zoNNscan: a boundary-entropy index for zone inspection of neural models
Résumé
The training of deep neural network classifiers results in decision boundaries whose geometry is still not well understood. This is in direct relation with classification problems such as so called corner case inputs. We introduce zoNNscan, an index that is intended to inform on the boundary uncertainty (in terms of the presence of other classes) around one given input datapoint. It is based on confidence entropy, and is implemented through Monte Carlo sampling in the multidimensional ball surrounding that input. We detail the zoNNscan index, give an algorithm for approximating it, and finally illustrate its benefits on three applications.
Origine : Fichiers produits par l'(les) auteur(s)