Enforcing low confidence class predictions for out of distribution data in deep convolutional networks

Sammanfattning: Modern discriminative deep neural networks are known to perform high confident predictions for inputs far away from the training data distribution, commonly referred to as out-of-distribution inputs. This property poses security concerns for the deployment of deep learning models in critical applications like autonomous vehicles because it hinders the detection of such inputs. The aim of this thesis is to investigate the problem of out-of-distribution inputs detection and to propose a solution based on a novel method to enforce low confidence far away from the training data in a supervised setting. To do so, samples lying on the classifier decision boundaries are generated by backpropagating the gradient of an appropriately designed loss function to the input and are used as out-of-distribution examples. At the end of the proposed iterative training procedure, the network high confidence region overlaps with the training data distribution support, resulting in low confidence everywhere else and in an improved detection capability of out of distribution inputs. We first evaluate the method on a synthetic 2-dimensional dataset to have a hindsight of its functioning by visualizing the model confidence. To verify its ability to scale up to higher dimensionality settings, we then apply it to more complex datasets. The experimental results on the MNIST dataset show comparable performance to the current state-of-the-art approach. When tested on the CIFAR-10 dataset, the proposed method is not able to reach entirely satisfactory results. Some of the considered metrics however suggest that through further experimentation we might improve the method capabilities.