Disentangling Computation in Multi-Task Neural Networks with the Green's Operator
The paper frames a trained recurrent network’s reuse of computation as a global perturbation-routing problem.
James Hazelden proposes using a finite-horizon Green’s operator to map where perturbations enter a trajectory and how they affect downstream network states. The reductions described in the abstract give task-to-task and time-to-time views of the same learned computation without building the full operator. In a flexible multitask recurrent network, the method is said to expose reused computational motifs, causal pathways, and how those pathways appear during training. Accepted as a NeurReps 2026 poster. ArXiv · AI/CL/LG's note
James Hazelden proposes using a finite-horizon Green’s operator to map where perturbations enter a trajectory and how they affect downstream network states. The reductions described in the abstract give task-to-task and time-to-time views of the same learned computation without building the full operator. In a flexible multitask recurrent network, the method is said to expose reused computational motifs, causal pathways, and how those pathways appear during training. Accepted as a NeurReps 2026 poster. ArXiv · AI/CL/LG's note
score 4