Improved winner-take-all neural network

Analogue Globally Stable WTA Neural Circuit

“…In fact, this unique steady state does not depend on the initial conditions because it is built on the basis of second order blocks unlike other WTA circuits where resetting has to be made whenever inputs change [5]. This is convenient for a neural circuit running in a real time [6].…”

“…This net is built using subnets arranged in a layered binary tree to reduce the number of nodes required and includes N-1 comparator subnets which are arranged in roughly log2 N layers when the maximum of N inputs must be selected. Since the network is implemented on the basis of comparators consequently it has a slow convergence rate [6]. T is an 2(N -1) x 2(N -1) matrix.…”

CMOS current-mode implementation of spatiotemporal probabilistic neural networks for speech recognition

“…In the output neuron, the function is back again as fp(X) = exp(Ip) (9) It can be seen from (5) that the SPNN not only inherits on-line real time training capability of the PNN, but provides a faster computation than the PNN in the retrieving phase. This is due to a simpler similarity estimation of the distance measurement instead of the complicated computation as in (4).…”

“…The exemplar STPs are previously encoded in binary format and stored in the weighted memory Wp~, W m ..... WpL. After the conversion of input patterns X, the SPNN then estimates the similarity score fl (X), f2(X) ..... fM(X) between the input STPs and the M-class template STPs according to (6) and (9). In the same time slice, each switch turns on the neuron to instantaneously compute the similarity degree, then immediately turns off.…”

“…Thus, the WTA circuits usually require several iterations to settle down the race, and sometimes this serious competition process fails to decide the final winner [8]. In [9], an adaptive algorithm/MAX was proposed to speed up the convergence rate in each iteration. However, this algorithm is not easily implemented in hardware form.…”