Dear friends,
I've stumbled upon a controversial (to me) manuscript:
https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/excess-capacity-learning/909013EF15575CF119FD511904CDF0C3 (Excess
Capacity Learning, by Dubova and Sloman, appearing on CUP-edited
Behavioral and Brain Sciences).
In essence the claim is: double descent in NN training smells like the
human cognitive system. My very rough distillation, following a cursory
read, is: three regimes of representational capacity are laid out.
i. constrained - not enough: I love food, remember burgers taste good to
me, sort of can tell a burger joint if I see one;
ii. sufficient - I go to burger-making places to have one;
iii. excess - I retain more information than necessary (e.g. day of week
on which I had a bad burger).
I don't need to know the day of the week, but given a bad burger
experience on a given day, I will refrain from eating one on the same
day. Implication: maybe on Wednesdays there's a different burger
flipper at work.
Here's a proper synthesis:
https://www.santafe.edu/news-center/news/upending-assumptions-about-learning-inspired-by-an-ai-phenomenon
I don't buy it. For starters, I see a circular reasoning: the moment an
"excess signal" becomes predictive, it stops being "excess" and becomes
"sufficient", in their parlance.
The editors are requesting commentary here*, and I can't think of a more
apt congregation to throw this curveball.
Thank you for your attention to this matter,
Cheers,
-Matteo
*https://drive.google.com/file/d/1XkBQ2K0_hIz4Se8b0cOUKreM96ONBqqP/view <<<
link appears on the SFI webpage, assumedly safe
.- .-.. .-.. / ..-. --- --- - . .-. ... / .- .-. . / .-- .-. --- -. --. / ...
--- -- . / .- .-. . / ..- ... . ..-. ..- .-..
FRIAM Applied Complexity Group listserv
Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p Zoom
https://bit.ly/virtualfriam
to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
FRIAM-COMIC http://friam-comic.blogspot.com/
archives: 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/
1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
