A two-layer Perceptron with 2×2 input neurons, 2×2 hidden neurons, and one output neuron is electro-optically implemented. Holographic lenslet arrays are used to implement neural interconnections, and time-sharing of single layer interconnections is utilized. In addition, positive and negative interconnection weights are encoded by using subrows in space and error back propagation is optically implemented. The possibility of solving nonlinear separable problem of pattern classification using this scheme is experimentally demonstrated.