An orthogonal delta weight estimator for MLP architectures

We present in this paper an extension of the OWE neural network architecture (orthogonal weight estimator). The OWE architecture permits to implement the modelization of context-dependent behavior by dynamically estimating the synaptic weights of a MLP with respect to external parameters named here “context” parameters. The principle of this extension, named ODWE (orthogonal delta weight estimator), is based not on an estimation but on a modulation of the synaptic efficiencies. The interest of this approach is firstly to view the context dependent behavior as a general behavior modulated by the context parameters, and secondly to bring closer ODWE principle and neurobiological knowledge. An illustration on the modelization of a mathematical function is shown at the end of the paper