James Warren <
jwwar...@gmail.com> writes:
> Gosh Mike. Even Quantum Physics is statistical at the bottom.
> Individual quantum events can't be predicted exactly only their
> statistical properties. Only the probabilities are knowable.
Yeah, yeah, I know. An electron isn't knowably anywhere in particular
until it's "observed". The wave equation is a probability wave, a
wave-like function that asserts the the loci or surface in which the
electron has a non-zero probability of being found. I didn't
understand thermodyanmics until I took a course that dealt with
statistical mechanics. I failed the course due to my weak math
ability but learned a lot.
> You will agree that statistics is part of mathematics, right?
Oh, sure. Despite the fact that there remains (AFAICT) an unresolved
controversy within the math world over what, exactly, "probability"
is. :-)
The problem arises when the subject domain is sufficiently complex
and/or emerges from insufficiently understood events/interactions that
common sense fails, analytical methods result in insoluble equations
and/or no one knows how to collect, organize or filter data that might
promise understanding.
So you count something and do statistics.
The current Farcebook fiasco is, AFAICT, an example. FB has worked
out an algorithm that catalogs clicks (or some more complicated
measure of "engagement") and their contexts, selects contexts that
increase the probability of increased "engagement" and optimize for
those contexts. The daily 1000 (?) clicks by each of a billion users
is more than sufficient to make statistical analysis workable. It
turns out that the most engaging contexts are ones that exploit
credulity, gullibility, resentment, fear and a sense of victimhood.
Business is booming thanks to statistics.
But, OTOH, when this is seen to be harmful and even destructive,
propagating and (apparently) validating misinformation, propaganda,
overt lies, deranged fantasies and other social/cognitive pathologies,
they're stumped. The tools of statistics fail to encompass the
domain of natural language semantics, even moreso the neural and
social domains of mind.
"Big data" and "artificial intelligence" -- today that means neural
nets -- kind of end-run statistics per se and notionally offer a new
math-related option for sufferers from physics envy. But the early
promise that emerged from stuff like image recognition is looking
shaky when applied to more difficult domains. Neural nets have to be
"trained". Starting with random numbers to determine outputs in
response to inputs, a human or a human-devised catalog of input +
yes/no answer pairs must be run through the NN and the outputs compared
to the catalog. When the output is wrong (unwanted, deprecated) the
matrices of initially random numbers are nudged ever so slightly in
the direction of ones that would have given desired output. After a
large number of such training cycles (a million? a billion? limited
only by the power of the computer and the size of & confidence in the
"training data" corpus) the NN/AI can be given real-world inputs with
some expectation (possibly specious) that the output will be "good"
(in whatever sense "good" has in the context).
This has already been seen to fail of its glowing, quasi-magical
promise and be fraught with pitfalls. NNs trained on a corpus of
natural language or other data of the sort that FB might expect to
encounter has emerged from training exhibiting racism & sexism and
telling lies. [1] No one knows -- it's probably intrinsically
unknowable -- exactly why this should be because the details of
discrimination are buried in vast, impenetrable matrices of real
numbers no one of which has any significance. That probably won't be
an more tidily "fixed" by employing "better data scientists" that the
failures in, say, medicine are fixed by better statisticians.
AFAIK from bumbling around the net, no one has a clear handle on why
ca. 40% of the US adult population should be terminally gullible.
Demographic statistics haven't helped. We've known about con games
for millennia, about "Never drop the con; die with the lie" for at
least a century. Why it's become a mass phenomenon hasn't been
resolved by statistics, even if, as it appears, it's been at least
partly caused by the application of statistics. Cognitive and neuro
science offer some tentative explanations for the force of religion but
statistics isn't helping that much AFAIK.
The behavior of people *can* be examined by statistical tools. But
any definitive explanations lie in the brain (we don't understand the
human brain) and mind (no one knows what a mind is) so statistical
tools are prone to shingling off onto the mathematical fog.
Complexity -- the emergent behavior of systems with a large number of
components, very many pairs of which influence each other -- adds yet
another difficulty. Complex systems in the abstract are difficult to
comprehend. More tangibly, we're still struggling with biology
because the endocrine system, the immune system or even the community
of gut microflora are such systems. When the components are a billion
minds, influencing one another over the noisy channel of natural
language, it's more difficult yet. Neither statistics nor NNs analyze
or explain the operation or principles of such systems and are useful
but blunt instruments. When it comes to social (as oposed to
biochemical) systems, none of us regards h{is,er} role/place in the
world as analogous to that of a particle of ideal gas in a box, nor do
we wish to be treated as if it were.
[1]
https://news.un.org/en/story/2020/12/1080192
or Gwglw for others.