AI
Research1
We
learned a lot about neural nets just by tinkering around with one. If you want
your own neural net, we suggest Tariq Rashid’s (2016) book “Make Your Own Network,”
available on Kindle. We finally got this
network to operate coherently by increasing the hidden layer nodes from 3
to 300. The hidden layer of a neural network, according to Gemini, “…acts as an
intermediate layer that transforms the raw input data into a more complex
representation (which no one really understands, it’s probably just error
correction), that the output layer can then use to make predictions.” This is
why when there is a lot of data, the GPUs have to get
more numerous and bigger.” The Meta network reportedly intends to add 350,000
parallel processing Nvidia H100 GPUs, within which the neural network model
resides. The investment implications, however, may be quite different, because
the trend now is to use smaller networks to explore more defined problems, with
an increasing number of competitors.
We
use a neural net to pose a simple and constant numerical question, computers
represent words by numbers, [.10,.10,.10], to a neural
net where an extreme leadership sets increasingly extreme “targets.” What’s
going to happen? Here is the result:
Run 1 Run 2 Run 3
Target
Training Range [0,0,0] to [.10,.10,.10] [.20,.20,.20] [.30,.30,.30]
Question [.10,.10,.10] [.10,.10,.10] [.10,.10,.10]
_
Network
Answers: x .10 .17 .26
Standard
Deviation s .012 .021 .043
This
table shows that as the range of answers to the same question become more
extreme, so does the volatility of “acceptable” answers (volatility 4x),
swamping the previous consensus. In Run 1, this table also suggests, for a
democratic society, there has to be a continuing consensus.
Researchers
in AI explore different kinds of architectures for different kinds of problems.
If you are a decisionmaker, you are looking to the researcher (network
architecture), the data, and the answers.
AND
“Simplicity is complexity resolved.” -
Constantin Brancusi