High speed color recognition with an analog neural network chip

The neuro chip introduced is a classificator which is intended for fast classification of signal vectors up to the width of 10. It consists of analog components. The width of the output vector is also 10. Due to the implementation of analog hardware, the chip works fully parallel and needs less than 5 /spl mu/s to recognize a pattern. The analog approach necessitates that capacitive storage elements are used for storing synaptic weights. The storage of analog voltages in a capacitor of only 1 pF with a precision of more than 6 bit is possible for a period of time of up to several minutes by suitable circuit technique. To fulfill vector-matrix multiplications, two arrays of 66 and 70 analog multipliers are integrated. The advantage of the analog approach in terms of speed, however, requires a high effort in modelling complex transfer function. We show that the circuit is able to perform color recognition tasks in combination with an analog sensor. Results show that color recognition can be achieved with a precision sufficient for the demands of the human eye. By segmentation of the color space, the neural network can be trained with a precision beyond the spectral resolution of the human eye.