Hugues Bersini

Twenty-five years ago, I wrote the first article dedicated to recurrent fuzzy systems that, by simple imitation of recurrent neural networks, extends the temporal memory of these fuzzy systems. As the article show, they appeared to work pretty well on simple time series predictions and non-linear process optimal control. After the so-called winter of neural networks, and on account of a new shaping that avoids gradient vanishing, these recurrent neural networks have regained a high level of popularity for language and music processing, pictures labelling, and session-based products recommendation (for instance the famous Netflix movie recommendation challenge). I will show and discuss the impressive good performance of GRU neural networks that we obtained for short term and long-term movie recommendation. Consequently, it is time now, once again, to adapt to fuzzy systems this new reshaping of recurrent neural networks and the new learning mechanism associated to them.