Early classification method and electrical device for multi-objective optimization
An early classification method with multiple-objectives optimization is provided. The method includes: dividing time series data into multiple snippets; input a snippet into a first machine learning model to obtain a spatial feature vector; calculating, by a recurrent neural network, a current spatiotemporal feature vector according to the spatial feature vector and a previous spatiotemporal feature vector; determining whether to perform early classification according to the spatiotemporal feature vector; if determined not to perform the early classification, processing a subsequent snippet; if determined to perform the early classification, input the spatiotemporal feature vector into a second machine learning model to calculate a predicted label of the time series data. Accordingly, multiple-objectives optimization is achieved.
1 . An early classification method performed by an electrical device, wherein the early classification method comprises:
obtaining time series data, and dividing the time series data into a plurality of snippets;
inputting one of the snippets into a first machine learning model to obtain a first spatial feature vector;
calculating, by a recurrent neural network, a first spatiotemporal feature vector according to the first spatial feature vector and a previous spatiotemporal feature vector;
determining whether to perform an early classification according to the first spatiotemporal feature vector, comprising:
calculating a stopping probability according to the first spatiotemporal feature vector;
if the one of the snippets is a first one, substituting the stopping probability into a function and determining if an output of the function is greater than a first threshold; and
if the output of the function is greater than the first threshold, determining to perform the early classification, otherwise determining not to perform the early classification;
if determining not to perform the early classification, obtaining a subsequent snippet of the snippets, inputting the subsequent snippet into the first machine learning model to obtain a second spatial feature vector, and calculating, by the recurrent neural network, a second spatiotemporal feature vector according to the second spatial feature vector and the first spatiotemporal feature vector; and
if determining to perform the early classification, inputting the first spatiotemporal feature vector into a second machine learning model to calculate a predicted label of the time series data.
2 . The early classification method of claim 1 , wherein the time series data is an electrocardiogram signal, the time series data comprises a plurality of sampling points, each of the sampling points comprises a plurality of variables, each of the variables corresponds to a sensing electrode, and each of the snippets corresponds to a heartbeat.
3 . The early classification method of claim 1 , wherein the first machine learning model is a convolution neural network, the recurrent neural network is a long short-term memory (LSTM) network, and the second machine learning model is a fully connected layer.
4 . The early classification method of claim 1 , wherein the step of determining whether to perform the early classification according to the first spatiotemporal feature vector further comprises:
if the one of the snippets is not the first one, normalizing the stopping probability to obtain a normalized probability;
calculating a maximum value among at least one preceding normalized probability; and
determining if the normalized probability plus a difference between the normalized probability and the maximum value is greater than or equal to a second threshold, and determining to perform the early classification if yes, otherwise determining not to perform the early classification.
5 . The early classification method of claim 4 , further comprising:
setting a reward according to a following equation:
R
=
{
max
(
1
,
(
log
(
T
-
log
(
τ
)
)
)
,
if
correct
classification
-
max
(
1
,
(
log
(
T
-
log
(
τ
)
)
)
,
otherwise
wherein R is the reward, τ is a time point of determining whether to perform the early classification, T is a number of the snippets.
6 . The early classification method of claim 5 , further comprising:
setting a score objective function in a training phase based a following equation:
Score
reward
=
1
N
∑
i
=
1
N
R
(
i
)
wherein N is a number of a plurality of training samples, R (i) is the reward of an i-th training sample of the training samples.
7 . The early classification method of claim 6 , further comprising:
setting an accuracy objective in the training phase based on a following equation:
Accuracy
=
1
N
∑
i
=
1
N
(
y
(
i
)
==
y
^
(
i
)
)
wherein y (i) is a ground truth of the i-th training sample, ŷ (i) is the predicted label of the i-th training sample.
8 . The early classification method of claim 7 , further comprising:
setting an earliness objective in the training phase based on a following equation:
Earliness
snippet
=
1
N
∑
i
=
1
N
τ
(
i
)
T
(
i
)
wherein T (i) is a number of the snippets of the i-th training sample.
9 . An electrical device comprising:
a memory storing a plurality of instructions; and
a processor communicatively connected to the memory and configured to execute the instructions to perform a plurality of steps:
obtaining time series data, and dividing the time series data into a plurality of snippets;
inputting one of the snippets into a first machine learning model to obtain a first spatial feature vector;
calculating, by a recurrent neural network, a first spatiotemporal feature vector according to the first spatial feature vector and a previous spatiotemporal feature vector;
determining whether to perform an early classification according to the first spatiotemporal feature vector, comprising:
calculating a stopping probability according to the first spatiotemporal feature vector;
if the one of the snippets is a first one, substituting the stopping probability into a function and determining if an output of the function is greater than a first threshold; and
if the output of the function is greater than the first threshold, determining to perform the early classification, otherwise determining not to perform the early classification;
if determining not to perform the early classification, obtaining a subsequent snippet of the plurality of snippets, inputting the subsequent snippet into the first machine learning model to obtain a second spatial feature vector, and calculating, by the recurrent neural network, a second spatiotemporal feature vector according to the second spatial feature vector and the first spatiotemporal feature vector; and
if determining to perform the early classification, inputting the first spatiotemporal feature vector into a second machine learning model to calculate a predicted label of the time series data.
10 . The electrical device of claim 9 , wherein the time series data is an electrocardiogram signal, the time series data comprises a plurality of sampling points, each of the sampling points comprises a plurality of variables, each of the variables corresponds to a sensing electrode, and each of the snippets corresponds to a heartbeat.
11 . The electrical device of claim 9 , wherein the first machine learning model is a convolution neural network, the recurrent neural network is a long short-term memory (LSTM) network, and the second machine learning model is a fully connected layer.
12 . The electrical device of claim 9 , wherein the step of determining whether to perform the early classification according to the first spatiotemporal feature vector further comprises:
if the one of the snippets is not the first one, normalizing the stopping probability to obtain a normalized probability;
calculating a maximum value among at least one preceding normalized probability; and
determining if the normalized probability plus a difference between the normalized probability and the maximum value is greater than or equal to a second threshold, and determining to perform the early classification if yes, otherwise determining not to perform the early classification.
13 . The electrical device of claim 12 , wherein the plurality of steps further comprise:
setting a reward according to a following equation:
R
=
{
max
(
1
,
(
log
(
T
-
log
(
τ
)
)
)
,
if
correct
classification
-
max
(
1
,
(
log
(
T
-
log
(
τ
)
)
)
,
otherwise
wherein R is the reward, τ is a time point of determining whether to perform the early classification, T is a number of the snippets.
14 . The electrical device of claim 13 , wherein the plurality of steps further comprise:
setting a score objective function in a training phase based on a following equation:
Score
reward
=
1
N
∑
i
=
1
N
R
(
i
)
wherein N is a number of a plurality of t training samples, R (i) is the reward of an i-th training sample of the training samples.
15 . The electrical device of claim 14 , wherein the plurality of steps further comprise:
setting an accuracy objective in the training phase based on a following equation:
Accuracy
=
1
N
∑
i
=
1
N
(
y
(
i
)
==
y
^
(
i
)
)
wherein y (i) is a ground truth of the i-th training sample, ŷ (i) is the predicted label of the i-th training sample.
16 . The electrical device of claim 15 , wherein the plurality of steps further comprise:
setting an earliness objective in the training phase based on a following equation:
Earliness
snippet
=
1
N
∑
i
=
1
N
τ
(
i
)
T
(
i
)
wherein T (i) is a number of the snippets of the i-th training sample.