Re: QI and MQ Coder: First real-life experiences

> However, the general approach used here is typical
> for my applications for entropy coder back-ends:
> Decorrelation, context modelling and entropy coding.

Having worked with audio and video codecs, as well as on
numerous other compression and coding projects, I realize
that is how it is done. And I do appreciate great deal
that you took the trouble to adapt the [QIC] code for the
audio data tests (I'll probably end up playing with it
too, now that you have opened up the topic).

While the coder is nominally a small part of the whole task,
the larger pattern itself has grown around the AC way of
coding (its probabilistic parametrization and interface,
predictive modeling constraint, its one dimensional, one
symbol at a time per full coder+modeler cycle...), wrapping
and shaping its every new layer closely to the one below,
all faithfully propagating the genetic imprint of the
little seed pattern at their core.

Maybe I didn't communicate the central point well enough
in the summary ([M1] or [N7] p.9 [T3]), but QI is an
entirely different seed pattern, almost an antipode
of the AC pattern on every key element, [M1]. { The
two coders are, of course, not the real 'master genes'
of their respective schemes. The original AC idea was
suggested first by Shannon [2] and enumerative
coding, hence its offspring QI, by Kolmogorov [31].
Each algorithm carries an implicit blueprint for
the larger pattern of one of the two kinds of
information theory frameworks, which it was meant
to unfold and in which it can be at its best.}

The present release of the QI source kit [QIC] is just
the core seed pattern which has plenty of growth and
unfolding ahead before it bears all the 'marvelous
fruits' painted in [M1], on the application levels.
The functions included in [QIC] are merely the initial
form of the core building blocks which were meant
to fit and work well within a descriptive coding and
modeling pattern. Thus trying to fit them into an
existent AC way of coding+modeling would be quite a
challenge, having to go against the grain on both
ends. And they're certainly not aiming to be a 'drop
in' replacement for AC within its conventional setups.

But they are sufficient to demonstrate the core QI
strengths (A1)-(A4), in the low noise, 'clean room'
environment, where these unique purely coding aspects
can be measured and analyzed with high precision,
in isolation from any extraneous factors, however
important and large part of the application these
other components may be. Once the core strengths
are explored, quantified and charted, one gets the
basic assurance and a much clearer idea what are
the outer bounds, the potentials in different fields,
and what needs to be built and how, on top and around
to realize these potentials in specialized applications.

While much of the present coding and modeling is
done within the AC's predictive and probabilistic
pattern, there are only the isolated elements of
the QI's modeling superstucture above, such as BW
transform, along with the vastly rich substratum
below, consisting of the highest quality mathematical
results and algorithms of enumerative combinatorics,
extremal and general finite set theory and other
areas of finite/discrete math almost entirely
untapped in this context. Hence, there will be
many pretty blossoms from this seed, before the
season for fruit arrives.


-- References ( http://www.1stworks.com/ref/RefLib.htm )

M1. -- Summary of QI algorithmic advances (A1)-(A4)
http://groups.google.com/group/comp.compression/msg/27c6f329038a1bdc

2. C. E. Shannon "A mathematical theory of communication"
Bell Syst. Tech. J. 27, 79-423, 1948
http://cm.bell-labs.com/cm/ms/what/shannonday/shannon1948.pdf

31. A. N. Kolmogorov "Three approaches to the quantitative
definition of information" (in Russian)
Problems of Inform.Transmission, 1(1), 1-7, 1965
http://www.1stworks.com/ref/kolmogorov65.pdf

QIC. -- QI source code, tech. reports [T1]-[T3], references
http://www.1stworks.com/ref/qi.htm

.

Relevant Pages