Tackling Multiple Tasks with One Single Learning Framework
This work addresses a specific bottleneck in sequential multi-task learning, offering an incremental improvement for researchers and practitioners in this domain.
The paper tackles the challenge of finding optimal knowledge sharing in deep multi-task learning for sequential problems, where task relationships change over time, by proposing the HTAN-SPD framework, which outperforms state-of-the-art methods in experiments on several applications.
Deep Multi-Task Learning (DMTL) has been widely studied in the machine learning community and applied to a broad range of real-world applications. Searching for the optimal knowledge sharing in DMTL is more challenging for sequential learning problems, as the task relationship will change in the temporal dimension. In this paper, we propose a flexible and efficient framework called HierarchicalTemporal Activation Network (HTAN) to simultaneously explore the optimal sharing of the neural network hierarchy (hierarchical axis) and the time-variant task relationship (temporal axis). HTAN learns a set of time-variant activation functions to encode the task relation. A functional regularization implemented by a modulated SPDNet and adversarial learning is further proposed to enhance the DMTL performance. Comprehensive experiments on several challenging applications demonstrate that our HTAN-SPD framework outperforms SOTA methods significantly in sequential DMTL.