Baseband Filter Banks for Neural Prediction

We propose in this paper a new prediction paradigm, which is based on filter banks for subband decomposition of the sequences to be predicted. Filter banks allow the implementation of a parallel computing system, taking the advantage of a faster and more accurate implementation. In particular, we introduce a novel subband decomposition method yielding baseband sequences that are easier to be predicted. The core of the prediction system is based on a neural model, which is trained for each subband using specific embedding techniques. The latter are used in order to optimize the prediction performances when dealing with real-world data sequences, which often possess a chaotic behavior.