Neural network construction

1 . A method of image processing comprising:

generating a population of software hox genomes, each software hox genome encoding a neural network and comprising, for each zone of a plurality of zones of the neural network, a hox gene and at least one norm gene, where the norm gene encodes the zone and the hox gene controls at least one of: a quantity of instances of the zone in the neural network, a position of the zone in the neural network, a gradient of weights in the zone, a depth of nested zones within the zone, an amount of bias to be applied to weights in the zone;

applying an evolutionary algorithm to the population of software hox genomes where a fitness function of the evolutionary algorithm comprises a performance of the neural network encoded by a software hox genome on an image processing task, where the neural network has been trained for the image processing task using training images;

outputting from the evolutionary algorithm, one of the software hox genomes and constructing the neural network encoded by the software hox genome; and

processing an image using the constructed neural network in order to perform an image processing task.

2 . The method of claim 1 , wherein the fitness function of the evolutionary algorithm also rates smaller neural networks over larger neural networks so as to encourage the smaller neural networks to be evolved.

3 . The method of claim 1 , wherein applying the evolutionary algorithm comprises:

computing a fitness of each of the software hox genomes using the fitness function, by, for each software hox genome, by:

configuring the neural network according to the software hox genome;

training the neural network using the training images;

measuring performance of the trained neural network at the image processing task and computing the fitness of the software hox genome according to the measured performance; and

wherein the evolutionary algorithm further comprises:

selecting a plurality of the software hox genomes as parents which have a fitness above a threshold;

generating offspring software hox genomes from the parents; and

replacing least-fit software hox genomes in the population with the offspring software hox genomes.

4 . The method of claim 1 , wherein said each zone comprises a plurality of nodes and a plurality of edges of the neural network.

5 . The method of claim 1 , farther comprising: rendering a representation of the constructed neural network on a display, the representation comprising marks indicating the zones and topology of the zones in the neural network.

6 . The method of claim 1 , further comprising: allocating individual ones of the plurality of zones to different components of the image processing task.

7 . The method of claim 3 , wherein configuring the neural network according to the software hox genome comprises reusing one or more neural network zones.

8 . The method of claim 1 , wherein the software hox genome is a sequence of characters, and wherein each hox gene begins with a unique character combination and terminates with a specified character.

9 . The method of claim 8 , wherein the sequence of characters comprises one or more norm genes, each norm gene comprising characters encoding a zone of the neural network by encoding a number of nodes in the zone, a topology of the zone, an activation function of the nodes in the zone.

10 . The method of claim 9 , wherein each of the norm genes encodes the zone of the neural network as one or more of: a fully connected layer, a convolutional layer, a recurrent layer, an encoder neural network, a decoder neural network, an autoencoder neural network.

11 . The method of claim 1 , wherein constructing the neural network comprises pruning an existing neural network.

12 . A method of detecting security breach events in a telecommunications network comprising:

generating a population of software hox genomes, each software hox genome encoding a neural network and comprising, for each zone of a plurality of zones of the neural network, a hox gene and at least one norm gene, where the norm gene encodes the zone and the hox gene controls at least one of: a quantity of instances of the zone in the neural network, a position of the zone in the neural network, a gradient of weights in the zone, a depth of nested zones within the zone, an amount of bias to be applied to weights in the zone;

applying an evolutionary algorithm to the population of software hox genomes where a fitness function of the evolutionary algorithm is a performance of the neural network encoded by a software hox genome on detection of the security breach events in the telecommunications network, where the neural network has been trained using labeled event data from the telecommunications network;

outputting from the evolutionary algorithm, one of the software hox genomes and constructing the neural network encoded by the software hox genome; and

processing measurements from the telecommunications network using the constructed neural network in order to detect the security breach events.

