Jacobian descent (JD) uses the Jacobian matrix of vector-valued objectives to resolve conflicting gradients in multi-objective optimization. The method guarantees convergence to the Pareto front in smooth convex cases and supports instance-wise risk minimization.
HOW THIS AFFECTS YOU
●
researcherYou can improve multi-objective optimization stability by projecting gradients to resolve conflicts directly.