How SheShe Works

SheShe transforms a probabilistic model into an explorer of its own decision landscape. It searches for local maxima in class probability or predicted value and can outline regions around the discovered modes.

Mathematical foundation

Optimisation follows gradient-ascent updates x_{k+1} = x_k + α∇f(x_k) until the gradient norm falls below a small ε.

Boundary scans shoot rays from each peak and locate inflection points where d²f/dr² = 0 with a change in concavity, or a percentile drop is reached, delimiting regions around every mode. See ShuShu and ModalBoundaryClustering for details.

Components

ModalBoundaryClustering

This clustering approach scans score surfaces for separated modes and draws supervised boundaries around each one. It performs its own optimisation and does not depend on ShuShu or CheChe.

ShuShu

ShuShu is a lightweight gradient-ascent optimiser that can probe any score function. It locates local maxima independently and can be combined with other tools as needed.

CheChe

CheChe projects optimisation paths into two dimensions and draws convex-hull frontiers for selected feature pairs. It operates separately from ModalBoundaryClustering and ShuShu to provide visual analysis.

These independent components let SheShe reveal the structure of complex models by following their peaks and mapping boundaries in whichever way best suits the task.