Symbolic regression typically hunts through combinations of a fixed set of operators—plus, times, sine, and so on—to find an expression that fits data. MorphingNet, described in a new arXiv paper, takes a different route: instead of searching over discrete expression trees, it trains a network whose units gradually alter their own functional form.

According to the abstract, each unit in MorphingNet changes its function continuously over the course of training. This means the architecture is not constrained by a predefined operator library; the form of the network itself is learned. The authors position this as a shift from discrete search to continuous optimization for symbolic regression.

The paper is short on details in its public abstract, so the full mechanics of how the morphing is parameterized or regularized are not yet clear. But the core idea is distinctive: the final network, after training, is the formula, rather than a separate expression extracted from a search process. If the approach scales, it could offer a new way to discover compact mathematical relationships from data.