|
|
 | | From: | Jim Leonard | | Subject: | PAQ6 AnalogModel question | | Date: | 19 Jan 2005 12:51:48 -0800 |
|
|
 | I'm familiar with most lossless audio compression schemes, the basic
operation being the encoding of the output of a predictor. But in
browsing through the PAQ6 code I found the following description of the
AnalogModel context submodel:
"Contexts drop the low order bits, and include the position within the
file modulo 2, 3, or 4."
Forgive my inexperience in this area, but why are the low order bits
dropped? I understand this makes the high order bits easier to
compress, but the low order bits are mostly noise if taken out of
context, right? Can anyone explain why this is done?
|
|
| | From: | Matt Mahoney | | Subject: | Re: PAQ6 AnalogModel question | | Date: | 20 Jan 2005 07:55:05 -0800 |
|
|
|
Jim Leonard wrote:
> I'm familiar with most lossless audio compression schemes, the basic
> operation being the encoding of the output of a predictor. But in
> browsing through the PAQ6 code I found the following description of
the
> AnalogModel context submodel:
>
> "Contexts drop the low order bits, and include the position within
the
> file modulo 2, 3, or 4."
>
> Forgive my inexperience in this area, but why are the low order bits
> dropped? I understand this makes the high order bits easier to
> compress, but the low order bits are mostly noise if taken out of
> context, right? Can anyone explain why this is done?
Dropping the low order bits uses the approximate value as the context.
The model can be trained faster when there are fewer distinct contexts
and generalize to similar contexts. The exact contexts are still
available in the other models (CharModel, RepeatModel, SparseModel),
and these are blended into the prediction.
-- Matt Mahoney
|
|
|
