Neural network hardware: what does the future hold?

Neural network technology has reached the stage where practical applications are appearing. However, the computational resources required to perform the data manipulation are large and complex. Hardware to perform neural network computations has been slow to appear. The causes for these limitations are reviewed, along with the state-of-the-art in neural network hardware. Novel methods to circumvent these limitations are suggested: hybrid optical/electronic systems neural network algorithms specifically formatted to match to digital semiconductor chips; fine grained parallel simulators using a novel neuron value transfer; and massive analog semiconductor chips.<
>