J. Bernardi

An atlas of models taken apart

Julien
Bernardi

AI/ML student at Université PSL. I take models apart to see what they actually learned.

CurrentlyM2 IASD @ PSL · AI engineer @ AMIAD · Paris region

Fig. 1. The site as a loss landscape. Drag the ball into a basin, or pick a label.Pick a section.

Latest project

NLP · 2026

Interpretable text classification at INSEE

Lightweight, explainable classifiers that file French company descriptions under ~700 NAF codes, with a model that predicts all five levels at once.

“Completely wrong” errors, multi-level heads29–34%

Read the project→All work→

Small print

Keyboard & errata

g i
Go to the index
g w
Go to work
g p
Go to the field log
g a
Go to about
g o
Go to now
t
Paper / darkroom
[ ]
Previous / next project
?
This sheet
  1. Index For “models taken apart”, read “models that took me apart, a little, each time”.
  2. GANs Some rejected digits were, on reflection, perfectly good digits.
  3. Work The list shows every project but one. See the missing one →