Neural Slack Variables for Shape Constraints

In three linesNew method to enforce inequality constraints (monotonicity, convexity) in neural networks using neural slack variables. Couples primary network with jointly learned auxiliary network, converting constraint enforcement into regression problem. Achieves zero measured violations on monotonicity/convexity test cases, outperforming penalty and primal-dual baselines.

Summary generated by Claude — human-verified