The algorithm of backpropagation error allows training a hidden layer neural network (perceptron type), but still far from the complexity of biological networks in which the property reentrancy is paramount and for which it would be unrealistic want to follow the flow of nerve impulses. This is why global methods have been proposed on networks almost completely connected (ie where the axon of each neuron is connected to almost all the dendrites of other neurons) as the Kohonen networks.


Assuming that an organism is satisfied if it is in agreement with its environment (that is to say if the perception that it has of the result of its actions on this environment is in agreement with them) one can make the hypothesis that the flow of his actions is consistent with the flow of his perceptions. In terms of signal processing we will say that these two flows are in phase.
For almost periodic signals we will try to align their periods and, for any signals, we will seek to parallelize their variations (they grow and decrease at the same time).


I employed this method in the connexionist mime making coherent the video image analyse of the real mime and the analyse of the virtual mime actions.