While Multi-Task Learning (MTL) usually focuses on finding commonalities between related tasks, this research demonstrates that incorporating unrelated auxiliary tasks can actually improve performance on a principal task. By forcing the representations of unrelated task groups to be orthogonal, the model effectively "screens out" noise and task-specific quirks, leading to more robust and informative features.
The Problem: Negative Transfer
Standard MTL relies on the assumption that tasks share a common underlying structure. However, if you force unrelated tasks to share features, performance often degrades—a phenomenon known as "negative transfer."
The authors tackle a specific scenario: you have a set of principal tasks (e.g., recognizing emotions) and a set of auxiliary tasks that use the same input data but are known to be unrelated (e.g., recognizing the identity of the person). In small datasets, a model might mistake a person's unique facial structure for a sign of an emotion. The challenge is to use the knowledge that these tasks are unrelated to prevent the model from confusing their respective features.
Methods: OrthoMTL
The authors propose a framework called OrthoMTL (Orthogonal Multi-Task Learning). It builds on existing multi-task feature learning methods with three key components:
- Shared Low-Dimensional Representations: Within each group (principal or auxiliary), tasks are encouraged to share a small set of features.
- Orthogonality Penalty: The core innovation is a regularization term that penalizes the inner product between the predictor functions of tasks from different groups. This encourages the model to find two distinct, non-overlapping subspaces of features—one for the principal tasks and one for the auxiliary tasks.
- Optimization: Because the orthogonality term makes the problem non-convex, the authors introduce a "strongly convex" version using an Elastic Net-style penalty (OrthoMTL-EN).
Key Insights: Unrelatedness as a Filter
Knowing that two tasks are unrelated is just as valuable as knowing they are related. By explicitly modeling "unrelatedness" through orthogonality, the auxiliary tasks act as a filter, "soaking up" the features that would otherwise act as noise for the principal task.
This benefit is most pronounced when the training sample size per task is small. In these cases, models are prone to overfitting on "idiosyncrasies" (e.g., mistaking a specific person's squint for "anger"). OrthoMTL prevents this by assigning those identity-specific features to the auxiliary task's subspace.
Experimental Results
The researchers evaluated their approach on synthetic data, the JAFFE (Japanese Female Facial Expression) dataset, and the UNBC-McMaster Shoulder Pain Expression Archive.
Key findings include:
- Superior Accuracy: On the JAFFE dataset, OrthoMTL-EN significantly reduced misclassification error compared to standard MTL and Ridge Regression, especially at low training volumes (30-50 samples).
- Effective Separation: Visualization of the learned task correlation matrices showed that the model successfully drove the correlation between unrelated groups toward zero.
- Transfer Learning: In tests on the UNBC-McMaster archive, the method outperformed specialized "Bilinear Models" in transfer learning scenarios, predicting emotions for individuals not seen during training.
Conclusion
OrthoMTL provides a robust way to leverage auxiliary data even when that data is unrelated to the primary goal. While the method currently requires prior knowledge of task groupings, it opens new doors for automatic task relationship estimation and more efficient learning in data-scarce environments.
Summary updated