Residual Reservoir Memory Networks: A New Class of Untrained RNNs
A new type of untrained Recurrent Neural Networks, called Residual Reservoir Memory Networks (ResRMNs), has been introduced by researchers within the Reservoir Computing framework. This innovative model integrates a linear memory reservoir with a non-linear counterpart that includes residual orthogonal connections across time, improving the propagation of long-term inputs. The dynamics of the reservoir state are examined through linear stability analysis, and various configurations of temporal residuals are explored. Empirical tests on time-series and pixel-level 1-D classification tasks demonstrate superior performance compared to traditional RC models. This research is detailed in a paper available on arXiv (2508.09925) under the category of Computer Science > Machine Learning.
Key facts
- Residual Reservoir Memory Networks (ResRMNs) are a new class of untrained RNNs.
- ResRMN combines a linear memory reservoir with a non-linear reservoir.
- Non-linear reservoir uses residual orthogonal connections along the temporal dimension.
- Reservoir state dynamics studied via linear stability analysis.
- Tested on time-series and pixel-level 1-D classification tasks.
- Outperforms conventional Reservoir Computing models.
- Paper available on arXiv with ID 2508.09925.
- Submitted to Computer Science > Machine Learning.
Entities
Institutions
- arXiv