Neuron-Wise Inhibition Zones and Auditory Experiments

Spatiotemporal processing, including many tasks such as speech recognition, object tracking, and natural language processing, still faces great challenges. The developmental network-1 (DN-1) was proposed as a spatiotemporal processing algorithm with general purposes in the sense of turing machines (TM). DN-1 used emergent motor vectors as contexts for top-down attention, which is conducive for temporal processing. This also enables finite automaton (FA), which is the control of TM, to incrementally emerge inside DN-1. But the hidden areas of DN-1 have a fixed boundary. In the proposed new model-developmental network-2 (DN-2)-each hidden neuron automatically discovers its neuron-wise inhibition zone based on incremental statistics. Therefore, not only the number of hidden areas is dynamic, so are their interconnections. We first present this new network through a visual toy problem. Then, we conducted phoneme recognition experiments to investigate DN-2's performance in various settings. Based on the experimental results, we analyzed how the new mechanisms enabled DN-2 to automatically generate a dynamic hierarchy and improve the performance. This paper is a step toward our goal of applying DN-2 to a wide variety of spatiotemporal processing tasks.