Time-adaptive vector A/D conversion

Traditional analog-to-digital (A/D) conversion is a limited scaler coding technique, unfit for compression, but which can be implemented with very fast, simple, and small analog devices. This work extends the concept of A/D conversion from scalers to vectors. A vector A/D converter offers a level of complexity similar to that of usual scalar A/D converters but operates on a block of k analog inputs so as to perform a mixed conversion/compression task. To adapt to changing source statistics, a backward, unsupervised learning rule is proposed. The rule, called BAR_r(k), attempts to minimize the rth power law distortion in the high resolution case. Adaptation is only determined by the transmitted codeword and can be performed simultaneously at both sides of the channel without any side information. The learning rule includes a forgetting factor to increase robustness in case of transmission errors and initial encoder/decoder mismatches