Can training data be safety-critical events? 训练数据能否用安全关键事件？

[Yiru-Jiao]

This article uses normal interactions as training data. Could the crashes and near-crashes in SHRP2 serve as training data?

——This question is asked by Xudong Ren from Southeast University via email.

本文采用的是正常交互数据作为训练数据，是否可以采用SHRP2 的碰撞与接近碰撞数据作为训练数据呢。

——本问题由来自东南大学的任旭东通过邮件提出。

[Yiru-Jiao]

Using crashes and near-crashes to train a risk quantification model could be okay, but it deviates from the core ideas in this article.

• To train a powerful deep learning model requires a large amount of data. However, data on safety-critical events are difficult to collect and extremely valuable. This conflict leads to the scalability challenge in collision risk quantification if a model's training relies on crashes and near-crashes. In other words, we think training on crashes and near-crashes is not sustainable in reality or for practice.
• GSSM extrapolates collision risk from normal interactions because its design is based on the safety pyramid theory. If a model attempts to learn collision risk from crashes and near-crashes, it resembles the Traffic Accident Anticipation (TAA) task as reviewed in the Introduction section of our paper. Such training would face challenges not only in scalability, but also generalisability.

使用碰撞和近碰撞来训练风险量化模型是可以的、但不一定可行，它偏离了本文的核心思想。

• 要训练一个强大的深度学习模型需要大量的数据，然而安全关键事件的数据难以收集且极其宝贵。这种冲突导致如果模型的训练依赖于碰撞和近碰撞，就会在碰撞风险量化中面临规模化的挑战。换句话说，我们认为在碰撞和近碰撞上进行训练在现实或实践中并不可持续。
• GSSM 从正常交互中外推碰撞风险，因为其设计基于安全金字塔理论。如果一个模型试图从碰撞和近碰撞中学习碰撞风险，这类似于我们论文引言部分中回顾的交通事故预警（TAA）任务。这种训练不仅在规模化方面面临挑战，在泛化性方面也会存在问题。

[xudongren-seu]

Thank you for your insightful comments—I fully agree with your perspective.