Visual explanation of simple neural networks using interactive rainbow boxes

Jean-Baptiste Lamy (1) , Rosy Tsopra (2, 3)

1 LIMICS - Laboratoire d'Informatique Médicale et Ingénierie des Connaissances en e-Santé
2 UFR SMBH - UFR Santé, Médecine et Biologie Humaine
3 AP-HP - Assistance publique - Hôpitaux de Paris (AP-HP)

Artificial neural networks are machine-learning algorithms inspired by biological neural networks. Their main inconvenient is their "black-box" nature: while they are very efficient for making predictions, it is difficult to explain these predictions. In this paper, we propose a visual translation of the reasoning performed by simple neural networks, i.e. without hidden layers. This visualization relies on rainbow boxes, a recently-introduced technique for set visualization, and on three improvements we propose for rainbow boxes, including interactivity. We also present a small application of the proposed approach to decision support in antibiotherapy, for helping a physician to choose an antibiotic in urinary infections. Index Terms-Explainable artificial intelligence (XAI), artificial neural networks, medical decision support, set visualization.

Domaines: Santé publique et épidémiologie; Intelligence artificielle [cs.AI]; Interface homme-machine [cs.HC]

Jean-Baptiste Lamy, Rosy Tsopra. Visual explanation of simple neural networks using interactive rainbow boxes. 23rd International Conference Information Visualisation (iV), Jul 2019, Paris, France. ⟨hal-02264225⟩