Multitask Learning in Deep Neural Networks

Multitask learning in deep neural networks is an approach in which a single model is trained to perform multiple related tasks concurrently, exploiting commonalities and differences across tasks to improve generalisation and efficiency. By sharing internal representations, models learn to extract features that capture underlying structure relevant to all tasks, reducing the risk of overfitting and enabling knowledge transfer. Key challenges include balancing task-specific objectives, managing conflicting gradients and ensuring that no single task dominates the optimisation process. Recent advances have focused on adaptive loss weighting, dynamic architecture search and integration with meta-learning schemes to automate bias discovery and facilitate rapid adaptation. Practical applications span computer vision, natural language processing, healthcare and robotics, where multitask frameworks yield more robust and data-efficient solutions than isolated single-task models.

Research from Nature Portfolio

Recent studies have demonstrated that combining meta-learning with multitask objectives can enable models to adapt swiftly to new tasks with minimal additional training. One approach employs a bilevel optimisation scheme in which a shared feature extractor is refined by task-specific heads, allowing efficient generalisation across disparate domains such as genomics and clinical imaging. Another work introduces a transformer-based multitask architecture that unifies vision and language tasks by sharing attention modules, resulting in improved performance on image captioning, visual question answering and scene understanding benchmarks. A third investigation explores curriculum learning strategies to sequence task exposure, revealing that organising tasks by difficulty accelerates convergence and enhances final accuracy across tasks with varying complexity.

Research from all publishers

Research outside the Nature portfolio has contributed diverse strategies for effective multitask learning. A recent study in the International Journal of Computer Vision presents universal representations that distil multiple task- and domain-specific networks into a single backbone, achieving state-of-the-art performance on dense prediction and classification benchmarks while demonstrating robust cross-domain few-shot generalisation. An investigation in IEEE Access systematically evaluates loss weighting strategies—including uniform summation, dynamic weight averaging and uncertainty-based weighting—showing that no single method uniformly outperforms others, and emphasising the need for task-pair selection and adaptive schemes. In remote sensing, a boundary-aware multitask framework jointly addresses semantic segmentation, height estimation and edge detection, using auxiliary boundary maps to regularise feature learning and refine object contours, thereby improving both segmentation accuracy and geometric consistency in aerial imagery.

Multitask Learning in Deep Neural Networks publication trend

The graph below shows the total number of articles in multitask learning in deep neural networks across all publications each year (not limited to Nature Index journals).

Technical terms

Multi-task learning: A framework in which a single model is trained to perform several related tasks simultaneously, leveraging shared representations.

Parameter sharing: The technique of using common network weights across tasks to promote information transfer and reduce model complexity.

Loss weighting: Methods for balancing the contribution of each task’s loss during joint optimisation to ensure stable and effective training.

Meta-learning: A paradigm that enables models to rapidly adapt to new tasks by learning how to learn across tasks.

Universal representation: A unified feature embedding that supports multiple tasks and domains through shared network structures.

References

1. An overview of multi-task learning. National Science Review (2017).
2. Universal Representations: A Unified Look at Multiple Task and Domain Learning. International Journal of Computer Vision (2023).
3. A Comparison of Loss Weighting Strategies for Multi task Learning in Deep Neural Networks. IEEE Access (2019).
4. Boundary-Aware Multitask Learning for Remote Sensing Imagery. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2020).