Parametric Sensitivity and Scalability of k-Winners-Take-All Networks with a Single State Variable and Infinity-Gain Activation Functions
In recent years, several k-winners-take-all (kWTA) neural networks were developed based on a quadratic programming formulation. In particular, a continuous-time kWTA network with a single state variable and its discrete-time counterpart were developed recently. These kWTA networks have proven properties of global convergence and simple architectures. Starting with problem formulations, this paper reviews related existing kWTA networks and extends the existing kWTA networks with piecewise linear activation functions to the ones with high-gain activation functions. The paper then presents experimental results of the continuous-time and discrete-time kWTA networks with infinity-gain activation functions. The results show that the kWTA networks are parametrically robust and dimensionally scalable in terms of problem size and convergence rate.
J. Wang and Z. Guo, "Parametric Sensitivity and Scalability of k-Winners-Take-All Networks with a Single State Variable and Infinity-Gain Activation Functions," Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 6063 LNCS, no. PART 1, pp. 77 - 85, Springer Verlag, Jun 2010.
The definitive version is available at https://doi.org/10.1007/978-3-642-13278-0_11
7th International Symposium on Neural Networks, ISNN 2010 (2010: Jun. 6-9, Shanghai, China)
Research Grants Council of the Hong Kong Special Administrative Region, China
Keywords and Phrases
K Winners-Take-All; Optimization; Parametric Sensitivity; Recurrent Neural Networks; Scalability
International Standard Book Number (ISBN)
International Standard Serial Number (ISSN)
Article - Conference proceedings
© 2010 Springer Verlag, All rights reserved.
01 Jun 2010
The work described in this paper was supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project no. CUHK417608E).