Learning complex temporal sequence using bi-directional spatiotemporal neural network

The authors propose a bi-directional spatiotemporal neural network (BSNN) paradigm capable of learning complex temporal sequences. The parallel architecture employs a fully connected structure, in that an input layer holds the temporal weights (i.e., long-term memory LTM) incorporated with a set of short-term memory (STM) units to provide sequence detecting capability. A two-pass training algorithm is developed to efficiently learn LTM during the forward pass, and thresholding values during the backward pass. Experimental results show that accurate sequence detecting power and rejection to erroneous input sequences are obtainable with BSNN.