Nicolas Szilas Research -->   Recurrent Neural Networks

Learning in recurrent nets is difficult...

Coupling is a new idea for facilitate learning.

The coupling-decoupling learning algorithm.

Recurrent Neural Networks can be used to learn to reproduce a given temporal sequence or signal. No satisfactory algorithm exists for fully recurrent neural networks, in terms of time and storage needed to achieve learning.
We investigated the idea of coupling the network with the external object producing the signal (in our application, a mechanical object - see my Physical Modeling page). Unfortunately, with coupling, the exact calculation of learning formulae is impossible, in most of cases. But the challenge is to examine whether coupling, a basic form of interactive learning, could facilitate learning.
Interestingly, coupling is related to the concept of physical guidance, in human learning...
Finally we investigated a new learning algorithm where the function that is minimized is not the difference (mean squared error for example) between the networks and the learned signal, but the difference between a first coupled network and a second "free" network.
   RR 972-I


( HOME )