Has anyone explored simulating independent dendritic pattern detectors by modifying standard ML neurons?
For each neuron, assign I independent sets of fully connected input connections (randomly initialise weights). When forward propagating, take the independent set which delivers the max combined output (Z_i) - ignore the rest. Pass value through activation function (increase the activation threshold to renormalise network activation sparsity). The network will train/optimise only those independent pattern detectors which provide maximum output in a given context.