The emergence of large-scale data-driven machine learning and optimization methods has led to successful applications in areas as diverse as finance, marketing, retail, and health care. Yet, many application domains remain out of reach for these methods, when applied in isolation. In the area of medical robotics, for example, it is crucial to develop systems that can recognize, guide, support, or correct surgical procedures. This is particularly important for next-generation trauma care systems that allow life-saving surgery to be performed remotely in the presence of unreliable bandwidth communications. For such systems, machine learning models have been developed that can recognize certain patterns, but they are unable to perform under complex physical or operational constraints. Using constraint-based optimization methods, on the other hand, would allow the generation of feasible surgical plans; but currently, there is no mechanism to represent and evaluate such knowledge under complex environments. To leverage the required capabilities for real-life applications, this project develops an integrated method that Embeds Constraint Reasoning in Machine Learning (ECOR-ML). The researchers intend to demonstrate the effectiveness of ECOR-ML in the context of medical robotics. Prior research indicates that the integration of constraint reasoning and machine learning is essential for the development of safe and efficient technologies in this domain. The project aims to advance both machine learning and constraint reasoning technology, and will promote the cross-fertilization of formal and applied research in the areas of machine learning, constraint learning, and robotics.
The approach in this project provides a scalable method for machine learning over structured domains. The core idea is to augment machine learning algorithms with a constraint reasoning module that represents physical and operational requirements. Specifically, this research proposes to embed decision diagrams, a popular constraint reasoning tool, as a fully-differentiable layer in deep neural networks. By enforcing the constraints, the output of generative models can now provide assurances of safety, correctness, and/or fairness. Moreover, ECOR-ML possesses a smaller modeling space than traditional machine learning approaches, allowing machine learning algorithms to learn faster and generalize better.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
