Residual Reservoir Memory Networks

Researchers have introduced two new neural network architectures: Residual Reservoir Memory Networks (ResRMNs) and Vision Hopfield Memory Network (V-HMN). ResRMNs combine linear and non-linear reservoirs for enhanced long-term input propagation, while V-HMN integrates hierarchical memory mechanisms for improved interpretability and data efficiency.

ResRMNs, a novel class of untrained Recurrent Neural Networks, were introduced on arXiv^[1] with a submission date of August 13, 2025, and last revised on May 29, 2026. The architecture combines a linear memory reservoir with a non-linear reservoir based on residual orthogonal connections along the temporal dimension. This design allows for enhanced long-term propagation of input signals. The proposed approach was empirically assessed on time-series and pixel-level 1-D classification tasks, showing advantages over conventional Reservoir Computing models. Separately, researchers proposed V-HMN, a brain-inspired foundation backbone that incorporates local and global Hopfield modules for associative memory dynamics and episodic memory, respectively^[2]. V-HMN achieved competitive results in public computer vision benchmarks while offering better interpretability and data efficiency compared to existing self-attention- or state-space-based approaches.