Reservoir computing adapts to system changes with low-rank updates
LoRA-RC: Reservoir Computing with Low-Rank Adaptation
Artificial IntelligenceMachine Learning
Summary
Reservoir computing is a method that predicts future data by training a simple part of a fixed network, which makes it fast but not flexible to changes. The authors introduce LoRA-RC, which allows the internal network to adjust slightly in a controlled way when the system it predicts changes. This new method keeps the network stable and reliable while improving its accuracy after sudden shifts in the data. Tests showed LoRA-RC greatly reduces errors compared to the standard method and a version that only adapts the output part.
What this means in practice
- •For real-time prediction engineers: Improve accuracy of online predictions in systems that experience sudden changes by applying low-rank updates to reservoir computing models.
- •For control system developers: Maintain stability and performance of recurrent models in adaptive control settings by enforcing constraints on network updates.
Authors
Wenbin Wan
Abstract
Reservoir computing (RC) trains only a linear readout over a fixed recurrent layer, making it fast and data-efficient for online prediction. However, a static reservoir degrades under system drift, readout-only adaptation is then insufficient, and unconstrained reservoir adaptation can destroy the echo-state and incremental stability properties that make RC reliable. This paper proposes LoRA-RC, which adapts the recurrent matrix through a low-rank correction driven by streaming prediction errors. The base reservoir and adaptation bases are fixed offline; a small core matrix is adapted online, projected onto a spectral-norm ball, and low-pass filtered at each step. The projection guarantees that every applied recurrent 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, with path-independent rate and gain. On a Lorenz system with an abrupt parameter drift, LoRA-RC cuts post-drift prediction error by 56% versus a fixed RC and 51% versus readout-only adaptation; ablations over 20 seeds show that removing the projection inflates this error by more than a factor of 40.