13 . An image processing apparatus comprising:

one or more processors; and

computer storage memory having computer-executable instructions stored thereon which, when executed by the one or more processors, implement a method comprising:

generating a population of software hox genomes, each software hox genome encoding a neural network and comprising, for each zone of a plurality of zones of the neural network, a hox gene and at least one norm gene, where the norm gene encodes the zone and the hox gene controls at least one of: a quantity of instances of the zone in the neural network, a position of the zone in the neural network, a gradient of weights in the zone, a depth of nested zones within the zone, an amount of bias to be applied to weights in the zone;

applying an evolutionary algorithm to the population of software hox genomes where a fitness function of the evolutionary algorithm comprises a performance of the neural network encoded by a software hox genome on an image processing task, where the neural network has been trained for the image processing task using training images;

outputting from the evolutionary algorithm, one of the software hox genomes and constructing the neural network encoded by the software hox genome; and

processing an image using the constructed neural network in order to perform an image processing task.

14 . The image processing apparatus of claim 13 , wherein the fitness function of the evolutionary algorithm also rates smaller neural networks over larger neural networks so as to encourage the smaller neural networks to be evolved.

15 . The image processing apparatus of claim 13 , wherein applying the evolutionary algorithm comprises:

computing a fitness of each of the software hox genomes using the fitness function, for each software hox genome, by:

configuring the neural network according to the software hox genome;

training the neural network using the training images;

measuring performance of the trained neural network at the image processing task and computing the fitness of the software hox genome according to the measured performance; and

wherein the evolutionary algorithm further comprises:

selecting a plurality of the software hox genomes as parents which have a fitness above a threshold;

generating offspring software hox genomes from the parents; and

replacing least-fit software hox genomes in the population with the offspring software hox genomes.

16 . The image processing apparatus of claim 13 , wherein:

the software hox genome is a sequence of characters, and wherein each hox gene begins with a unique character combination and terminates with a specified character; and

the sequence of characters comprises one or more norm genes, each norm gene comprising characters encoding a zone of the neural network by encoding a number of nodes in the zone, a topology of the zone, an activation function of the nodes in the zone.

17 . The image processing apparatus of claim 16 , wherein each of the norm gene encodes the zone of the neural network as one or more of: a fully connected layer, a convolutional layer, a recurrent layer, an encoder neural network, a decoder neural network, an autoencoder neural network.

18 . The image processing apparatus of claim 13 , wherein constructing the neural network comprises pruning an existing neural network.

19 . An apparatus in a telecommunications network for detecting security breach events, the apparatus comprising:

one or more processors; and

computer storage memory having computer-executable instructions stored thereon which, when executed by the one or more processors, implement a method comprising:

generating a population of software hox genomes, each software hox genome encoding a neural network and comprising, for each zone of a plurality of zones of the neural network, a hox gene and at least one norm gene, where the norm gene encodes the zone and the hox gene controls at least one of: a quantity of instances of the zone in the neural network, a position of the zone in the neural network, a gradient of weights in the zone, a depth of nested zones within the zone, an amount of bias to be applied to weights in the zone;

applying an evolutionary algorithm to the population of software hox genomes where a fitness function of the evolutionary algorithm comprises a performance of the neural network encoded by a software hox genome on detection of the security breach events in the telecommunications network, where the neural network has been trained using labeled event data measured from the telecommunications network;

outputting from the evolutionary algorithm, one of the software hox genomes and constructing the neural network encoded by the software hox genome; and

processing measurements from the telecommunications network using the constructed neural network in order to detect the security breach events.

20 . A telecommunications network for detecting security breach events, the telecommunications network comprising an apparatus, wherein the apparatus comprises:

one or more processors; and

computer storage memory having computer-executable instructions stored thereon which, when executed by the one or more processors, implement a method comprising:

generating a population of software hox genomes, each software hox genome encoding a neural network and comprising, for each zone of a plurality of zones of the neural network, a box gene and at least one norm gene, where the norm gene encodes the zone and the hox gene controls at least one of: a quantity of instances of the zone in the neural network, a position of the zone in the neural network, a gradient of weights in the zone, a depth of nested zones within the zone, an amount of bias to be applied to weights in the zone:

applying an evolutionary algorithm to the population of software hox genomes where a fitness function of the evolutionary algorithm comprises a performance of the neural network encoded by a software hox genome on detection of the security breach events in the telecommunications network, where the neural network has been trained using labeled event data measured from the telecommunications network;

outputting from the evolutionary algorithm, one of the software hox genomes and constructing the neural network encoded by the software hox genome; and

processing measurements from the telecommunications network using the constructed neural network in order to detect the security breach events.