Probabilistic Inference and Deep Learning

虽然神经网络在图像和语言理解、机器人技术和医学方面取得了巨大的性能进步，但这些算法有时仍然存在校准错误、不透明和难以配置的问题。该研究旨在通过借鉴贝叶斯统计和信息几何的技术来提高神经网络的鲁棒性、可解释性和可配置性。结果将是网络对敌对输入具有鲁棒性，并且知道他们不知道的东西。这项研究也将导致自动设计神经网络架构的技术。

While neural networks have led to dramatic performance gains in image and language understanding, robotics, and medicine, the algorithms are still sometimes miscalibrated, opaque, and difficult to configure. The proposed research aims to improve the robustness, interpretability, and configurability of neural networks by borrowing techniques from Bayesian statistics and information geometry. The result will be networks which are robust to adversarial inputs and which know what they dont know. This research will also lead to techniques for automatically designing neural network architectures.