A supervised strategy for deep kernel machine

Florian Yger (1) , Maxime Berar (1) , Gilles Gasso (2) , Alain Rakotomamonjy (2)

1 DocApp - LITIS - Equipe Apprentissage
2 LITIS - Laboratoire d'Informatique, de Traitement de l'Information et des Systèmes

This paper presents an alternative to the supervised KPCA based approach for learning a Multilayer Kernel Machine (MKM) . In our proposed procedure, the hidden layers are learnt in a supervised fashion based on kernel partial least squares regression. The main interest resides in a simplified learning scheme as the obtained hidden features are automatically ranked according to their correlation with the target outputs. The approach is illustrated on small scale real world applications and shows compelling evidences.

Domaines: Apprentissage [cs.LG]

Florian Yger, Maxime Berar, Gilles Gasso, Alain Rakotomamonjy. A supervised strategy for deep kernel machine. European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Apr 2011, Bruges, Belgium. pp.501-506. ⟨hal-00668302⟩