DISentangled Counterfactual Visual interpretER (DISCOVER) generalizes to natural images

TL;DR

DISCOVER is a generalized method for visual interpretability of image classifiers, successfully applied to both biomedical and natural images, revealing meaningful discriminative features.

Contribution

This paper extends DISCOVER's applicability to natural images, demonstrating its effectiveness in interpreting facial features across different domains.

Findings

DISCOVER identified key facial features distinguishing dogs and cats.

It successfully interpreted facial traits like cheeks, jawline, and eyes.

The method proved effective across biomedical and natural image domains.

Abstract

We recently presented DISentangled COunterfactual Visual interpretER (DISCOVER), a method toward systematic visual interpretability of image-based classification models and demonstrated its applicability to two biomedical domains. Here we demonstrate that DISCOVER can be applied to the domain of natural images. First, DISCOVER visually interpreted the nose size, the muzzle area, and the face size as semantic discriminative visual traits discriminating between facial images of dogs versus cats. Second, DISCOVER visually interpreted the cheeks and jawline, eyebrows and hair, and the eyes, as discriminative facial characteristics. These successful visual interpretations across two natural images domains indicate that DISCOVER is a generalized interpretability method.