A learning scheme for bipartite recurrent networks and its performance
Itsuo Kumazawa, M. Fukuda · 2002
A new learning scheme specialized to recurrent neural networks with the bipartite topology is proposed. The scheme is expected to have better convergence than the general Boltzmann machine learning. This improvement results from the restricted form of the network topology and an energy form devised to have a dominant global minimum. Compared to the recurrent backpropagation algorithm, the scheme is simple, more suitable for hardware realization and its probabilistic nature reduces the effect of spurious local minima. The performance of the scheme is partly demonstrated by simulations of associative memory and compared with the general Boltzmann machine learning.>