Neuro-Symbolic AI Systems for Safe Autonomous Driving

该项目属于EPSRC人工智能和机器人研究领域，旨在提高自动驾驶系统的态势感知，最终导致更安全和值得信赖的模型。自动驾驶系统严重依赖神经网络来处理视觉数据，虽然这些都是强大的工具，但它们仍然是黑匣子，经常做出不可预测的决定，这在实践中很容易导致灾难性的结果。在这个研究项目中，我将重点讨论神经网络的两个警告，数据贪婪和缺乏推理能力，并讨论通过将领域知识整合到神经网络中来使用逻辑来解决这些限制的方法。理想情况下，我们希望我们的神经网络系统能够保证满足获取领域知识的安全要求，同时使用较少的注释数据。确保神经网络满足要求是人工智能中一个长期存在的问题，但直到最近才开始重新引起研究界的兴趣。先前关于神经符号整合的工作表明，通过逻辑约束为神经网络配备推理能力使其能够指导学习，从而使模型符合约束条件，而且还可以有效地从较小的注释数据集中学习。然而，现有的神经符号方法是为小型合成数据集设计的，不能扩展到更复杂的现实世界场景，例如自动驾驶的对象检测。为了在自动驾驶的背景下解决上述两个挑战，本项目提出了一种在能够扩展到目标检测场景的损失函数中嵌入逻辑约束的方法，以及使用逻辑约束来解决数据贪婪问题的方法。这些提出的方法将把逻辑约束集成到神经网络中，以指导学习过程，并纠正神经网络的预测。此外，关于自动驾驶车辆和其他交通参与者的安全的许多约束可以表示为线性不等。因此，该项目还致力于开发稳定地构建约束的可表现性的方法，以便支持更复杂的需求。

This project falls within the "EPSRC Artificial Intelligence and Robotics" research area and aims at improving the situational awareness of autonomous driving systems, ultimately leading to safer and trustworthy models. Autonomous driving systems heavily rely on neural networks to process visual data and, while these are powerful tools, they are still black boxes, often taking unpredictable decisions, which could easily lead to disastrous outcomes in practice. In this research project, I will focus on two caveats of neural networks, data greediness and lack of reasoning capabilities, and discuss approaches to address these limitations by using logic as means of incorporating domain knowledge into the neural networks. Ideally, we would like our neural network systems to guarantee the satisfaction of safety requirements capturing domain knowledge and, at the same time, use less annotated data. Ensuring that neural networks satisfy requirements is a long-standing problem in AI, but only recently it started to regain interest from the research community. Prior work on neuro-symbolic integration showed that equipping neural networks with reasoning capabilities via logical constraints allows them to guide the learning, so that the models are compliant with the constraints, but also to efficiently learn from smaller annotated datasets. However, the existing neuro-symbolic methods were designed for small, synthetic datasets and would not scale to more complex, real-world scenarios such as object detection for autonomous driving. To address the above two challenges in the context of autonomous driving, this project proposes an approach to embed logical constraints into the loss function that is able to scale to object detection scenarios, and a way of addressing the data greediness problem using logical constraints. These proposed approaches will integrate logical constraints into neural networks to guide the learning process, but also to correct the neural network's predictions. Additionally, numerous constraints concerning the safety of autonomous vehicles and of other traffic participants can be expressed as linear inequalities. Therefore, the project also aims at developing methods that steadily build up the expressivity of the constraints in order to support more complex requirements.