Re: Analytical approach to temporal coding...)

> It's an interesting question to figure out the constraints imposed by
> the AN spike representation and how much of the original sound waveform
> information is available to subsequent processing.  But if the answer
> is "all of it" (as I understand Ramdas Kumaresan's post to be
> arguing), that's actually, for me, a disappointing result.
> The function of auditory scene analysis -- and audition in general --
> is to extract *useful* information (defined according to the goals of
> the organism), and throw away the rest.  It may be that, ultimately,
> the cochlea doesn't contribute much to this higher-level processing
> i.e. it is obliged to provide a near-complete representation because
> the useful selection between relevant and irrelevant detail can't be
> implemented at such a low level of analysis.
> But for me the real interest lies in finding the representations that
> *do* throw away information, that sacrifice the ability to make full
> reconstructions, then seeing which parts are kept.  That't the kind of
> description we need to help us build intelligent automatic sound
> analysis systems.
> -- DAn Ellis <dpwe@ee.columbia.edu> http://www.ee.columbia.edu/~dpwe/
>    Dept. of Elec. Eng., Columbia Univ., New York NY 10027 (212) 854-8928