the workshop

Kolmogorov–Arnold Networks

the code my audit

My MSc thesis. I used a KAN to fill in missing measurements of a quantum detector before rebuilding it, then audited my own results.

f(x) q q( p q,p(xp) )

a learnt curve on every connection

.01546 .00573 2.7.00589 .00572 1.0

predict

Keep a quarter of the brightness levels. What happens to the error?

answer

Raw goes from .0057 to .0155; the KAN stays at .0059. But my audit found raw is smoothed 4 times harder there, so I’m rerunning it fairly.

the 93 brightness levels i measured.005.010.015.020100604025levels kept, in 100error on the levels held outrawkan2.4 times
fig. 1  a detector, starved of dataDrag the pencil to keep fewer levels.

25 in 100: raw .0141, kan .0059

my audit: at 25 in 100 the raw fit is smoothed 4 times harder, so it isn’t a fair race yet

challenge find where the KAN starts to pay off

key words

brightness level
one strength of light the detector is tested with
SDP
the solver that turns clicks into a valid detector