Supervised Architectures for Internal Simulation of Perceptions and Actions

We present a study of supervised neural network architectures capable of internal simulation of perceptions and actions. These architectures employ the novel Associative Self-Organizing Map (A-SOM) as a hidden layer (for the representation of perceptions), and a neural network adapted by the delta rule as an output layer (for the representation of actions). The A-SOM develops a representation of its input space, but in addition it also learns to associate its activity with an arbitrary number of additional (possibly delayed) inputs. We test architectures, with as well as without, recurrent connections. The simulation results are very encouraging. The architecture without recurrent connections correctly classified 100% of the training samples and 80% of the test samples. After ceasing to receive any input the best of the architectures with recurrent connections was able to continue to produce 100% correct output sequences for 28 epochs (280 iterations), and then to continue with 90% correct output sequences until epoch 42.