Fair neural networks
A system is disclosed that includes a computer that includes a processor and a memory, the memory including instructions executable by the processor to input an image acquired by a sensor to a neural network to output a prediction regarding an object included in the image. The neural network can be trained based on (a) a distributed robust optimization that minimizes an expectation applied to probability distributions of loss functions to select training images that yield a solution with a selected uncertainty level and (b) generating additional input images based on adversarial images.
1 . A system, comprising:
a computer that includes a processor and a memory, the memory including instructions executable by the processor to:
input an image acquired by a sensor to a neural network to output a prediction regarding an object included in the image; and
wherein the neural network is trained based on (a) a distributed robust optimization that minimizes an expectation applied to probability distributions of loss functions to adversarial training images that yield a solution with a selected uncertainty level and (b) generating additional input images based on the adversarial images; wherein the additional input images are generated based on a structural similarity index measure.
2 . The system of claim 1 , wherein outputting the prediction regarding the image includes outputting an object identity and an object location.
3 . The system of claim 1 , wherein the neural network outputs a confidence value that indicates a probability that the prediction regarding the image is correct.
4 . The system of claim 1 , wherein the loss function is determined based on comparing output from the neural network with ground truth data based on the input image.
5 . The system of claim 1 , wherein the neural network includes convolutional layers and fully connected layers.
6 . The system of claim 1 , wherein the input images are respectively input to the neural network a plurality of times and the loss functions are backpropagated through layers of the neural network to select weights that minimize the loss functions.
7 . The system of claim 1 , wherein generating the additional input images based on the adversarial images includes determining imperceptible differences based on the structural similarity index measure.
8 . The system of claim 7 , wherein the structural similarity index measure generates images that are not perceptibly different to a human observer but can cause the neural network to fail.
9 . The system of claim 1 , wherein generating the additional input images based on the adversarial images includes determining imperceptible differences based on processing the input image with a PieAPP neural network.
10 . The system of claim 9 , wherein the PieAPP neural network generates images that are not perceptibly different to a human observer but can cause the neural network to fail.
11 . A method, comprising:
inputting an image acquired by a sensor to a neural network to output a prediction regarding an object included in the image; and
wherein the neural network is trained based on (a) a distributed robust optimization that minimizes an expectation applied to probability distributions of loss functions to adversarial training images that yield a solution with a selected uncertainty level and (b) generating additional input images based on the adversarial images; wherein the additional input images are generated based on a structural similarity index measure.
12 . The method of claim 11 , wherein outputting the prediction regarding the image includes outputting an object identity and an object location.
13 . The method of claim 11 , wherein the neural network outputs a confidence value that indicates a probability that the prediction regarding the image is correct.
14 . The method of claim 11 , wherein the loss function is determined based on comparing output from the neural network with ground truth data based on the input image.
15 . The method of claim 11 , wherein the neural network includes convolutional layers and fully connected layers.
16 . The method of claim 11 , wherein the input images are respectively input to the neural network a plurality of times and the loss functions are backpropagated through layers of the neural network to select weights that minimize the loss functions.
17 . The method of claim 11 , wherein generating the additional input images based on the adversarial images includes determining imperceptible differences based on the structural similarity index measure.
18 . The method of claim 17 , wherein the structural similarity index measure generates images that are not perceptibly different to a human observer but can cause the neural network to fail.
19 . The method of claim 11 , wherein generating the additional input images based on the adversarial images includes determining imperceptible differences based on processing the input image with a PieAPP neural network.
20 . The method of claim 19 , wherein the PieAPP neural network generates images that are not perceptibly different to a human observer but can cause the neural network to fail.