A connectionist approach to trend detection

A connectionist architecture for the detection of the onset of transients is presented. The approach utilizes the temporal information to achieve a robust trend detection. It does not require any knowledge of the underlying process that produces the signal. The linearly increasing or decreasing, and step changes as well as constant differences between the current and the previous value in time are accounted for by primitives. Two networks, the Shift-TNN which processes the actual signal and the other, the Diff-TNN which processes the differenced signal, are developed along with their corresponding primitives. The results indicate that the presented approach is capable of detecting trends at onset even at low signal-to-noise ratios.