Introduction
Traditional AI ethics often relies on "post-hoc audits"—checking for bias after a model has been deployed. This paper, presented at AAAI-20, proposes a shift toward "active learning" through a framework called Embedding Principles of ethics by Design (EPbD). By integrating formal ethical ontologies directly into the neural network training process, the authors demonstrate how to build models that are inherently aware of ethical constraints.
The Limitations of Reactive Ethics
The authors argue that reactive approaches are insufficient for three primary reasons:
- Search Space Complexity: It is practically impossible to test every input state in complex systems.
- Data Scarcity: In domains like digital lending, the lack of data on "false negatives" (e.g., rejected applicants who would have been reliable) complicates post-hoc validation.
- Alignment Gap: Human values should be integrated during development, not treated as an afterthought.
The challenge lies in managing the "graded" nature of ethics, where decisions exist on a spectrum of acceptability rather than a simple binary.
Methodology: The Ethical by Design Neural Network (EbDNN)
The core contribution is the Ethical by Design Neural Network (EbDNN) architecture, which uses multi-task learning to optimize for both business objectives and ethical compliance.
The architecture consists of three main stages:
- Ethics Encoder: Fully connected layers that learn a latent representation of input data specifically tuned for ethical judgment.
- Multi-Task Experts:
- Business Expert (BE): A sub-network trained to predict standard "gold standard" outcomes (e.g., loan default risk).
- Ethics Expert (EE): Reconstructs an "ethical signature" using an Ethical Ontology (EO)—formal rules mapping raw features to ethical dimensions like social inclusiveness.
- Ethics-aware (EA) DNN: Integrates signals from both experts to estimate the joint probability of a decision and its associated ethical risks.
The system includes tunable parameters: Smoothing ($\alpha$) to allow deviations from biased historical data, and Tweaking ($\beta$) to adjust the ethical signal's influence during training.
Experimental Results
The framework was evaluated using the German Credit dataset, with an ontology focused on "Motherhood Fostering" and "Cultural Inclusiveness." The authors introduced a new metric, Acceptability ($EAcc_\gamma$), to quantify the trade-off between business performance (F1-score) and ethical compliance.
Key findings include:
- Baseline Performance: A standard Multi-Layer Perceptron (MLP) achieved 66.4% F1 but only 77.8% ethical compliance.
- Full Compliance: The Ethics-Constrained model ($EA^{EC}$) reached 100% ethical compliance, though business accuracy dropped significantly to 41.2% F1.
- The Optimal Balance: A configuration with $\alpha=0.3$ and $\beta=0.1$ achieved 60.6% F1 and 85.1% ethical compliance, resulting in a superior overall Acceptability score of 80.2% compared to the baseline's 75.5%.
Conclusion
This research demonstrates that ethical principles can act as a form of regularization, steering models away from biased patterns in training data. While the current study uses a small dataset and relatively simple ethical rules, it provides a foundation for more complex implementations, including potential applications in Reinforcement Learning where agents learn through interaction.
Summary updated