Low-Offset Neural Winner-take-all Network

Título: Low-Offset Neural Winner-take-all Network

Autores: Pedroni, Volnei A.

Resumo: Winner-take-all (wta) circuits are common building blocks in neural networks, vector quantizers, and other analog parallel signal processing systems. We present a wta circuit that employs a Hopfield-like architecture for the transmission of the positive-feedback coefficients over the 2-D computing array. The properties of this kind of network are further illustrated by means of a 32-input VLSI implementation on a 2.0 μm CMOS chip. Experimental results show a high voltage gain (so digital outputs are immediately available) and very small offsets (under 10mV in the worst-case scenario), with an analog dynamic range resolution of approximately 50dB.

Palavras-chave:

Páginas: 4

Código DOI: 10.21528/CBRN1994-028

Artigo em PDF: CBRN1994-paper28.pdf

Arquivo BibTex: CBRN1994-paper28.bib