Batch Learning Growing Neural Gas for Sequential Point Cloud Processing

Abstract

This papers describes a learning algorithm for growing neural gas to construct a topology-preserving map from a 3D point cloud whose topology can change dynamically. Growing Neural Gas with Utility Factor (GNG-U) has been presented as a method for learning the topology of a 3D space environment and applying it to non-stationary or dynamic data distribution. However, when a node is added to an existing network after several errors with sampling data have accumulated, it is difficult for a standard GNG-U to considerably boost learning speed. As a result, we propose a revolutionary growth strategy that dramatically accelerates learning and convergence. This method immediately adds a sample of data as a new node to an existing network based on the likelihood of node addition estimated by the distance to the third closest node and the first and second closest nodes at maximum. Experiment findings show that the proposed algorithm’s network can quickly adapt to represent the topology of non-stationary input distributions.