Liquid State Machines (LSMs), a reservoir computing model based on recurrent spiking neural networks, provide a powerful framework for solving spatiotemporal classification tasks by leveraging rich temporal dynamics and event-driven processing. Although the traditional LSM formulation assumes a fixed, randomly generated reservoir, recent research has explored optimization strategies to improve liquid dynamics. However, most existing approaches focus primarily on optimizing synaptic connectivity or reservoir structure, while the role of neuron-level parameters remains largely underexplored. This work proposes a neuroevolutionary strategy based on a Genetic Algorithm (GA) that encodes both neuron configurations and their spatial positions, explicitly treating neuron-level parameters as optimization targets. By evolving neuron-specific parameters and spatial positions, the method induces diverse reservoir dynamics. Unlike approaches that directly optimize synaptic weights, the proposed representation maintains an encoding whose dimensionality scales linearly with the number of neurons. The approach was evaluated on four synthetic benchmark tasks, including one Frequency Recognition task and three Pattern Recognition tasks, using compact reservoirs composed of only 20 Leaky Integrate-and-Fire neurons. Despite the small reservoir size, the method achieved state-of-the-art or highly competitive performance, reaching mean accuracies of up to 99.71%. In the most challenging case (PR12), performance improved when the reservoir size was increased to 64 neurons. The method was further evaluated on two real-world datasets, N-MNIST and the Free Spoken Digit Dataset (FSDD), using reservoirs of 300 neurons, achieving 90.65% and 81.47% accuracy, respectively, while using substantially fewer neurons than many existing LSM-based approaches. These results highlight the potential of evolving neuron configurations and spatial organization to produce compact and effective liquid reservoirs.

Abstract Recurrent spiking neural networks (RSNNs) can be implemented very efficiently in neuromorphic systems. Nevertheless, training of these models with powerful gradient-based learning algorithms is mostly performed on standard digital hardware using Backpropagation through time (BPTT). However, BPTT has substantial limitations. It does not permit online training and its memory consumption scales linearly with the number of computation steps. In contrast, learning methods using forward propagation of gradients operate in an online manner with a memory consumption independent of the number of time steps. These methods enable SNNs to learn from continuous, infinite-length input sequences. In addition, approximate forward propagation algorithms have been developed that can be implemented on neuromorphic hardware. Yet, slow execution speed on conventional hardware as well as inferior performance has hindered their widespread application. In this work, we introduce HYbrid PRopagation (HYPR) that combines the efficiency of parallelization with approximate online forward learning. Our algorithm yields high-throughput online learning through parallelization, paired with constant, i.e., sequence length independent, memory demands. HYPR enables parallelization of parameter update computation over subsequences for RSNNs consisting of almost arbitrary non-linear spiking neuron models. We apply HYPR to networks of spiking neurons with oscillatory subthreshold dynamics. We find that this type of neuron model is particularly well trainable by HYPR, resulting in an unprecedentedly low task performance gap between approximate forward gradient learning and BPTT.

This paper explores the implementation of Traveling Wave Protection (TWP) in microgrids through the integration of Internet of Things (IoT) technologies and a Spiking Recurrent Neural Network (SRNN). Microgrids present unique fault de-tection challenges, as conventional protection techniques can be hindered by reduced fault currents, bidirectional power flow, and communication latency. By leveraging high-frequency traveling wave signals, TWP offers rapid and precise fault localization. In parallel, IoT-enabled sensing provides real-time data acquisition and decentralized decision-making. The proposed SRNN further enhances fault classification and location accuracy by combining spiking neuron dynamics with recurrent memory. Hardware-in-the-loop experiments on both simplified and complex microgrids demonstrate the method’s effectiveness in minimizing misclas-sification while maintaining low latency and reduced power consumption. This work extends our previous IoT-based TWP research by adopting a neuromorphic framework suitable for microgrid edge deployments, paving the way for more adaptive and robust protection solutions in modern distribution networks.

