Resource-aware training for neural networks
In implementations of resource-aware training for neural network, one or more computing devices of a system implement an architecture optimization module for monitoring parameter utilization while training a neural network. Dead neurons of the neural network are identified as having activation scales less than a threshold. Neurons with activation scales greater than or equal to the threshold are identified as survived neurons. The dead neurons are converted to reborn neurons by adding the dead neurons to layers of the neural network having the survived neurons. The reborn neurons are prevented from connecting to the survived neurons for training the reborn neurons.
1. In a digital medium environment to optimize architecture of a neural network, a method implemented by a computing device, the method comprising:
identifying survived neurons of the neural network, the survived neurons having an activation scale greater than or equal to a threshold;
identifying dead neurons of the neural network, the dead neurons having an activation scale less than the threshold;
adding a particular dead neuron to a layer of the neural network having at least some of the survived neurons; and
training the survived neurons and the particular dead neuron separately by training the survived neurons on a set of training examples and training the particular dead neuron on adversarial training examples generated from the set of training examples.
2. The method as described in claim 1 , further comprising adding an additional dead neuron to an additional layer of the neural network having at least one of the identified survived neurons.
3. The method as described in claim 2 , wherein the particular dead neuron and the additional dead neuron are connected.
4. The method as described in claim 3 , further comprising:
preventing the particular dead neuron from connecting to the at least one identified survived neuron; and
preventing the additional dead neuron from connecting to the at least some identified survived neurons.
5. The method as described in claim 1 , wherein the threshold is determined as a percentage of a maximum neural activation scale.
6. The method as described in claim 1 , wherein the adding comprises expanding a width of the layer of the neural network by a maximum expansion rate, the maximum expansion rate defined by a constraint on at least one computing resource of the computing device.
7. The method as described in claim 6 , wherein the constraint is a defined resource limit.
8. In a digital medium environment to optimize architecture of a neural network, modules implemented at least partially in hardware of one or more computing devices of a system comprising:
a utilization monitoring module implemented to:
identify survived neurons of the neural network, the survived neurons having an activation scale greater than or equal to a threshold; and
identify dead neurons of the neural network, the dead neurons having an activation scale less than the threshold;
a network morphing module implemented to add a particular dead neuron to a layer of the neural network having at least some of the survived neurons; and
a life extension module implemented to train the survived neurons and the particular dead neuron separately by training the survived neurons on a set of training examples and training the particular dead neuron on adversarial training examples generated from the set of training examples.
9. The system as described in claim 8 , wherein the network morphing module is implemented to add an additional dead neuron to an additional layer of the neural network having at least one of the identified survived neurons.
10. The system as described in claim 9 , wherein the particular dead neuron is connected to the additional neuron.
11. The system as described in claim 10 , wherein the life extension module is implemented to prevent the particular dead neuron from connecting to the at least one identified survived neuron and prevent the additional dead neuron from connecting to the at least some identified survived neurons.
12. The system as described in claim 8 , wherein the threshold is determined as a percentage of a maximum neural activation scale.
13. The system as described in claim 8 , wherein the network morphing module is implemented to expand a width of the layer of the neural network by a maximum expansion rate, the maximum expansion rate defined by a constraint on at least one computing resource of the one or more computing devices.
14. The system as described in claim 13 , wherein the constraint is a defined resource limit.
15. In a non-transitory computer-readable medium to optimize architecture of a neural network, a computer implemented method comprising:
a step for identifying survived neurons of the neural network;
a step for identifying dead neurons of the neural network;
adding a particular dead neuron to a layer of the neural network having at least some of the identified survived neurons;
adding an additional dead neuron to an additional layer of the neural network having at least one of the survived neurons; and
training the survived neurons and the particular dead neuron separately by training the survived neurons on a set of training examples and training the particular dead neuron on adversarial training examples generated from the set of training examples.
16. The method as described in claim 15 , wherein the particular dead neuron and the additional dead neuron are connected.
17. The method as described in claim 15 , further comprising a step for determining a maximum neural activation scale.
18. The method as described in claim 15 , further comprising a step for expanding a width of the layer of the neural network by a maximum expansion rate, the maximum expansion rate defined by a constraint on at least one computing resource of a computing device.
19. The method as described in claim 18 , wherein the expanding is linear.
20. The method as described in claim 18 , wherein the constraint is a defined resource limit.