MGDA
Multiple Gradient Descent Algorithm (MGDA) is a multi-objective optimization algorithm that seeks a common gradient direction along which all objective functions simultaneously decrease, used to find Pareto-optimal solutions in multi-task and multi-objective learning settings.
Timeline (last 90 days)
The gradient correction method MGDA and its variants model multi-client objectives as multi-objective optimization to find Pareto-optimal solutions, but their common descent direction is not equivalent to the global aggregation direction, making global accuracy loss unavoidable.
Representative methods for gradient balancing include PCGrad (projecting conflicting gradients onto a conflict-free subspace) and MGDA (a multiple gradient descent algorithm that finds a direction allowing all objectives to descend simultaneously).
All Facts (2)
The gradient correction method MGDA and its variants model multi-client objectives as multi-objective optimization to find Pareto-optimal solutions, but their common descent direction is not equivalent to the global aggregation direction, making global accuracy loss unavoidable.
50%UnverifiedRepresentative methods for gradient balancing include PCGrad (projecting conflicting gradients onto a conflict-free subspace) and MGDA (a multiple gradient descent algorithm that finds a direction allowing all objectives to descend simultaneously).
50%