From: malcolm@interval.com (Malcolm Slaney)
Date: Mon, 26 Sep 1994 17:07:09 -0800
Subject: Auditory Grouping at CCRMA
Message-Id: <aaad1eea060210046b14@[199.170.106.94]>

This week, David Rosenthal will be presenting his view of auditory grouping
and sound separation.  Some of you will remember David's work (while he was
at the Media Lab) on Rhythm Perception.  This work was presented at the
Hearing Seminar a year or two ago.

Over the past few weeks we've heard from a number of people about their
sound separation techniques.  David has expanded his rhythm perception work
to do more general sound separation and will be talking about his work.
His rhythm work was great, I expect this will be too.

        Who:    David F. Rosenthal (International Media Research Foundation)
        What:   Generic Perceptual Grouping Techniques
        When:   Thursday Sept. 29 at 10:30 AM
        Where:  CCRMA Library (Top Floor of the Knoll)

In coming weeks, Dick Duda will be talking about his computer model of
onset detectors, and I'll be reviewing a paper on approaches for modeling
computer vision and how they apply to audio.

See you at CCRMA!

-- Malcolm


From: "David F. Rosenthal" <dfr@brazil.imrf.or.jp>
Title: Generic Perceptual Grouping Techniques

Abstract:

In this talk I will make the case for the following claim: A number of
apparently diverse problems in computational perception -
particularly, problems of machine audition and vision - have some
interesting commonalties which deserve more attention than they're
currently getting.

Machine perception problems are often usefully viewed as problems of
_hierarchically grouping primitive elements_. The primitive elements
in question may be pixels in the case of a vision understanding
system, or concentrations of energy in certain time-frequency regions
in the case of an audio understanding system.

Many of the considerations involved in re-representing and grouping
the primitive elements are domain-dependent: so, for example, brightness,
spatial-location, color, and so on, dictate the grouping of pixels, while
energy, frequency, temporal location, etc. dictate the grouping of
audio data.

There are, however, important considerations, such as efficient
representation of multiple grouping hypotheses and integrating
evidence from multiple sources, where the particular perceptual domain
is relatively unimportant.  It's while tackling problems like these
that researchers in ostensibly different fields tend to be working on
essentially the same problem - usually without being aware of each
other's efforts.

In my talk I'll show how we successfully applied computational
structures built to solve one machine perception problem - machine
understanding of musical rhythm - to an apparently completely
different problem - that of automatic sound source separation.  I'll
show, furthermore, that the techniques in question are generic, in
that they apply to a variety of problems in machine audition and
vision.


Bio:

David Rosenthal is Principal Researcher at the International Media
Research Foundation in Tokyo.  He has a B.A. and and M.A. in
Mathematics from Brandeis University and a Ph.D. in Media Technology
from the MIT Media Laboratory.  Before completing his graduate studies
he worked as a software engineer at Computervision, Inc., as a
researcher at Thinking Machines, Inc., and as a consultant to the CYC
Project at MCC.  While earning his doctorate at MIT he also completed
a B. Mus. in musical composition from the Rubin Academy of Music in
Jerusalem.

Dr. Rosenthal's research aims at embodying aspects of human perception
and cognition as computer programs.  His particular interests are
sound source separation, musical rhythm parsing and other problems of
machine understanding of audio.  His publications and technical
reports include work on computational modeling of human rhythm
parsing, sound source separation, machine learning, and text
compression.  In January, 1994 he proposed, organized and co-chaired
the International Media Technology Workshop on Abstract Perception,
where a group of machine vision and audition researchers from major
American and Japanese research laboratories explored new approaches to
computational perception.  At IMRF, Dr. Rosenthal heads the Abstract
Perception Group, which investigates common aspects of various
computational models of perception.