LoRA-RC introduces a method for adapting reservoir computing (RC) systems to handle system drift. The approach focuses on adapting the recurrent matrix within the reservoir, minimizing disruption to the echo state while maintaining prediction accuracy. The core adaptation process involves a low-rank correction driven by streaming prediction errors. This correction is applied to a small core matrix, projected onto a spectral-norm ball, and subsequently low-pass filtered at each step. This projection ensures the adapted matrix remains within a certified contraction set, and an incremental input-to-state stability bound is established for the reservoir along each online adaptation path.
Experiments were conducted on the Lorenz system with an abrupt parameter drift. LoRA-RC achieved a 56% reduction in post-drift prediction error compared to a fixed RC system. Readout-only adaptation achieved a 51% reduction in post-drift prediction error. Ablation studies, performed over 20 seeds, demonstrated that removing the projection resulted in an increase in prediction error exceeding 40 times the original error.
The method’s stability guarantees are key to its reliability. The projection and low-pass filtering contribute to maintaining incremental stability, preventing the reservoir from losing its echo-state properties during adaptation. This stability is crucial for online prediction tasks, particularly in dynamic environments where system parameters change over time.
Source: https://arxiv.org/abs/2609.12327
