Apparatus for machine learning
First Claim
Patent Images
1. A computer system for machine learning of a time dependent pattern sequence y(t) comprising:
- input means for receiving a plurality, indexed by i, of time dependent inputs xi (t) and a meta-step-size parameter θ
;
calculation means for calculating from said time dependent inputs a predicted value, y*, of said pattern sequence;
a computer memory associated with the said means for calculating;
said calculating means further including a learning rate, ki, exponentially related to an incremental gain β
i (t) and a derivation means for deriving the incremental gain β
i (t) from previous values of β
i (t) and having means forInitializing hi, a per input memory parameter, to 0 and weight coefficients, wi, and β
i, the incremental gain parameter, to chosen values, i=1, . . . , n,Repeating for each new inputs (x1, . . . , xn, y*) the steps of;
calculating, ##EQU5## calculating,
space="preserve" listing-type="equation">δ
=y *-yRepeating for i=1, . . . , n where Ki is an input learning rate and θ
a positive constant denoted the meta-learning rate;
calculating,
space="preserve" listing-type="equation">β
.sub.i =β
.sub.j +β
δ
x.sub.i h.sub.i ##EQU6##
space="preserve" listing-type="equation">w.sub.i (t+1)=w.sub.i (t)+k.sub.i (t)δ
(t)x.sub.i (t)
space="preserve" listing-type="equation">h.sub.i (t+1)= h.sub.i (t)+k.sub.i (t)δ
(t)! 1-k.sub.i (t)x.sub.i (t)!.sup.+.
2 Assignments
0 Petitions
Accused Products
Abstract
An apparatus is disclosed for machine learning of a pattern sequence which is derived from a plurality of inputs. The pattern sequence is predicted from learning rate parameters that are exponentially related to an incrementally calculated gain parameter for each input. The gain parameter are increased or decreased in real time in correlation with the accuracy of the learning process. The disclosed apparatus are advantageously utilized in signal processing, adaptive control systems, and pattern recognition.
-
Citations
1 Claim
-
1. A computer system for machine learning of a time dependent pattern sequence y(t) comprising:
-
input means for receiving a plurality, indexed by i, of time dependent inputs xi (t) and a meta-step-size parameter θ
;calculation means for calculating from said time dependent inputs a predicted value, y*, of said pattern sequence; a computer memory associated with the said means for calculating; said calculating means further including a learning rate, ki, exponentially related to an incremental gain β
i (t) and a derivation means for deriving the incremental gain β
i (t) from previous values of β
i (t) and having means forInitializing hi, a per input memory parameter, to 0 and weight coefficients, wi, and β
i, the incremental gain parameter, to chosen values, i=1, . . . , n,Repeating for each new inputs (x1, . . . , xn, y*) the steps of; calculating, ##EQU5## calculating,
space="preserve" listing-type="equation">δ
=y *-yRepeating for i=1, . . . , n where Ki is an input learning rate and θ
a positive constant denoted the meta-learning rate;calculating,
space="preserve" listing-type="equation">β
.sub.i =β
.sub.j +β
δ
x.sub.i h.sub.i ##EQU6##
space="preserve" listing-type="equation">w.sub.i (t+1)=w.sub.i (t)+k.sub.i (t)δ
(t)x.sub.i (t)
space="preserve" listing-type="equation">h.sub.i (t+1)= h.sub.i (t)+k.sub.i (t)δ
(t)! 1-k.sub.i (t)x.sub.i (t)!.sup.+.
-
Specification