Browse Prior Art Database

Improved Neural Network Circuit

IP.com Disclosure Number: IPCOM000101764D
Original Publication Date: 1990-Aug-01
Included in the Prior Art Database: 2005-Mar-16
Document File: 2 page(s) / 65K

Publishing Venue

IBM

Related People

Schulz, RA: AUTHOR

Abstract

Analog circuit implementations of neural networks are attractive because the large number of multiplications required for the synapse weightings can be done efficiently and quickly. On the other hand, field programmable implementations require digital storage because analog memories are not practical. The circuit shown in the figure performs analog multiplications and also efficiently employs stored digital synapse weightings.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 55% of the total text.

Improved Neural Network Circuit

       Analog circuit implementations of neural networks are
attractive because the large number of multiplications required for
the synapse weightings can be done efficiently and quickly.  On the
other hand, field programmable implementations require digital
storage because analog memories are not practical.  The circuit shown
in the figure performs analog multiplications and also efficiently
employs stored digital synapse weightings.

      For simplicity, the circuit in the figure is shown with only
one input, V.  The weighting function for this input is stored
digitally in a recirculating shift register 10 excited by a fixed
high frequency clock 12.  The stored weighting data is not
conventional binary code.  Weighting information is contained in the
duty cycle of the data stored in the shift register.  A large
negative weighting signal would consist of mostly zero's, a large
positive  weighting signal is mostly one's, and zero is represented
as an equal number of one's and zero's.  Four bit accuracy, for
example, requires a string of sixteen bits.  This format is practical
only because high resolution is not needed in neural network
applications.  This form of data avoids the conversion circuits which
would otherwise be needed by the multipliers.

      Multiplication is accomplished with Q4 and Q5 by pulse width
modulating the input, a continuous voltage, by the weighting
function.  A continuous voltage represent...