Method for synthesizing a neural network

1. A method for synthesizing a neural network from a plurality of component neural networks, each component neural network comprising a combination of convolutional and/or fully-connected layers between at least one common input and at least one common output, the method comprising the steps of:

mapping each component neural network to a respective component neural network graph with each node of each component neural network graph corresponding to a layer of said component neural network and with each node connected to each other node in accordance with connections of the layers in said component neural network;

providing a first structural label for each node in accordance with the structure of the corresponding layer of the component neural network and a distance of the node from one of a given input or output;

merging the graphs for each component neural network into a single merged graph by merging nodes from component neural network graphs having the same first structural label;

providing a second structural label for each node of the single merged graph in accordance with the structure of the corresponding layer of the or each component neural network and a distance of the node from the other of a given input or output;

providing a contracted-merged graph by merging nodes of the merged graph having the same second structural label; and

mapping the contracted-merged graph to a synthesized neural network with each node of the contracted-merged graph corresponding to a layer of said synthesized network and with each layer of the synthesized network connected to each other layer in accordance with connections of the nodes in the contracted-merged graph.

2. A method according to claim 1 wherein at least one of said component neural networks comprise one of a pooling or an un-pooling layer and wherein said first structural label and second structural labels for convolutional, pooling and un-pooling layers indicate an extent of pooling preceding such layers of a network.

3. A method according to claim 2 wherein said first structural label and second structural labels for fully-connected layers are independent of the extent of pooling preceding a fully-connected layer of a network.

4. A method according to claim 2 wherein a distance property of the first and second structural labels for each node is unaffected by pooling or un-pooling layers between said node and an input or output.

5. A method according to claim 2 wherein at least some of said convolutional layers involve striding and said first and second structural labels for convolutional, pooling and un-pooling layers indicate an extent of striding preceding such layers of a network.

6. A method according to claim 1 further comprising training the synthesized neural network.

7. A method according to claim 1 wherein said at least one common input comprises a P channel image.

8. A computer program product comprising a computer readable medium on which computer executable instructions are stored which when executed on a computing device are arranged to perform the method of claim 1 .

9. A synthesized neural network produced according to the method of claim 1 .