Neural network modeling supports a theory on the hierarchical control of prehension

A theory on the hierarchical organization of the control of human prehension (grasping and manipulation of a hand-held object) was tested by comparing the performances of neural networks of different designs. The inputs into the networks were external torque, handle width, and thumb location, and the outputs were the individual digit forces. The networks differed only in their architecture: Nl was a classical three-layer network; N2 was a hierarchical two-tier network with single projections, in which the outputs of the first tier were used as inputs for the second tier, that yielded the individual digit forces; and N3 was a hierarchical two-tier network with dual projections, where the inputs to the second tier were the outputs of the first tier-as in N2-plus the inputs into the first tier (external torque, handle width, and thumb location). Each tier of N2 and N 3 consisted of one three-layer network. The N3 network showed the best performance, supporting the idea that the control of prehension is hierarchically organized.