Training a convolutional neural network

1 . A method for use in training a convolutional neural network using a training data set, wherein each item of training data in the training data set comprises numerical data and a corresponding label for the respective numerical data, comprising:

for each item of training data, converting the numerical data into a matrix, wherein elements in the matrix represent values of features in the numerical data;

determining an arrangement for the numerical data in the matrices that decreases a similarity of matrices comprising numerical data with different labels, and/or increases a similarity of matrices comprising numerical data with the same labels, based on one or more measures of entropy, wherein determining comprises using a machine learning method to: (i) determine a transformation that decreases a measure of entropy between matrices comprising numerical data with the same labels, and/or (ii) determine a transformation that increases a measure of entropy between matrices comprising numerical data with different labels;

arranging the numerical data in each matrix of each item of training data, according to the determined arrangement; and

training the convolutional neural network, based on the arranged matrices and the respective labels, to predict labels for previously unseen numerical data, wherein the previously unseen numerical data is provided to the neural network in the determined arrangement.

2 . The method of claim 1 , wherein the step of converting the numerical data into a matrix comprises representing each feature in the numerical data as a column or a row of elements in the corresponding matrix.

3 . The method of claim 1 , wherein the step of converting the numerical data into a matrix comprises representing the numerical data as a two-dimensional image or a three-dimensional image.

4 . The method of claim 1 , wherein the step of determining an arrangement for the numerical data in the matrices comprises determining an order for the columns or rows of numerical data in the matrices.

5 . The method of claim 1 , wherein the machine learning method comprises at least one of a genetic algorithm method, or an information gain method.

6 . The method of claim 1 , wherein the step of determining an arrangement for the numerical data in the matrices comprises determining a transformation of the form:

X

′

=

X

⁢

∏

k

=

1

n

!

E

k

φk

,

wherein X comprises a matrix, X′ comprises a transformed matrix, E K comprises a column transformation matrix of rank n, and φ k ∈{0,1} comprises a binary parameter.

7 . The method of claim 6 , wherein the parameter φ k ∈{0,1} is determined by optimizing a fitness function, the fitness function comprising a first term for increasing an intra-label similarity of the numerical data represented by the matrices in the training data set and/or a second term for decreasing an inter-label similarity of the numerical data represented by the matrices in the training data set.

8 . The method of claim 6 , wherein the parameter φ k ∈{0,1} is optimized based on a fitness function, G, of the form:

G

⁡

(

Φ

*

)

=

∑

w

=

1

n

-

1

M

⁡

(

x

iwk

′

,

x

i

⁡

(

w

+

1

)

⁢

k

′

)

+

1

2

⁢

∑

c

⁢

1

=

1

C

∑

c

⁢

2

=

1

C

D

⁡

(

X

c

⁢

1

′

,

X

c

⁢

2

′

)

+

∑

k

=

1

n

!

-

φ

k

wherein M comprises a measure of the entropy between matrices having the same label, and D comprises a measure of the entropy of between matrices having different labels.

9 . The method of claim 1 , wherein the numerical data comprises data output by a machine, and wherein the label comprises a ground truth indication of whether the machine is likely to fail within a predetermined time frame.

10 . An apparatus for use in training a convolutional neural network using a training data set, wherein each item of training data in the training data set comprises numerical data and a corresponding label for the respective numerical data, comprising:

a processor configured to communicate with a memory having instruction data representing a set of instructions and to execute the set of instructions, wherein the set of instructions, when executed by the processor, cause the processor to:

for each item of training data, convert the numerical data into a matrix, wherein elements in the matrix represent values of features in the numerical data;

determine an arrangement for the numerical data in the matrices that decreases a similarity of matrices comprising numerical data with different labels, and/or increases a similarity of matrices comprising numerical data with the same labels, based on one or more measures of entropy, wherein determining comprises using a machine learning method to: (i) determine a transformation that decreases a measure of entropy between matrices comprising numerical data with the same labels, and/or (ii) determine a transformation that increases a measure of entropy between matrices comprising numerical data with different labels;

arrange the numerical data in each matrix of each item of training data, according to the determined arrangement; and

train the convolutional neural network, based on the arranged matrices and the respective labels, to predict labels for previously unseen numerical data, wherein the previously unseen numerical data is provided to the neural network in the determined arrangement.

11 . A non-transitory computer program product comprising a computer readable medium, having computer readable code embodied therein, the computer readable code being configured such that, on execution by a processor, the processor is caused to perform the method as claimed in claim 1 .