An analysis of craniocentric and oculocentric coding stages in a neural network model of the saccadic system

We have shown earlier in a neural network study of the saccadic system, how retinal error and an efference copy signal of eye position may give rise to distributed coding of target position in craniocentric coordinates at one level, and of motor error in oculocentric coordinates at another stage. In the present paper, the coding properties of units in the model's two hidden layers were investigated, in order to understand at a more abstract level, how they handle their inputs and how the two different target representations at subsequent stages emerge. In particular, we hoped to understand better how craniocentric and oculocentric target representations can be constructed by merging a retino-topically coded visual signal and recruitment-coded eye position information. In the first hidden layer, we found that inputs from both visual and oculomotor signals were nicely matched in showing similar directional selectivity. Computationally, the net input of each hidden unit can be represented by the dot product between a fixed sensitivity vector, embodied by the unit's input weights, and the two-dimensional input signal encoded by the population activity. Scaling of the resulting total net input signal through a sigmoidal nonlinearity then yields the activity of the hidden unit. The fact that the sensitivity vectors for retinal and oculomotor signals in the first hidden layer were roughly aligned and matched in amplitude is the basic underlying principle for a rough craniocentric coding at this level. Units in the second hidden layer represent motor error. This can similarly be understood on the basis of the previously mentioned dot product characterization of the hidden unit's connectivity. The combined process of tuned projection and compression by the unit's sigmoidal nonlinearity also captures the gain-field properties of units in the first hidden layer. Our study suggests that the approach underlying the present analysis of an artificial network may be a useful tool to describe real networks and to allow a direct comparison of simulated and neurophysiological data. Copyright (C) 1996 Elsevier Science Ltd.