DisjunctiveNet: Neural Symbolic Learning via Differentiable Convexified Optimization Layers
Researchers have developed DisjunctiveNet, a novel framework for integrating hard, input-dependent mixed-integer linear constraints into neural networks. This approach addresses limitations in existing neuro-symbolic methods by representing rules as disjunctive constraints and employing hierarchical convex relaxations. The resulting convex hull formulations allow for exact rule satisfaction and end-to-end differentiability, demonstrating strong predictive performance on real-world datasets. AI
IMPACT Enables more robust integration of domain knowledge into AI models, potentially improving performance in scientific and engineering applications with sparse data.