AWS's ControlG uses PID controllers to fix multitask graph learning conflicts
ControlG assigns computation to one objective at a time, eliminating gradient compromise and negative transfer.
Amazon Web Services researchers, including Amazon PhD fellow Karish Grover and Amazon Scholar Christos Faloutsos, presented ControlG at ICML to tackle a core multitask learning problem: conflicting objectives dragging model training in different directions. Traditional graph self-supervised learning (graph SSL) combines objectives like link prediction, feature reconstruction, and contrastive learning by blending gradients at every step. This per-step mixing forces compromise, causing three failure modes: disagreement (negative transfer), drift (objectives losing relevance over time), and drought (some objectives starved of weight).
ControlG borrows proportional-integral-derivative (PID) controller principles from industrial control systems to separate objectives temporally. It operates across three time scales: measuring each objective's difficulty via spectral-demand and interference metrics, optimizing per-epoch computation allocation using log-hypervolume sensitivity, and tracking that plan with PID feedback loops. The result is an interpretable schedule—early training explores all objectives, mid-training prioritizes mutual information, and late training focuses on reconstruction that was lagging. This eliminates the tug-of-war between gradients and lets each objective receive dedicated capacity, improving downstream performance on graph tasks.
- ControlG uses PID controllers to allocate computational capacity sequentially to graph SSL objectives, avoiding per-step gradient compromise.
- The framework operates on three time scales: spectral-demand metrics, log-hypervolume sensitivity optimization, and PID feedback tracking.
- It eliminates three common multitask failures: disagreement (negative transfer), drift, and drought, with an interpretable training schedule.
Why It Matters
ControlG offers a practical way to train multitask models without negative transfer, improving graph representation learning for real-world applications.