Home▸Data Science▸Gradient Surgery for Conflicting Tasks: PCGrad Projection on Two Quadratic Losses and Pareto Stationarity

Gradient Surgery for Conflicting Tasks: PCGrad Projection on Two Quadratic Losses and Pareto Stationarity

Two tasks pull one shared parameter vector in opposite directions. Project each gradient onto the normal plane of the other whenever they conflict, and compare the path with plain gradient summing. Change the curvature mismatch, the distance between the task optima and the step size, then read the conflict fraction, both losses and the Pareto stationarity gap. A Blender-built plotting easel with exact orthogonality, descent and finite-difference checks built in.

Data Science3DModerate60 FPS📱 Mobile-adapted🟠 Blender⇄ 2D version
Gradient Surgery for Conflicting Tasks: PCGrad Projection on Two Quadratic Losses and Pareto Stationarity ↗ Open standalone

Two tasks pull one shared parameter vector in opposite directions. Project each gradient onto the normal plane of the other whenever they conflict, and compare the path with plain gradient summing. Change the curvature mismatch, the distance between the task optima and the step size, then read the conflict fraction, both losses and the Pareto stationarity gap. A Blender-built plotting easel with exact orthogonality, descent and finite-difference checks built in.

⚙ Under the hood

Two tasks pull one shared parameter vector in opposite directions. Project each gradient onto the normal plane of the other whenever they conflict, and compare the path with plain gradient summing. Change the curvature mismatch, the distance between the task optima and the step size, then read the conflict fraction, both losses and the Pareto stationarity gap. A Blender-built plotting easel with exact orthogonality, descent and finite-difference checks built in.

Gradient SurgeryPCGradMulti-Task LearningConflicting GradientsPareto Stationarity

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install

What did you find?

Add reproduction steps (optional)