Path: ...!feeds.phibee-telecom.net!2.eu.feeder.erje.net!feeder.erje.net!fu-berlin.de!uni-berlin.de!individual.net!not-for-mail
From: rbowman <bowman@montana.com>
Newsgroups: comp.os.linux.advocacy
Subject: Re: If you're a fucking moron
Date: 10 Oct 2024 21:38:47 GMT
Lines: 36
Message-ID: <lmqvr7FiegjU1@mid.individual.net>
References: <dvhpfjd1rh7uheoien02arle31q9fhcd57@4ax.com>
	<8m6dnczkO_GAPpz6nZ2dnZfqnPSdnZ2d@giganews.com>
	<slrnvgdh2b.2nfc6.candycanearter07@candydeb.host.invalid>
	<pan$78b32$3adf0bd2$1deb5e7a$b8c37cfa@linux.rocks>
	<slrnvgg5g4.1gkvp.candycanearter07@candydeb.host.invalid>
Mime-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
X-Trace: individual.net dYwDYdUF1mYFD3VFil+nIwCJZYotQQwWdHpkRnkR3ksmou2aGI
Cancel-Lock: sha1:Jp0qJw10hRSB5pkPg37pmSfMOfI= sha256:77RTtQMX/d8XMupo5mkYdrcqq8epme8pHgvgH7MCT7Y=
User-Agent: Pan/0.149 (Bellevue; 4c157ba)
Bytes: 2487

On Thu, 10 Oct 2024 18:00:07 -0000 (UTC), candycanearter07 wrote:

>> AI, or more properly, pattern matching, can be quite valuable in image
>> sharpening, background removal, and other image processing tasks.
> 
> Agreed, and /some/ tedious tasks, like captioning.

Unfortunately image recognition starts with the tedious task of 
captioning. The 'hello world' of image recognition is a huge dataset of 
dogs and cats, each labeled 'dog' or 'cat'.  Guess who does the labeling. 

Voice recognition is as bad if not worse. There are a variety of accents 
to deal with as well as languages. A real problem is the very skewed 
nature of the available data. English has the most data while Tagalog data 
is sparse/

If you have some spare time:

https://commonvoice.mozilla.org/en

The 'listen' side is revealing. You're shown the words and listens to the 
speaker's pronunciation and score it. Some people try really hard but what 
they produce is unintelligible. The 'speak' side can be humbling. I 
wouldn't make a good politician since it's apparent I can't read from a 
teleprompter :) At least you get a redo. 

There are unsupervised learning techniques that do not require massive 
amounts of stat but they seldom are in the spotlight. An example would be 
a system receiving sound, temperature, vibration, and other data from a 
car engine. In time it learns the baseline parameters of a healthy engine 
and can raise a red flag when you're getting ready to drop a rod.