Título: A Capacitive Adaptive Memory for Neural Applications
Autores: Pedroni, Volnei A.
Resumo: Due to their simplicity and relatively small silicon area, hardware realizations of neural networks (NN) often rely on capacitors for storing the synaptic weights. The weights, however, must be constantly changed (adapted) during the learning process. Circuits capable of incrementing-decrementing these stored (analog) values on-board are generally complex, resulting in large-area MOS implementations. For this reason, learning is often done off-board, being the weights loaded onto the NN later. This paper addresses a capacitive circuit which implements an incrementing-decrementing scheme directly on the storage medium. The circuit makes use of a second (small) capacitor, which transfers charge to or removes charge from the main (storage) capacitor, hence increasing or decreasing its voltage. The variation of the stored voltage is exponential, changing faster in the beginning of the adaptation procedure and slower at the end.
Código DOI: 10.21528/CBRN2001-009
Artigo em pdf: 5cbrn_009.pdf
Arquivo BibTex: 5cbrn_009.bib