US 11,636,667 B2Granted

Pattern recognition apparatus, pattern recognition method, and computer program product

AssigneeKabushiki Kaisha Toshiba

InventorTakashi Masuko

Life of this patent

Estimating the term…

Loading patent…

Abstract

According to an embodiment, a pattern recognition apparatus includes an RNN layer, as a middle layer, that includes an input converting unit and an RNN processor. The input converting unit performs conversion, for each step, on an input vector and a recurrent input vector, and calculates and outputs a converted vector of which the number of dimensions is smaller than the sum of the numbers of dimensions of respective the input vector and the recurrent input vector. The input vector is formed of a feature vector output from an input layer or an output of the RNN processor included in a lower RNN layer. The recurrent input vector is formed of an output of a previous step of the RNN processor. The RNN processor calculates an RNN output vector from the converted vector calculated in the input converting unit and outputs the RNN output vector, for each step.

Claims16

Dependent claims

A pattern recognition apparatus including at least one recurrent neural network (RNN) layer serving as a middle layer forming a recurrent neural network between an input layer and an output layer, the pattern recognition apparatus being configured to output, from the output layer, a pattern series corresponding to a feature vector series input to the input layer, wherein

the RNN layer includes an input converting unit and an RNN processor,
the input layer successively outputs feature vectors forming the feature vector series one at a time for each step,
the input converting unit of the RNN layer performs conversion, for each step, on an input vector and a recurrent input vector, and calculates and outputs a converted vector of which the number of dimensions (I) is smaller than the sum of the number of dimensions (N) of the input vector and the number of dimensions (M) of the recurrent input vector, the input vector being formed of the feature vector output from the input layer or an output of the RNN processor included in the lower RNN layer, the recurrent input vector being formed of an output of a previous step of the RNN processor included in the RNN layer, the number of dimensions (I) of the converted vector satisfying a condition indicated in Expression
I
<
(
N
+
M
)
·
M
N
+
2
⁢
M
,
the RNN processor of the RNN layer calculates an RNN output vector from the converted vector calculated in the input converting unit included in the RNN layer and outputs the RNN output vector, for each step, and
the output layer calculates the pattern series from the RNN output vector calculated in the RNN processor included in an uppermost layer of the RNN layer and outputs the pattern series,
wherein the input converting unit calculates the converted vector vt at a step t, from the input vector xt at the step t and the recurrent input vector mt−1, with
vt=Wvx·xt+Wvm·mt−1
(Wvx is a weight matrix for the input vector xt;
and Wvm is a weight matrix for the recurrent input vector mt−1), or
vt=k(Wvx·xt+Wvm·mt−1+bv)
(k(⋅) is a nonlinear function applied to each vector;
Wvx is a weight matrix for the input vector xt, Wvm is a weight matrix for the recurrent input vector mt−1;
by is a bias vector having I dimensions where the number of dimensions of the converted vector vt is I).

25 more elements

A pattern recognition method executed by a pattern recognition apparatus including at least one recurrent neural network (RNN) layer serving as a middle layer forming a recurrent neural network between an input layer and an output layer, the pattern recognition apparatus being configured to output, from the output layer, a pattern series corresponding to a feature vector series input to the input layer, wherein the RNN layer includes an input converting unit and an RNN processor, the pattern recognition method comprising:

