In a standard neural network, we optimize a single loss function. In Multi-Task Learning (MTL), we often try to do the same by simply summing the losses of different tasks: $\mathcal{L}_{total} = \sum w_i \mathcal{L}_i$.
However, this "naive" summation creates a hidden competition. Tasks are not equal; some have steeper loss landscapes, larger reward scales, or more frequent data points. When we aggregate their gradients, the "loudest" task—the one with the largest gradient magnitude—effectively hijacks the shared parameters. This leads to a phenomenon researchers call gradient dominance, where the model ignores subtle tasks to satisfy the most demanding one.
The Magnitude Imbalance
The most immediate problem in MTL is that different tasks operate on different scales. For instance, in autonomous driving, a depth estimation task might produce gradients orders of magnitude larger than a semantic segmentation task. As Isabelle Leang and Ganesh Sistu observe, this discrepancy allows dominant tasks to "drown out" critical but smaller learning signals.
To fix this, researchers have moved toward dynamic normalization. Jana Kosecka introduced SLAW, which estimates gradient magnitudes from simple loss statistics to balance task importance efficiently. In reinforcement learning, where reward scales can vary wildly, Myungsik Cho proposed Adaptive Reward Scaling (ARS) to "equalize" these signals, preventing the agent from overfitting to high-reward tasks while treating difficult, low-reward tasks as mere noise.
Resolving Directional Conflicts
Even if magnitudes are equalized, tasks can still disagree on which direction the parameters should move. This is the "tug-of-war" or conflicting gradients problem. If Task A wants to move "left" and Task B wants to move "right," a simple average might result in no movement at all, or worse, a direction that harms both.
Several strategies have emerged to negotiate these conflicts:
- Orthogonal Decomposition: Xin Dong and colleagues developed GDOD, which decomposes gradients into shared and conflicting components. By removing the conflicting parts, the update only moves in directions that are mutually beneficial.
- Conflict-Averse Descent: Bo Liu and Peter Stone introduced CAGrad, which looks for an update direction that balances individual task improvements while minimizing the average loss, ensuring the optimizer doesn't oscillate between high-loss regions.
- Rotation and Alignment: Adrián Javaloy and Isabel Valera created Rotograd, which uses task-specific rotation matrices to actively align the directions of task gradients, forcing collaboration rather than cancellation.
Fairness and Game Theory
A recent trend treats multi-task optimization not as a math problem to be solved, but as a negotiation between competing agents. If we view each task as a "player," we can apply economic principles to find a stable compromise.
Hao Ban and Kaiyi Ji introduced FairGrad, which uses $\alpha$-fairness—a concept from communication networks—to ensure that no task is "starved" of progress. Similarly, Jing Wu and Mehrtash Harandi applied Nash Bargaining theory to resolve conflicts, deriving a closed-form solution that prevents any single objective from dominating the update.
Accounting for Gradient Noise
While most methods focus on magnitude and direction, Caoyun Fan and Yaohui Jin argue that stochastic noise is a neglected factor. In their MaxGNR framework, they point out that tasks with small gradients are particularly vulnerable to being "drowned" by the collective noise of other tasks. By maximizing the Gradient-to-Noise Ratio (GNR), they ensure that the learning signal remains clear even for the most difficult objectives.
Why It Works
The intuition behind these diverse methods is shared: a balanced backbone is a better generalizer. When one task dominates, the shared layers of the neural network become specialized for that task's specific features, leading to "negative transfer" where the model's performance on other tasks actually degrades. By enforcing "impartiality"—whether through geometric alignment, fair resource allocation, or noise-aware weighting—we force the model to find a representation that serves all tasks, leading to the "universal representations" sought by researchers like Wei-Hong Li.