Denoising Autoencoders for Overgeneralization in Neural Networks

摘要: Despite recent developments that allowed neural networks to achieve impressive performance on a variety of applications, these models are intrinsically affected by the problem overgeneralization, due their partitioning full input space into fixed set target classes used during training. Thus it is possible for novel inputs belonging categories unknown training or even completely unrecognizable humans fool system classifying them as one known classes, with high degree confidence. This can lead security problems in critical and closely linked open recognition 1-class recognition. paper presents way compute confidence score using reconstruction error denoising autoencoders shows how correctly identify regions close distribution. The proposed solution tested benchmarks ‘fooling’, constructed from MNIST Fashion-MNIST datasets.