Claims
- 1. A neural net comprising:
- a signal net and a constraint net; wherein
- said signal net comprises:
- a plurality N.sub.s of signal net legs, said signal net legs being numbered, respectively, from i=1 to N.sub.s ;
- each of said signal net legs comprises an exponential amplifier of transfer function g;
- the output of the ith of said signal net legs is O.sub.i ; and
- wherein said constraint net comprises:
- a plurality N.sub.c of constraint net legs, said constraint net legs being numbered, respectively, from j=1 to N.sub.c ;
- each of said constraint net legs comprising an amplifier of gain K;
- the input of each of said signal net legs has a shunt impedance having a real component R,
- wherein said neural net comprises feedback means for causing the output of the jth of said constraint net legs to be fed to the input of the ith of said signal net legs via a series transconductance -T.sub.ij, and for causing O.sub.i to be fed to the input of the jth of said constraint net legs via series transconductance T.sub.ij ; and
- wherein said neural net further comprises means for causing an input to the ith of said signal net legs to be (1/R)log(M.sub.i), M.sub.i being the ith element of a preselected data set.
- 2. A neural net circuit of the Tank-Hopfield kind, wherein said circuit comprises means for causing the stability function E of said circuit to be:
- E=(K/2).SIGMA..sub.j (.SIGMA..sub.i O.sub.i T.sub.ij -I.sub.j).sup.2 +(1/R).SIGMA..sub.i [O.sub.i log(O.sub.i /M.sub.i)-O.sub.i ]
- where M.sub.i is the ith element of a preselected data set of N.sub.s members, i=1 to N.sub.s, O.sub.i is the output of the ith leg of the signal net of said circuit, I.sub.j is the input to the jth leg of the constraint net of said circuit, K is the gain of each said leg of said constraint net, T.sub.ij is the interconnect strength between the ith leg of said signal net and the jth leg of said constraint net, and R is the real part of the input shunt impedance of each of said signal net legs.
- 3. A circuit of the Tank-Hopfield kind, wherein:
- the signal net of said circuit comprises a plurality N.sub.s of circuit legs, each of said circuit legs having an exponential transfer function g, and
- wherein said circuit comprises means for causing the input of the ith of said circuit legs to be (1/R)log(M.sub.i ), where i=1 to N.sub.s, M.sub.i is the ith element of a preselected data set M having N.sub.s elements, each M.sub.i is selected to be a prior estimate of a signal O.sub.i, and R is the real portion of the input shunt impedance for each of said legs.
- 4. A method of deconvolving a data set I.sub.j having noise corruption, j=1 to N using circuit of the Tank-Hopfield kind, wherein:
- the signal net of said circuit comprises a plurality N.sub.s of circuit legs, called signal legs,
- the constraint net of said circuit comprises a plurality N.sub.c of circuit legs called constraint legs
- wherein said method comprises steps for:
- causing the transfer function g of each of said signal legs to be exponential; and
- imputting to the ith of said circuit legs a signal (1/R)log(M.sub.i), where i=1 to N.sub.s, M.sub.i is the ith element of a preselected data set M having N.sub.s elements, and R is the real portion of the input shunt impedance for each of said legs.
- 5. The method of claim 4, further comprising using the outputs of said constraint legs to estimate said noise corruption.
Parent Case Info
This is a continuation in part of application Ser. No. 07/374,851, now abandoned.
US Referenced Citations (7)
Non-Patent Literature Citations (2)
Entry |
Michael Peter Kennedy et al., "Circuit Theoretic Solutions for Neural Netks", Jun. 21-24, 1987. |
Electronic "Neural" Net Algorithm for Maximum Entropy Solutions of Ill-Posed Problems, by Christie R. K. Marrian and Martin C. Peckerar, IEEE Transactions on Circuits and Systems, vol. 36, No. 2, Feb. 1989. |
Continuation in Parts (1)
|
Number |
Date |
Country |
Parent |
374815 |
Jul 1989 |
|