Strong response consistency for more robust, disentangled and generalizable machine vision

点击翻译按钮获取中文摘要

Deep learning is the workhorse of today's machine intelligence that has produced numerous success stories all over science, industry, and society. However, this rapid and wide-spread success has masked four severe and fundamental issues that limit deep learning to comparatively few, narrowly defined application areas. For one, deep neural networks (DNNs) are not robust and even small input perturbations like JPEG compression or imperceptible but carefully engineered noise can derail their behavior. Second, the representations of DNNs are not disentangled and cannot be easily related to concepts understandable by humans. Third, the representations of DNNs do not generalize well to other tasks or domains without additional supervised training. Finally, we have little theoretical understanding of the representations learned by DNNs, making it difficult to solve the aforementioned issues.To make headway in these areas, I here propose a set of four consistency conditions that the visual representations of DNNs should fulfill. These conditions provide a strong theoretical grounding for unsupervised representation learning that opens up new avenues to study and enhance robustness, disentanglement, and generalization. By targeting these fundamental issues of deep learning, this work will reduce the gap between human and machine vision, increase reliability and explainability of DNNs, accelerate method development, and open up new application areas.