Generative Models from the perspective of Continual Learning

Timothée Lesort (1) , Hugo Caselles-Dupré (1) , Michael Garcia-Ortiz (2) , Jean-François Goudou (3) , David Filliat (1, 4)

1 U2IS - Unité d'Informatique et d'Ingénierie des Systèmes
2 SoftBank Robotics Europe
3 Thales Services
4 Flowers - Flowing Epigenetic Robots and Systems

Which generative model is the most suitable for Continual Learning? This paper aims at evaluating and comparing generative models on disjoint sequential image generation tasks. We investigate how several models learn and forget, considering various strategies: rehearsal, regularization, generative replay and fine-tuning. We used two quantitative metrics to estimate the generation quality and memory ability. We experiment with sequential tasks on three commonly used benchmarks for Continual Learning (MNIST, Fashion MNIST). We found that among all models, the original GAN performs best and among Continual Learning strategies, gener-ative replay outperforms all other methods.

Domaines: Apprentissage [cs.LG]

Timothée Lesort, Hugo Caselles-Dupré, Michael Garcia-Ortiz, Jean-François Goudou, David Filliat. Generative Models from the perspective of Continual Learning. IJCNN - International Joint Conference on Neural Networks, Jul 2019, Budapest, Hungary. ⟨hal-01951954⟩