Estimating similarity of neural network dynamics 
TitleEstimating similarity of neural network dynamics
Publication TypeReport
Year of Publication2010
AuthorsKomosinski, M, Rosinski, K
Document NumberRA-10/10
InstitutionPoznan University of Technology, Institute of Computing Science

This report concerns estimation of the similarity between neural networks of any topology. Motivations and benefits of having an automated and quantitative network comparison mechanism are presented. The concept of neural network dynamics (neuron output signal) is considered. A measure is proposed for estimating similarity of active (i.e., working) neural networks. Properties of the measure are analyzed theoretically and verified empirically. The experiments have been performed on a set of evolved networks responsible for controlling 3D structures (agents, robots). These experiments demonstrate the capabilities and the limitations of the proposed measure as a mechanism to support humans in analyzing large sets of neural networks.