Training of artificial neural networks using safe mutations based on output gradients

1. A computer implemented method comprising:

determining a threshold value representing a limit on an amount of divergence of response for a neural network;

identifying a set of weights of the neural network, the set of weights corresponding to an initial set of weights; and

training the neural network by:

determining a safe mutation representing a perturbation that results in a response of the neural network that is within the threshold value, wherein the perturbation, without modification, would result in a perturbation that is not within the threshold value, and wherein determining the safe mutation comprises:

feeding the perturbation into a gradient model;

determining an output of the gradient model;

determining a sensitivity of a parameter that is to be perturbed based on the output; and

modifying the perturbation by normalizing the perturbation based on the sensitivity; and

modifying the set of weights of the neural network in accordance with the safe mutation.

2. The computer implemented method of claim 1 , wherein determining the threshold value comprises:

accessing an archive of historical responses of the neural network to historical perturbations;

determining, from the historical responses, a maximum amount that the neural network diverged without causing an error; and

determining the threshold value based on the maximum amount.

3. The computer implemented method of claim 1 , wherein the perturbation, without modification, would result in a perturbation that is not within the threshold value, wherein determining the safe mutation comprises determining a rescaling for the perturbation that, if applied to the perturbation, would result in a response of the neural network that is within the threshold value, and wherein modifying the set of weights of the neural networks in accordance with the safe mutation comprises applying the rescaling to the perturbation.

4. The computer implemented method of claim 3 , wherein determining the rescaling comprises:

decomposing the perturbation into a direction vector comprising a magnitude scalar and a direction; and

determining the magnitude scalar by using a line search to target an amount of divergence that is within the threshold divergence; and

assigning the magnitude as the resealing.

5. The computer implemented method of claim 1 , wherein determining the sensitivity comprises:

generating a sensitivity vector comprising the sensitivities of each weight of the set of weights; and

determining the sensitivity comprises selecting the sensitivity from the sensitivity vector based on the parameter.

6. The computer-implemented method of claim 1 , further comprising:

receiving sensor data from one or more sensors, wherein the set of weights is derived from the sensor data; and

performing an action based on output of the neural network, wherein the output comprises the modified set of weights.

7. The computer-implemented method of claim 1 , wherein generating the set of weights comprises:

transmitting each respective weight of the set of weights to a different respective worker system of a plurality of worker systems, wherein each respective worker system of the plurality of worker systems generates a respective modified weight of the set of weights.

8. The computer-implemented method of claim 1 , wherein training the neural network comprises repeating the determining of the safe mutation, and the modifying of the set of weights.