Abstract
The cytomatrix neuron is a softened cellular automaton, roughly motivated by interactions that could occur in a molecular or cellular complex. Subcells exert graded influences on each other and provide a medium for the integration of input signals in space and time. A readout element located in a suitably activated subcell triggers an output. The neurons are trained through variation-selection learning that acts on multiple dynamical parameters. Extensive experimentation with the model shows that the dynamics can be molded to yield different structures of generalization. Dimensionality can be increased by increasing the number of dynamical parameters open to variation and selection. Learning algorithms that vary the greatest number of parameters were found to have a greater variability in the structures of generalization and to yield higher performance values and learning rates. The focus here is on n-bit exclusive- OR tasks that are known to be hard due to their linear inseparability. The system successfully learned 2-bit and 4-bit exclusive-OR functions. The higher dimensional algorithms exhibited a relatively good performance on the 8-bit exclusive-OR function.