successively outputting, by the input layer, feature vectors forming the feature vector series one at a time for each step,
performing conversion, by the input converting unit of the RNN layer, for each step, on an input vector and a recurrent input vector, and calculating and outputting a converted vector of which the number of dimensions (I) is smaller than the sum of the number of dimensions (N) of the input vector and the number of dimensions (M) of the recurrent input vector, the input vector being formed of the feature vector output from the input layer or an output of the RNN processor included in the lower RNN layer, the recurrent input vector being formed of an output of a previous step of the RNN processor included in the RNN layer, the number of dimensions (I) of the converted vector satisfying a condition indicated in Expression
I
<
(
N
+
M
)
·
M
N
+
2
⁢
M
,
by the RNN processor of the RNN layer, calculating an RNN output vector from the converted vector calculated in the input converting unit included in the RNN layer and outputting the RNN output, for each step, and
by the output layer, calculating the pattern series from the RNN output vector calculated in the RNN processor included in an uppermost layer of the RNN layer and outputting the pattern series,
wherein the input converting unit calculates the converted vector vt at a step t, from the input vector xt at the step t and the recurrent input vector mt−1, with
vt=Wvx·xt+Wvm·mt−1
(Wvx is a weight matrix for the input vector xt;
and Wvm is a weight matrix for the recurrent input vector mt−1), or
vt=k(Wvx·xt+Wcm·mt−1+bv)
(k(⋅) is a nonlinear function applied to each vector;
Wvx is a weight matrix for the input vector xt, Wvm is a weight matrix for the recurrent input vector mt−1;
bv is a bias vector having I dimensions where the number of dimensions of the converted vector vt is I).

24 more elements

A non-transitory computer program product having a computer readable medium including programmed instructions, wherein the instructions, when executed by a computer, cause the computer to function as a pattern recognition apparatus including at least one recurrent neural network (RNN) layer serving as a middle layer forming a recurrent neural network between an input layer and an output layer, the pattern recognition apparatus being configured to output, from the output layer, a pattern series corresponding to a feature vector series input to the input layer, the RNN layer including an input converting unit and an RNN processor, the instruction being configured to cause the computer to execute:

successively outputting, by the input layer, feature vectors forming the feature vector series one at a time for each step,
performing conversion, by the input converting unit of the RNN layer, for each step, on an input vector and a recurrent input vector, and calculating and outputting a converted vector of which the number of dimensions (I) is smaller than the sum of the number of dimensions (N) of the input vector and the number of dimensions (M) of the recurrent input vector, the input vector being formed of the feature vector output from the input layer or an output of the RNN processor included in the lower RNN layer, the recurrent input vector being formed of an output of a previous step of the RNN processor included in the RNN layer, the number of dimensions (I) of the converted vector satisfying a condition indicated in Expression
I
<
(
N
+
M
)
·
M
N
+
2
⁢
M
,
by the RNN processor of the RNN layer, calculating an RNN output vector from the converted vector calculated in the input converting unit included in the RNN layer, and outputting the RNN output vector, for each step, and
by the output layer, calculating the pattern series from the RNN output vector calculated in the RNN processor included in an uppermost layer of the RNN layer, and outputting the pattern series,
wherein the input converting unit calculates the converted vector vt at a step t, from the input vector xt at the step t and the recurrent input vector mt−1, with
vt=Wvx·xt+Wvm·mt−1
(Wvx is a weight matrix for the input vector xt;
and Wvm is a weight matrix for the recurrent input vector mt−1), or
vt=k(Wvx·xt+Wvm·mt−1+bv)
(k(⋅) is a nonlinear function applied to each vector;
Wvx is a weight matrix for the input vector xt, Wvm is a weight matrix for the recurrent input vector mt−1;
bv is a bias vector having I dimensions where the number of dimensions of the converted vector vt is I).

24 more elements

Dependencies are inferred from the claim text.

Facts

Application: 15/677,069Filed 15 Aug 2017
Priority: JP JP2016-22333316 Nov 2016
Published: US 2018/0137410 A117 May 2018
Granted: 25 Apr 2023
Term ends: Not estimatedThe source records don’t support an estimate.

Classification11 codes

G06N 3/044
Architecture, e.g. interconnection topology: recurrent networks, e.g. Hopfield networks
G06F 17/16
Complex mathematical operations: matrix or vector computation , e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
G06F 18/241
Classification techniques: relating to the classification model, e.g. parametric or non-parametric approaches
G06N 3/0442
Recurrent networks, e.g. Hopfield networks: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]