NCVX: A General-Purpose Optimization Solver for Constrained Machine and Deep Learning

Constraints are becoming essential for trustworthy AI: The paper identifies that imposing explicit constraints in deep learning is increasingly necessary for building trustworthy AI systems, particularly when performing robust optimization against adversarial attacks and ensuring models respect real-world physical laws.

Existing tools create a gap for responsible AI developers: Current deep learning frameworks like PyTorch lack built-in constraint handling, while traditional optimization software cannot efficiently work with neural networks. NCVX (specifically its PyGRANSO solver) fills this gap by combining constraint-handling capabilities with deep learning tools, making it easier for practitioners without optimization expertise to build constrained models.

Accessibility matters for responsible AI adoption: By making constraint-based optimization freely available, open-source, and well-documented, NCVX lowers barriers to implementing responsible AI practices. The paper emphasizes that reliable constraint handling should not require specialized optimization expertise, democratizing access to trustworthy AI development methods.

The tool addresses multiple responsible AI use cases: NCVX is designed for diverse applications including robustness optimization (defending against adversarial perturbations) and scientific/domain-specific constraints, suggesting that responsible AI requirements span both security concerns and real-world applicability across different fields.