Adding noise to gradients as a regularizer

ballpark.py

from keras.optimizers import SGD, Adagrad, RMSprop, Adadelta, Adam, Adamax
import keras.backend as K

DEFAULT_NOISE = 0.05


def ballpark_gradient(gradient, noise):
    return [g * K.random_normal(shape=K.shape(g), mean=1.0, std=noise) for g in gradient]


class BallparkSGD(SGD):
    def __init__(self, noise=DEFAULT_NOISE, **kwargs):
        self.noise = noise
        super(BallparkSGD, self).__init__(**kwargs)

    def get_gradients(self, loss, params):
        grads = super(BallparkSGD, self).get_gradients(loss, params)
        return ballpark_gradient(grads, self.noise)


class BallparkRMSprop(RMSprop):
    def __init__(self, noise=DEFAULT_NOISE, **kwargs):
        self.noise = noise
        super(BallparkRMSprop, self).__init__(**kwargs)

    def get_gradients(self, loss, params):
        grads = super(BallparkRMSprop, self).get_gradients(loss, params)
        return ballpark_gradient(grads, self.noise)


class BallparkAdagrad(Adagrad):
    def __init__(self, noise=DEFAULT_NOISE, **kwargs):
        self.noise = noise
        super(BallparkAdagrad, self).__init__(**kwargs)

    def get_gradients(self, loss, params):
        grads = super(BallparkAdagrad, self).get_gradients(loss, params)
        return ballpark_gradient(grads, self.noise)


class BallparkAdadelta(Adadelta):
    def __init__(self, noise=DEFAULT_NOISE, **kwargs):
        self.noise = noise
        super(BallparkAdadelta, self).__init__(**kwargs)

    def get_gradients(self, loss, params):
        grads = super(BallparkAdadelta, self).get_gradients(loss, params)
        return ballpark_gradient(grads, self.noise)


class BallparkAdam(Adam):
    def __init__(self, noise=DEFAULT_NOISE, **kwargs):
        self.noise = noise
        super(BallparkAdam, self).__init__(**kwargs)

    def get_gradients(self, loss, params):
        grads = super(BallparkAdam, self).get_gradients(loss, params)
        return ballpark_gradient(grads, self.noise)


class BallparkAdamax(Adamax):
    def __init__(self, noise=DEFAULT_NOISE, **kwargs):
        self.noise = noise
        super(BallparkAdamax, self).__init__(**kwargs)

    def get_gradients(self, loss, params):
        grads = super(BallparkAdamax, self).get_gradients(loss, params)
        return ballpark_gradient(grads, self.noise)

Interesting!
Wondering - any special reason that you are using mean=1.0 ?
why don't you want the noise to be symmetrical (same chance for negative values) ?