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.