Gated Target Propagation for Compositional Generalization in Continual Learning
GaTaP uses learned gates to reuse network modules for new task combinations while limiting forgetting.
The paper introduces a continual learning method where task-specific gating variables suppress or amplify parts of a neural network. Those gates are adapted with a closed-form inner-loop update, while the main parameters learn more slowly through target propagation. In class-incremental experiments with MLPs and convolutional networks, the authors report both retained performance on old tasks and few-shot compositional generalization to unseen ones. Their gating analysis finds related tasks producing similar gate patterns, which they interpret as reusable task structure. ArXiv · AI/CL/LG's note
The paper introduces a continual learning method where task-specific gating variables suppress or amplify parts of a neural network. Those gates are adapted with a closed-form inner-loop update, while the main parameters learn more slowly through target propagation. In class-incremental experiments with MLPs and convolutional networks, the authors report both retained performance on old tasks and few-shot compositional generalization to unseen ones. Their gating analysis finds related tasks producing similar gate patterns, which they interpret as reusable task structure. ArXiv · AI/CL/LG's note
score 3