At this point, the core question became unavoidable: how do you learn structure from motion without defining that structure in advance?
If I manually specified what a “good” stroke looks like, the analysis would simply reflect my own assumptions. That would defeat the purpose. What I needed was a system that could observe many strokes and discover, on its own, how they differ and how they repeat.
This is where sequence-based models became relevant. A putting stroke isn’t a static object—it’s a progression. Position, orientation, and timing all matter in relation to what comes before and after. Treating each frame independently would destroy that context.
The key idea I adopted was reconstruction. Instead of asking a model to judge or classify a stroke, I asked it to reproduce it. If a model could take an entire motion sequence, compress it into a smaller internal representation, and then reconstruct the original sequence with minimal error, that internal representation must be capturing something essential about the stroke.
Crucially, this process doesn’t require labels. There’s no notion of correct or incorrect, only whether the motion can be faithfully represented. The model isn’t told what smoothness is, or what consistency is—it has to infer those concepts implicitly, if they exist in the data at all.
This reframing changed how I thought about analysis. The goal was no longer to explain performance directly, but to learn a compact “fingerprint” of each stroke based purely on its motion dynamics.