AdaHAT: Adaptive Hard Attention to the Task in Task-Incremental Learning

AdaHAT extends Hard Attention to the Task (HAT) for task-incremental learning. The method adaptively updates parameters that would otherwise remain static, using information about their importance to previous tasks and the network’s remaining capacity. The authors evaluate AdaHAT on multiple datasets against HAT and other task-incremental learning baselines, reporting improved average performance across tasks, particularly for long task sequences. The paper frames the method as a way to balance stability and plasticity while reducing the capacity bottleneck caused by progressively frozen parameters. Code is available through the authors’ project page.