Chaotic resonance can amplify responses of weak external inputs through intrinsic nonlinear fluctuations; however, the effective control of this phenomenon becomes challenging when the attractor dynamics are structurally asymmetric. Building upon preliminary findings, prior studies have investigated whether an asymmetric modification of the double-Gaussian reducing region of the orbit (DG-RRO), allowing independent adjustment of the feedback strength for each branch, could enhance synchronization in asymmetric cubic maps. However, a formalized framework and evaluation under noisy conditions have not yet been provided. In this study, we formalize and extend this concept by introducing an innovative control method called double-Gaussian-asymmetric-filtered reducing-region-of-orbit (DGA-RRO). The DGA-RRO retains the localized, low-intrusiveness filtering of the original DG-RRO while assigning independent Gaussian gains to each side of the orbit, permitting branch-specific tuning and simultaneous attractor-merging bifurcations under pronounced asymmetry. Through systematic numerical experiments across varying asymmetry levels, input amplitudes, frequencies, and two noise models (additive and contaminant), we show that DGA-RRO achieves a higher input-output correlation than symmetric DG-RRO, enables attractor-merging bifurcations to occur under weaker feedback strength, and preserves strong synchronization over a broader parameter range. This method is particularly resilient to contaminant noise and exhibits a well-defined frequency band for optimal performance. These results position the DGA-RRO as a practical extension of earlier asymmetric feedback tests and suggest its applicability to engineering tasks requiring high sensitivity with minimal perturbation, such as neural computation, sensing, and low-power control.

Neural interactions under optimal excitatory/inhibitory (E/I) balance are among the most crucial mechanisms for realizing cognitive functions. Among the phenomena supported by this mechanism, the duration of a phenomenon known as perceptual alternation exhibits two representative characteristics: nondeterminism and the long-tailed property at the level of a large neural population. However, even in a system consisting of a single pair of excitatory and inhibitory neurons, called chaos-chaos intermittency (CCI), a similar intermittent alternation of neural activity emerges, involving intermittent transitions between multiple isolated attractors. We hypothesized that the characteristics of CCI dynamics in local excitatory-inhibitory neural circuits can describe the nondeterminism and long-tailed properties observed at a broad hierarchical level. We evaluated the changes in nondeterminism and long-tailed properties under different E/I balance conditions to test this hypothesis. First, we validated the determinism of two types of dynamics: 1) transitions between attractors and 2) behavior within attractors. This evaluation was performed using iterated amplitude-adjusted Fourier transform and multi-scale entropy analysis. Next, we characterize the long-tailed properties of the alternations. These properties were evaluated while gradually shifting the parameters from attractor-merging bifurcation. These results indicate that while behavior within attractors demonstrate determinism across all conditions, transitions between attractors lose nondeterminism as the predominance of excitatory neuron increases. Furthermore, the duration histograms lose their long-tailed properties as excitatory neurons become dominant. Consequently, the disappearance of determinism and long-tailed properties co-occurs, and the coexistence of nondeterminism and long-tailed properties is realized within specific domains of the E/I balance. This discovery contributes to our understanding of the importance of an optimal E/I balance for maintaining the characteristics of interactions between excitatory and inhibitory neurons. Copyright © 2025 The Institute of Electronics, Information and Communication Engineers.

Reservoir computing (RC), particularly the echo state network (ESN), is an efficient framework for time-series processing. However, its conventional form often struggles with tasks characterized by high dimensionality and multiscale temporal dynamics. We propose the heterogeneous assembly ESN (HetAESN), a novel architecture that extends the assembly ESN (AESN) by incorporating temporal heterogeneity. HetAESN partitions high-dimensional input and assigns each sub-reservoir an optimized, distinct time constant, enabling it to adapt to the specific temporal properties of its input components. We validated HetAESN using time-series prediction tasks on three chaotic systems: the two-coupled van der Pol (tc-VdP), the Hindmarsh–Rose (HR), and the two-coupled Lorenz (tc-Lorenz). HetAESN achieved superior prediction accuracy compared to conventional ESN and AESN models for the tc-VdP and HR tasks. To analyze these results, we employed delay capacity and multiscale fuzzy entropy. Our analysis revealed that the model’s effectiveness critically depends on the balance between the task’s dimensionality and the signal’s complexity within the reservoir’s effective memory range. This study clarifies the crucial relationship between architectural design—specifically dimension splitting and multiscale adaptation— and the computational capability of RC models, paving the way for developing more robust, generalized architectures for processing high-dimensional, multiscale time series. © 2013 IEEE.