Foundations
Section 2.1

The Rashomon Effect

Many equally good predictors can implement different mechanisms.

Suppose we define the set of nearly optimal models

Rϵ={f:L(f)L+ϵ}.\mathcal{R}_\epsilon = \left\{ f : L(f) \leq L^* + \epsilon \right\}.

This is a Rashomon set: a collection of models that are essentially equivalent according to the evaluation criterion.

Many good models

Good predictive performance need not identify a unique model.

Different mechanisms

Two members of the Rashomon set may rely on completely different features.

Why this matters

The benchmark tells us that the model works.

It does not tell us how it works.

Interpretable Intelligence synthesizes work from across the field, including research from Guide Labs. Relevant authors and results are cited throughout.