Boosting Reservoir Computing with Brain-inspired Adaptive Dynamics

TL;DR

This work shows that reservoir computers perform best in balanced or slightly inhibited dynamical regimes, and that explicit brain-inspired mechanisms—local inhibitory adaptation and firing-rate heterogeneity—greatly boost performance and robustness. By moving from static reservoir optimization toward dynamic E-I balance tuning (and a one-step design alternative), the authors achieve up to about 130% improvement across memory and nonlinear time-series tasks. The study integrates neurobiological principles, such as Dale’s law, sigmoid activation, and homeostatic-like plasticity, to enhance scalability and reduce hyperparameter tuning. The results illuminate a pathway for more robust, brain-inspired RCs applicable to real-time, large-scale computation while deepening our understanding of neural computation in dynamical regimes.

Abstract

Reservoir computers (RCs) provide a computationally efficient alternative to deep learning while also offering a framework for incorporating brain-inspired computational principles. By using an internal neural network with random, fixed connections$-$the 'reservoir'$-$and training only the output weights, RCs simplify the training process but remain sensitive to the choice of hyperparameters that govern activation functions and network architecture. Moreover, typical RC implementations overlook a critical aspect of neuronal dynamics: the balance between excitatory and inhibitory (E-I) signals, which is essential for robust brain function. We show that RCs characteristically perform best in balanced or slightly over-inhibited regimes, outperforming excitation-dominated ones. To reduce the need for precise hyperparameter tuning, we introduce a self-adapting mechanism that locally adjusts E/I balance to achieve target neuronal firing rates, improving performance by up to 130% in tasks like memory capacity and time series prediction compared with globally tuned RCs. Incorporating brain-inspired heterogeneity in target neuronal firing rates further reduces the need for fine-tuning hyperparameters and enables RCs to excel across linear and non-linear tasks. These results support a shift from static optimization to dynamic adaptation in reservoir design, demonstrating how brain-inspired mechanisms improve RC performance and robustness while deepening our understanding of neural computation.

Figures (6)

• Figure 1: Balanced to slightly inhibited reservoirs exhibit robust, high performance devoid of synchronized or saturation dynamics.(A) Three-layered schematic of the E-I RC architecture. Excitatory (solid gray) and inhibitory (striped) reservoir neurons with sigmoid input-output function form excitatory (solid) and inhibitory (dashed) connections, respectively. (B) RC evaluation across tasks : i) Memory task - RC runs in open-loop ($\ast$) mode, receiving input $u(t)$ and predicting it $d$ steps later. Accuracy is measured by $R^2$, with memory capacity as the area under the $R^2$ vs. delay curve. ii) Other tasks - RC runs in open-loop (NARMA-10) or closed-loop (Mackey-Glass, Lorenz) mode ($\ast$$\ast$), predicting one step ahead. Accuracy is measured by RMSE or Valid Prediction Time (VPT). (C) RC dynamics and performance in the memory test for $\theta=0$. Left: Mean firing rate, Middle: Neuronal entropy and Right: Memory capacity. Synchronized (inset, green shaded) and saturated (red shaded) regimes are highlighted. (D) RC dynamics and performance in the memory test with variable thresholds reveals dynamical landscape with three extreme regimes: silent ($\sim$0), saturated ($\sim$1), and globally synchronized ($\bar{C_{ij}}>0.9$, neither silent nor saturated). Neuronal entropy maps these regions, showing reduced entropy in all three. Memory capacity is highest in balanced/slightly inhibited states and declines in the over-excited regime.
• Figure 2: The inhibitory adaptation mechanism enhances both global and local balance, improving memory performance across homogeneous and heterogeneous target firing rates.(A) Schematic of adaptation mechanism: inhibitory connections strengthen for neurons with firing rates (fill color) above their target rate (border color) and weaken for those below, restoring local balance. (B) Evolution of global balance and memory capacity across adaptation steps for five initial conditions (ranging from over-inhibited: dark blue, to over-excited: dark red) with homogeneous target firing rates. All networks converge to a balanced state with improved memory. Even the initially globally balanced reservoir (green curve) has improved memory capacity without a visible shift in global balance, as correcting large local imbalances (inset) enhances performance. (C) Similar analysis with heterogeneous target firing rates (drawn from a beta distribution). The network converges to global balance and improved performance while preserving slight variations in local balance (inset).
• Figure 3: Adaptive reservoirs achieve optimal performance at different target firing rates for different tasks, significantly outperforming non-adaptive, globally tuned reservoirs. (A) Optimal performance (purple line: mean across 100 runs, shaded area: standard error of the mean) varies with homogenous target firing rate ($\rho_T$). Linear tasks, such as memory capacity, peak near the linear midpoint of the sigmoid (dashed line), while others such as NARMA-10, Mackey-Glass and Lorenz benefit from slight deviations above or below 0.5, leveraging nonlinear effects. (B) Performance of non-adaptive vs. adaptive reservoirs. Gray bars: best non-adaptive, globally tuned RC. Colored bars: adaptive reservoirs with tuned homogeneous (purple) or fixed heterogeneous (pink) target rates. Adaptive reservoirs significantly improve performance, with heterogeneous rates offering robust, task-general performance without requiring firing rate tuning.
• Figure 4: Adaptive mechanisms enhance RC performance by removing locally imbalanced, low-entropy neurons. (A) Scatter plots show that in a globally balanced, non-adaptive RC (gray) many neurons exhibit extreme firing rates and local imbalances. An adaptive RC with a homogeneous target rate of 0.4 (purple), on the other hand, shifts neurons to the desired rate reducing local imbalances. Heterogeneous targets (pink) avoid extreme imbalances while maintaining diverse firing rates. (B) Imbalanced neurons with extreme firing rates (saturated or silent) have low entropy. Adaptation mechanism eliminates these low-entropy neurons for both homogeneous and heterogeneous target rates. (C) Neurons with extreme firing rates and low entropy contribute minimally to predictive performance, as reflected in their low output weights. By avoiding these extreme rates, our inhibitory adaptation mechanism effectively utilizes these neurons, improving predictive performance (see also Fig. 3).
• Figure 5: The adaptive mechanism improves RC performance across input link scaling, with the most significant gains occurring near each task's optimal scaling, determined by the memory-nonlinearity tradeoff. Each plot represents a specific task, arranged from memory-dominant (left) to nonlinearity-dominant (right). Colored lines show the mean across 100 runs, with the shaded area indicating the standard error of the mean. Performance of the best non-adaptive, globally tuned system (gray) is compared to the adaptive system with homogeneous (purple) and heterogeneous (pink) target rates as a function of input link scaling, $\sigma_{in}$. As task nonlinearity increases, the optimal input link scaling for the non-adaptive system ($\blacktriangle$) shifts to higher values, while adaptation consistently improves performance, particularly near each task’s optimal input link scaling.