Hardware Implementations of Neural Network Models

Neural network models are receiving widespread attention as new computing architectures for applications such as pattern recognition and machine learning. To obtain the full benefit of these models, special-purpose hardware must be built. Simulations of the highly interconnected neural networks on standard computers are very time-consuming. For most applications simulated neural networks are far too slow to be considered useful. Various research groups have built neural network circuits or have proposed designs. Most of these circuits apply analog computation to some extent. The high interconnectivity and the moderate precision required make these networks well suited for analog computation. One of the key operation performed by neural network models is evaluating sums of products. Using analog computation a multiplication can be done with a single resistor and summing of currents is done `for free` on a wire. Therefore, an analog circuit that computes sums of products can be built more compactly than a digital circuit.