Week 9: What a model sees
Stroke-level machine learning as a second, tireless opinion --- and exactly what it can't tell you.
3 May 2027
Guest week, taught by Wren Okafor. Stroke-level attribution models — the line of work that includes Ahmed Elgammal’s stroke-based analysis of Picasso and Matisse drawings, and brush-stroke classifiers trained specifically on Chinese ink-wash painting — learn to separate hands from low-level features of how a line is drawn: curvature, pressure variation, the fine texture a human eye tends to read holistically rather than measure. Applied to a securely attributed training set, a model like this can flag a disputed painting as closer to one hand than another with a precision no connoisseur would claim for their own eye.
That precision is exactly the trap. The model only ever tells you what it was trained to recognise, which means it inherits every misattribution already baked into its training set, states its answer with total confidence regardless of whether the question was fair, and cannot explain its reasoning in terms a human argument can weigh against the seal or the colophon. Wren’s own framing, used all week: a model that agrees with the connoisseurs is not yet interesting, and one that disagrees is not yet trustworthy.
Outline#
- how stroke-level attribution models actually work, at a level you could explain to a non-technical reader
- what training-set bias does to a model’s apparent confidence
- running a small classifier against this course’s own shortlisted paintings, live
- where computational and human methods should and shouldn’t be combined
Teaching team
- Wren Okafor — guest teacher, week 9