Why the WCRR-NN learns the MCP
Abijith Jagannath Kamath, Visiting Ph.D. Student (IISc Bangalore)

Seminar • 2026-03-10

Abstract
The weakly-convex ridge regularisation neural network (WCRR-NN) is a technique for the explicit design of weakly-convex regularisation functions for solving linear inverse problems. A surprising observation is that the network learns the activation function that corresponds to the potential of the minimax concave penalty (MCP). This presentation demonstrates a candidate explanation for this observation via a design philosophy for weakly-convex functions that naturally results in the MCP, and discusses the implications in solving linear inverse problems.