Selection Criteria for Neuromanifolds of Stochastic Dynamics

Nihat Ay, Guido Montúfar, and Johannes Rauh

Contact the author: Please use for correspondence this email.
Submission date: 19. Apr. 2011
Pages: 6
published in: Advances in cognitive neurodynamics III : proceedings of the 3rd International Conference on Cognitive Neurodynamics 2011 ; [June 9-13, 2011, Hilton Niseko Village, Hokkaido, Japan] / Y. Yamaguchi (ed.)
Dordrecht : Springer, 2013. - P. 147 - 154
(Advances in cognitive neurodynamics) 
DOI number (of the published article): 10.1007/978-94-007-4792-0_20
Bibtex
MSC-Numbers: 62B10, 82C32
Keywords and phrases: learning, neural nets, information geometry
Download full preprint: PDF (2108 kB)

Abstract:
We present ways of defining neuromanifolds – models of stochastic matrices – that are compatible with the maximization of an objective function (reward in reinforcement learning, predictive information in robotics, information flow in neural networks). Our approach is based on information geometry and aims at the reduction of model parameters with the hope to improve gradient learning processes. We discuss advantages and shortcomings of this approach.