de07eddf1b
- Applied reviewer-requested changes - Quality gate pass (fix-from-feedback) Pentagon-Agent: Auto-Fix <HEADLESS>
30 lines
No EOL
2 KiB
Markdown
30 lines
No EOL
2 KiB
Markdown
---
|
|
type: claim
|
|
claim: machine-learning pattern extraction systematically erases outliers where vulnerable populations concentrate
|
|
domain: ai-alignment
|
|
confidence: established
|
|
description: Machine learning systems using empirical risk minimization systematically underfit to low-density regions of feature space where minority populations concentrate, resulting in higher prediction error for vulnerable groups. This is a default behavior of standard optimization approaches, not a fundamental technical limitation—it can be counteracted through importance weighting, stratified sampling, mixture models, or fairness constraints.
|
|
created: 2024-01-01
|
|
processed_date: 2024-01-01
|
|
source:
|
|
- ai4ci-national-scale-collective-intelligence
|
|
---
|
|
|
|
Machine learning systems optimize for patterns in training data through empirical risk minimization, which with finite samples systematically underfits to low-density regions of feature space. Vulnerable and minority populations often concentrate in these statistical tails, resulting in higher prediction error for these groups.
|
|
|
|
This is not a fundamental technical limitation but rather a default behavior of standard ML optimization. The AI4CI strategy document identifies this as a key challenge for collective intelligence systems and proposes technical countermeasures including:
|
|
|
|
- Importance weighting (upweighting minority examples)
|
|
- Stratified sampling (ensuring tail coverage)
|
|
- Mixture models (separate models for subpopulations)
|
|
- Fairness constraints (explicit tail performance requirements)
|
|
- Federated learning approaches
|
|
- Explicit outlier protection mechanisms
|
|
|
|
The challenge is primarily one of governance and prioritization—current systems often don't implement these solutions—rather than technical impossibility.
|
|
|
|
## Related
|
|
|
|
- [[RLHF and DPO fail to preserve diversity in human preferences]]
|
|
- [[partial connectivity preserves diversity in collective intelligence systems]]
|
|
- [[safe AI development requires building alignment mechanisms before scaling capability]] |