THE GENESIS MODEL PART III: From Trophic Webs and Hierarchical Energy Transformation Systems to Multilayer Neural Networks

Peter WINIWARTER (I) and Czeslaw CEMPEL (2)

Address -   (1) Bordalier Institute, 41270 BOURSAY, France, fax + 33 5480 1937, e-mail: winiwarter@bordalierinstitute.com

(2) Poznan University of Technology, al. Piotrowo 3, 60-965 Poznan, Poland, fax +48 6178 2307, e-mail: czeslaw.cempel@put.poznan.pl

Abstract.

The key idea put forward in this paper is the formal equivalence between energy transformation processors undergoing birth-and-death processes ("Birth-and-Death processors") and formal neurons. Both can be described as binary threshold automata. The organization of Birth&Death processors in hierarchical feedforward structures (trophic webs) is equivalent to the organization of multilayer feedforward networks of formal neurons. The error-backpropagation in such neural networks can be interpreted as equivalent to the feedback of downgraded energy in trophic webs. Assuming a correlation between feedback and the overall web performance in terms of global energy throughput, an ecosystem can be modeled as a feedforward artificial neural network with error backpropagation being trained through gradient descent. Hence all features of this class of ANNs concerning memory and learning can be applied to the trophic web of an ecosystem and to hierarchical natural and « artificial » energy transformation systems in general. Parcto-Zipf distributions are observed for feedback symptoms in all natural and « artificial » energy transformation hierarchies. Pareto-Zipf distributions are also observed for the backpropagated errors of formal neurons learning along gradient descent.

Key Words - energy transformation, information transformation / transmission, (birth&Death processors, threshold automata, Pareto-Zipf-Frechet-Weibull distributions, trophic webs, neural networks, memory, backpropagation, adaptation, learning, evolution