Online Unsupervised Neural-Gas Learning Method for Infinite Data Streams

Abstract : We propose an unsupervised online learning method based on the "growing neural gas" algorithm (GNG), for a data-stream configuration where each incoming data is visited only once and used to incrementally update the learned model as soon as it is available. The method maintains a model as a dynamically evolving graph topology of data-representatives that we call neurons. Unlike usual incremental learning methods, it avoids the sensitivity to initialization parameters by using an adaptive parameter-free distance threshold to produce new neurons. Moreover, the proposed method performs a merging process which uses a distance-based probabilistic criterion to eventually merge neurons. This allows the algorithm to preserve a good computational efficiency over infinite time. Experiments on different real datasets, show that the proposed method is competitive with existing algorithms of the same family, while being independent of sensitive parameters and being able to maintain fewer neurons, which makes it convenient for learning from infinite data-streams.