Abstract
Multi-task learning (MTL) is a powerful paradigm for sharing knowledge across related tasks, yet it faces significant challenges in domains like medicine due to label scarcity, noise, and extreme data imbalance. This paper introduces CO-TASK, a framework that enhances MTL by automatically generating auxiliary tasks through the combination of existing labels. By mapping fine-grained tasks into coarser auxiliary ones, CO-TASK provides "free" regularization without requiring additional human annotation. When paired with a novel task-aware imbalance data sampler, the framework demonstrates superior performance on benchmark image datasets and large-scale medical ECG records, notably outperforming human experts in screening for heart conditions.
Introduction
Multi-task learning (MTL) allows a single neural network to learn multiple related tasks simultaneously, leveraging shared representations to improve generalization. However, practical application—especially in healthcare—is hindered by three primary obstacles:
- Label Scarcity: Expert medical annotations are costly and rare.
- Label Noise: Human-provided labels often contain inconsistencies or errors.
- Extreme Imbalance: Medical datasets are typically dominated by "healthy" samples, with rare conditions often being ignored by standard models.
The CO-TASK framework addresses these issues by generating auxiliary "hints" for the model using information already present in the dataset, effectively performing label-level data augmentation.
Methodology: The CO-TASK Framework
The authors propose a two-pronged approach to bolster MTL performance: auxiliary task generation and specialized data sampling.
1. COmbination of TASK Labels (CO-TASK)
The core innovation lies in the realization that if a model can distinguish between specific classes, it should also be able to distinguish their unions. CO-TASK randomly selects pairs of existing tasks and maps their classes into new, coarser auxiliary labels. These auxiliary branches act as regularizers during training, forcing the shared layers of the network to learn more robust, generalized features. Crucially, these branches are used only during training and are discarded during inference, ensuring no increase in computational overhead during deployment.
2. Task-Aware Imbalance Data Sampler
To combat the varying imbalance ratios across different tasks, the authors developed a sampler that first selects a task (weighted toward major tasks) and then samples data using a weight function $f(x) = \log(x)/x$. This approach ensures that rare classes are sufficiently represented in training batches, preventing the model from converging on a solution that only favors common conditions.
Experimental Evaluation
The framework was evaluated across three diverse datasets: CIFAR-MTL (image classification), ECG-P18 (multi-label cardiovascular disease detection), and ECG-EchoLVH (predicting Left Ventricular Hypertrophy from ECG signals).
Key Results
- Generalization: On CIFAR-MTL, CO-TASK more than doubled the average per-task performance gain over single-task models (from 4.38% to 9.78%).
- Medical Phenotyping: In the ECG-P18 dataset, the "Exact" match rate for correctly identifying all 18 diseases in a patient jumped from 0.1925 to 0.3094.
- Clinical Impact: In the LVH prediction task, CO-TASK achieved a 7.1% increase in sensitivity over human doctor annotations while maintaining identical specificity, highlighting its potential as a superior screening tool.
Discussion and Conclusion
CO-TASK demonstrates that significant performance gains can be achieved by re-organizing existing label information. A key insight is the framework's robustness; the performance gap between CO-TASK and standard MTL actually widens as label noise increases, suggesting that combined "coarse" tasks provide a more stable learning signal than noisy "fine" tasks.
While highly effective, the approach does face diminishing returns; performance typically plateaus after approximately 50 auxiliary tasks. Future research may explore "smarter" task selection strategies rather than random combinations to further optimize knowledge transfer. Ultimately, CO-TASK provides a scalable, zero-cost method for improving multi-task models in data-constrained environments.
Summary updated