[FFmpeg-devel] VDD conference invitation - Dublin 22-24 Sept 2023

[Michael Niedermayer]

On Sat, Sep 09, 2023 at 09:23:35PM +0200, Tomas Härdin wrote:
> lör 2023-09-09 klockan 15:53 +0200 skrev Michael Niedermayer:
> > On Sat, Sep 09, 2023 at 09:04:25AM +0200, Tomas Härdin wrote:
> > > fre 2023-09-08 klockan 15:09 +0200 skrev Michael Niedermayer:
> > [...]
> > 
> > > > AI based filters are neglegted at a time everything is shifting
> > > > to neural networks and AI. 
> > > 
> > > Good. AI is a meme. Polynomial regression is just as good.
> > 
> > Theres a lot i could reply but lets pick 2 choices
> > 1. FFmpeg is about data compression so lets look at compression
> > ill make it easy, just a "toy" project from fabrice, beat this with
> > Polynomial regression
> > https://bellard.org/ts_server/ts_zip.html
> 
> I have actually had something like this in mind. Also better predictors
> for intra compression.
> 
> > 2. lets just ask AI, and while we can argue about this, i REALLY like
> > to see
> > your Polynomial regression producing anything that resembles english
> > text
> > heres chat gpts reply, first attempt:
> 
> I'm not sure what that wall of text is supposed to accomplish, a Markov
> chain can produce similar output. It's just mystified statistics, or
> machine laundered labour. Or AI voodoo as I've heard it referred to.
> See the paper "Polynomial Regression as an Alternative to Neural Nets"
> by Cheng et al

Well, 20 years ago i had a similar oppinion, NN are just a inefficient way
to approximate statistics.
But i changed my mind when artificial neural networks started to outperform
classical statistics.

I see no voodoo and no mystified statistics. What i do see is that the structures
and algorithms have advanced quite a bit while for some reason its still called
a "neural network". But the name used is meaningless, isnt it?

The paper you quote talks about neural networks as they where 25 years ago.
The field has advanced rapidly in recent years. I dont know where it will
go from here. maybe we move to algorithms that can no longer be called
neural networks even with one trying hard. Or maybe it will continue with
things resembling neural networks. To me this makes no difference.
I just find the technology cool

If you dont see the difference between modern networks and this paper, i suggest
you look at teh paper "Attention Is All You Need" this is the foundation to
systems like GPT.
And while one surely can cast anything into a frame work of additions and multiplcations
or NAND gates. That doesnt capture the structure in a usefull way.

thx

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Homeopathy is like voting while filling the ballot out with transparent ink.
Sometimes the outcome one wanted occurs. Rarely its worse than filling out
a ballot properly.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 195 bytes
Desc: not available
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20230914/a3f6a169/attachment.sig>