Hardware implementation of a `wired-once´ neural net in thin-film technology on a glass substrate

To prove the feasibility of implementing artificial neural networks on large inexpensive substrates, a net designed and fabricated on a glass wafer using hydrogenated-amorphous-silicon-based technology (a-Si:H) is discussed. The net functions as an autoassociative memory in which binary numbers corresponding to 28, 56, 112, and 224 are stored. Learning of the weight matrix is carried out with the associative memory algorithm using the delta rule. Phosphorus-doped microcrystalline silicon with a resistivity of 100 to 300 Ω-cm was used for the fabrication of the weight (synapse) resistors. Inverters with a beta of one were used to form negative-weight synapses, and inverters with a beta of 10 were used for the thresholding elements (neurons). The net functions surprisingly well; it filters both the learned numbers and some numbers of the form N=4k (with k an integer), and maps other random numbers to the closest one accepted, even though the experimental weight matrix is not identical to the theoretical one