Source-linked AI summary
"Curse of rarity" for autonomous vehicles
Henry X. Liu, Shuo Feng
TL;DR
Autonomous-vehicle safety is challenged by rare safety-critical events in high-dimensional driving environments, a problem the paper names the curse of rarity (CoR). It systematically analyzes how CoR affects AV development and discusses potential approaches, concluding that CoR underlies challenges across perception, prediction, decision making, and verification and validation. The paper also identifies unresolved difficulties in defining rare events and combining proposed solutions.
Problem
Rare safety-critical events in high-dimensional driving environments make safety-critical systems difficult to learn and evaluate, motivating the formulation of the curse of rarity.
Method
The paper systematically analyzes CoR across AV development stages and discusses potential approaches for addressing it.
Results
CoR underlies critical AV challenges in perception, prediction, decision making, and verification and validation, and hinders the application of deep learning methods.
Takeaways & Limitations
The paper provides a framework for understanding AV safety challenges and directions for further investigation into CoR solutions.
Takeaways & Limitations
Defining, identifying, and combining rare-event approaches remains challenging, particularly because AV tasks have spatiotemporal complexity and differing objectives.
Abstract
from arXiv · showhide
In this paper, we reveal that the rarity of safety-critical events in high-dimensional driving environments is the root cause of the safety challenge for autonomous vehicle development. We formulate it as "curse of rarity" (CoR) because it occurs ubiquitously in various safety-critical systems such as medical diagnosis and aerospace systems.
What is the “curse of rarity”?
The curse of rarity (CoR) arises when safety-critical events are both rare and embedded in high-dimensional environments, leaving most available data weakly informative. This produces high-variance learning and validation problems for autonomous vehicles, while identifying useful rare events remains difficult.
- Curse of dimensionality: As dimensionality increases, the available data become sparse and the data needed for reliable results often grows exponentially.This is the curse of dimensionality, which compounds rarity-related difficulties in AV development.
- Curse of rarity: CoR combines rare safety-critical events with high-dimensional driving environments, making sufficient informative data difficult to obtain.The paper distinguishes this challenge from the curse of dimensionality and frames it as a fundamental safety problem for AVs.
- Learning under rarity: Only a small portion of AV training data may contribute effectively to gradient estimation, while most normal scenarios contribute little.Near-miss events and crashes are rare but contain valuable information for safety training.
- Learning under rarity: Rare-event sampling can produce large gradient-estimation variance and mislead learning when critical events carry the informative signal.The formal partition separates normal events A from rare critical events B, with the normal-event contribution assumed to be zero.
- Identifying critical events: Defining useful rare events is challenging because AV objectives may require both crashes and successfully avoided near-misses, while only safety-critical trajectory portions contain relevant information.The event definition must reflect the gradient information and may need to account for temporal relations.
- AV safety implications: At higher safety targets, rarity becomes more severe: collecting a fatality event at human-driver rates would take 10^8 miles on average.The paper links decreasing exposure frequency with sharply increasing development cost and notes that CoR affects perception, prediction, planning, and validation.
What are the potential solutions toward solving the “curse of rarity”?
The paper identifies three potential approaches to the curse of rarity: train on rare events, infer rare-event information from normal events, or reduce the rare-event set. Each approach faces significant challenges, and combinations may be promising.
- The paper identifies three potential approaches for solving the curse of rarity and guiding future development.These approaches are illustrated in Figure 1.
- Approach #1: Training with B: Training with B uses rare-event data to reduce estimation variance, but requires task-specific methods to define, identify, and use those events.The challenge is heightened by the spatiotemporal complexity of autonomous driving and differing learning objectives across tasks.
- Approach #2: Reasoning from A: Reasoning from A seeks to infer information about rare events from normal events, potentially reducing reliance on large-scale data.The paper associates this approach with combining bottom-up, data-driven reasoning and top-down cognitive reasoning.
- Approach #3: Reducing the size of B: Reducing the size of B aims to make safety-critical events occur at an acceptable probability for human society.Infrastructure-assisted automated driving, including infrastructure sensors and vehicle-to-everything communication, is presented as one promising direction.
- All three approaches face significant challenges, but combining them may help solve the curse of rarity and accelerate large-scale AV deployment.
Conclusions
The paper concludes that the curse of rarity underlies critical autonomous-vehicle challenges and hinders deep-learning applications. It proposes potential approaches to address these challenges and calls for further investigation.
- The curse of rarity is identified as the underlying cause of critical AV challenges in perception, prediction, decision making, and verification and validation.
- The curse of rarity hinders the application of deep learning methods in autonomous vehicles.
- The paper proposes potential approaches to address these challenges and seeks further investigation by the research community.
Appendix
The appendix derives the variance relationship for an estimator based on rare events under an independence assumption.
- Under the stated independence assumption, the variance of μ2 is bounded by ρB times the variance of μ1.The derivation concludes with σ_μ2^2 ≤ ρBσ_μ1^2.