A Generalised Framework for Property-Driven Machine Learning

2025-05-01

AI Summary

Neural networks often fail to satisfy formal properties required in safety-critical applications. Method: This paper proposes an attribute-driven unified training framework that jointly integrates geometric constraints from adversarial training (generalized hyperrectangular input domains) and semantic constraints encoded via differentiable first-order logic, thereby translating arbitrary formal properties into differentiable loss terms. The framework jointly optimizes a property-weighted loss function and neural network controllers. Contribution/Results: It enables concurrent assurance of robustness and correctness under flexible, domain-specific regional specifications across diverse fields (e.g., control systems, NLP). Evaluated on a neural controller for unmanned aerial vehicles, the framework achieves significant improvement in formal property satisfaction rates. Open-sourced and fully reproducible, it supports a broad range of canonical formal properties, demonstrating strong generalizability and plug-and-play usability.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Adversarial Learning & Robustness
• Multiagent Systems: Adversarial Agents

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Neural networks have been shown to frequently fail to satisfy critical safety and correctness properties after training, highlighting the pressing need for training methods that incorporate such properties directly. While adversarial training can be used to improve robustness to small perturbations within $epsilon$-cubes, domains other than computer vision -- such as control systems and natural language processing -- may require more flexible input region specifications via generalised hyper-rectangles. Meanwhile, differentiable logics offer a way to encode arbitrary logical constraints as additional loss terms that guide the learning process towards satisfying these constraints. In this paper, we investigate how these two complementary approaches can be unified within a single framework for property-driven machine learning. We show that well-known properties from the literature are subcases of this general approach, and we demonstrate its practical effectiveness on a case study involving a neural network controller for a drone system. Our framework is publicly available at https://github.com/tflinkow/property-driven-ml.

Problem

• Ensuring neural networks satisfy safety and correctness properties post-training
• Extending adversarial training beyond ε-cubes to hyper-rectangles for diverse domains
• Unifying differentiable logics and region-based methods for property-driven ML

Innovation

• Unifies adversarial training with hyper-rectangle specifications
• Encodes logical constraints via differentiable logics
• General framework for property-driven ML training