Asymmetric Error Control in Safety-Critical Systems

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2704 papers970 conceptsfrozen 2026-10-04

In standard machine learning, we typically train models to minimize the average error. This "risk-neutral" approach assumes that all mistakes are created equal: misclassifying a cat as a dog is just as bad as misclassifying a dog as a cat.

However, in safety-critical domains, this symmetry collapses. As Bianca Zadrozny and Charles Elkan noted in their foundational work, real-world decisions involve costs that are often unknown and highly lopsided. In autonomous driving, confusing a pedestrian for a tree is a catastrophe; confusing a tree for a pedestrian is merely a cautious brake tap.

The research direction of Asymmetric Error Control focuses on moving beyond aggregate accuracy to build models that prioritize the avoidance of high-stakes "false negatives"—situations where a hazard is present but the system fails to detect it.

The Failure of Standard Objectives

Why can't we just use standard deep learning for these tasks? The problem lies in how modern models handle data. As Qiyuan Chen and colleagues point out in Rethinking Cost-sensitive Classification in Deep Learning via Adversarial Data Augmentation, over-parameterized neural networks are so powerful that they can often achieve near-zero training error. When a model fits the training data perfectly, traditional cost-weighted loss functions lose their "leverage"—there is no error left to weight differently.

This leads to a dangerous overconfidence. In robotics, Tianrui Guan and Dinesh Manocha observe that perception errors carry asymmetric risks. A robot navigating off-road must group terrain into navigability levels, where misclassifying a hazardous ditch as a safe path is significantly more dangerous than the reverse. Standard models, optimized for general accuracy, frequently fail to prioritize the minimization of these high-stakes false positives.

Shifting the Decision Boundary

To solve this, researchers have developed mechanisms to "push" the model's decision boundaries away from high-cost regions.

One approach is Adversarial Data Augmentation. Chen et al. introduced CSADA, which uses targeted adversarial examples to force the model to be more conservative. By simulating "near-miss" errors during training, the model learns to give high-cost classes a wider berth.

Another method involves Cost-Based Decision Rules. In the context of semantic segmentation for self-driving cars, Robin Chan and Fabian Hüger argue that the standard practice of picking the "most likely" class is actually a hidden ethical choice. They propose using a penalty matrix that explicitly weights the cost of confusing different objects (e.g., person vs. pole), allowing the car to adopt an "altruistic" or "egoistic" posture depending on the safety requirements.

Rigorous Statistical Guarantees

In fields like aviation or power grid management, "trying our best" to be safe isn't enough; we need mathematical proof.

  1. Neyman-Pearson Classification: Ye Tian and Yang Feng have bridged the gap between classical error-control theory and modern machine learning. Their framework allows practitioners to set a strict upper bound on specific error types (e.g., "the rate of missing a venomous snake must be less than 1%").
  2. PAC-Bayes Bounds: For vision-based robots, Alec Farid and Anirudha Majumdar use PAC-Bayes generalization theory to provide statistical guarantees on failure prediction. This allows a robot to "know what it doesn't know" and stop before a collision occurs.
  3. Conformal Prediction: Daniel Bethell and Simos Gerasimou developed COPPOL, which integrates conformal prediction into semantic segmentation. This generates hazard maps with rigorous guarantees on missed detections, significantly reducing violations in autonomous navigation.

The Power of Abstention

Sometimes the safest decision is to make no decision at all. In high-stakes classification, being forced to choose a label can lead to overconfident, dangerous errors.

Bhavya Kalra and Naresh Manwani introduced RISAN, a network that can "abstain" from predicting when uncertainty is high. Similarly, in web question answering, Jiafeng Guo and Yixing Fan emphasize that providing an incorrect direct answer is far worse than providing no answer. Their framework estimates predictive uncertainty to ensure the system "holds its tongue" when it isn't sure.

Domain-Specific Safety Margins

This asymmetric logic manifests differently across various industries:

  • Predictive Maintenance: In jet engine monitoring, Giovanni Iacca and Hyunho Mo note that optimistic errors (predicting an engine will last longer than it does) are far more costly than pessimistic ones.
  • Cybersecurity: In phishing detection, Atharva Bhargude and colleagues use adaptive prompting to ensure models maintain a "suspicious-first" posture, as missing one threat is worse than flagging a safe site.
  • Biological Identification: In the SnakeCLEF challenge, researchers like Feiran Hu prioritize the identification of venomous species, recognizing that a false negative in this context is a matter of life and death.

By explicitly modeling these asymmetries, the field is moving toward AI that doesn't just aim for the highest score, but for the safest outcome.

Papers in this version

Showing 60 of 704 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 735 papers
  2. Version 1 · 2026-10-04 · 704 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_asymmetric_misclassification_costs_f278c8ee,
  title = {asymmetric misclassification costs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-misclassification-costs/rdr_f278c8ee25d343f0b5ed91ffe25d2db4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_asymmetric_error_control_in_safety_criti_f278c8ee_v1,
  title = {Asymmetric Error Control in Safety-Critical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-error-control-in-safety-critical-systems/rdr_f278c8ee25d343f0b5ed91ffe25d2db4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}