This study introduces PicoSleepNet, an ultra-lightweight sleep stage classification method that utilizes a spiking neural network (SNN) with single-channel electroencephalogram (EEG) signals. Traditional methods use multi-bit Nyquist sampling and dense computing, which result in high complexity and power consumption, hindering their deployment on wearable devices. To address these limitations, we propose an innovative pipeline combining single-bit sub-Nyquist level-crossing sampling (LCS) and sparse computing based on SNN. First, LCS adaptively encodes EEG signals into event-driven spike sequences, reducing data volume by 6.98× while preserving essential signal characteristics compared to Nyquist sampling. Second, a sparse recurrent spiking neural network (RSNN) architecture, optimized by the masked backpropagation and sparse regularization (Masked-BPSR) technique, improves performance and reduces computational costs. Third, quantization-aware training (QAT) ensures that the model maintains high accuracy with low-bitwidth quantization, significantly reducing computational power consumption and enabling hardware-friendly deployment. Compared with current state-of-the-art sleep staging approaches, PicoSleepNet achieves competitive performance on three public datasets (Sleep-EDF-20, Sleep-EDF-78, and ISRUC-Sleep) with accuracies of 83.5%, 77.9%, 79.4% and macro-F1 scores of 75.2%, 68.1%, 77.2%, respectively. Meanwhile, by leveraging the computational sparsity design of RSNN and the joint optimization of Masked-BPSR and QAT, PicoSleepNet achieves an ultra-lightweight model with only 14.0-25.8 K parameters (reduced by nearly 2×) and 681.4-842.0 K operations (reduced by 27×), reducing computational power consumption by 1480×. This approach demonstrates the feasibility of deploying ultra-lightweight sleep staging systems in wearable devices and neuromorphic hardware, paving the way for broader applications in real-time health monitoring.

A bionic spiking recurrent neural network with sparse connections and Dale’s principle for image and speech recognition

Traditional deep neural networks exhibit high computational complexity during training and lack biological interpretability due to their reliance on backpropagation-based methods. Spiking Recurrent Neural Network (SRNN) performs well in processing spatio-temporal information by using discrete spike events. It attracts increasing attention in neural computing due to its biological plausibility and hardware implementation. To improve the performance of SRNN, we propose an excitation-inhibition balanced shallow SRNN (EI-SRNN), which is inspired by the balance of excitation and inhibition in the brain, by optimizing the input currents of reservoir neurons to achieve a tight balanced state. The proposed EI-SRNN achieves optimal accuracy while maintaining low computational complexity, debunking the conventional trade-off between accuracy and robustness. We analyze the neural encoding ability and information memory capacity of the EI-SRNN and compare the performance of the model under different degrees of excitation and inhibition. Our experiments demonstrate that EI-SRNN can have higher neural coding capacity and memory capacity under tight balanced excitatory and inhibitory balanced states, so it can achieve better accuracy while possessing stronger robustness. Furthermore, when the reservoir is dominated by excitatory influences, performance declines faster than when the reservoir is dominated by inhibitory influences.

Neuromorphic systems that employ advanced synaptic learning rules, such as the three-factor learning rule, require synaptic devices of increased complexity. Herein, a novel neoHebbian artificial synapse utilizing ReRAM devices has been proposed and experimentally validated to meet this demand. This synapse features two distinct state variables: a neuron coupling weight and an "eligibility trace" that dictates synaptic weight updates. The coupling weight is encoded in the ReRAM conductance, while the "eligibility trace" is encoded in the local temperature of the ReRAM and is modulated by applying voltage pulses to a physically co-located resistive heating element. The utility of the proposed synapse has been investigated using two representative tasks: first, temporal signal classification using Recurrent Spiking Neural Networks (RSNNs) employing the e-prop algorithm, and second, Reinforcement Learning (RL) for path planning tasks in feedforward networks using a modified version of the same learning rule. System-level simulations, accounting for various device and system-level non-idealities, confirm that these synapses offer a robust solution for the fast, compact, and energy-efficient implementation of advanced learning rules in neuromorphic hardware.

Path integration refers to the process by which animals continuously update spatial position during movement by integrating locomotor stimuli such as speed and head direction, thereby enabling real-time localization in continuous space. This process relies on the cooperation of spatial neurons, including grid cells and place cells. In recent years, continuous attractor networks and recurrent neural networks have provided useful insights into the mechanisms of path integration; however, they often rely on fixed-weight connections or exhibit a lack of biological plausibility. To explore more biologically plausible mechanisms, we propose a path integrator based on the recurrent spiking neural network (RSNN). Employing recurrently connected leaky integrate and fire (LIF) neurons, the RSNN encodes spatial positions as spike sequences via membrane potential reset mechanisms, enabling robust long-term path integration. Analysis reveals the spontaneous emergence of spatial and locomotor units, with some units exhibiting spatial-locomotor conjunctive properties, indicating synergistic computations underlying path integration. Ablation experiments confirm that stripe and border units have a larger effect on performance than other unit types under our experimental conditions. Under sparse spiking conditions, the network naturally develops diverse biologically inspired representations. The RSNN's performance provides novel insights into neuronal synergistic mechanisms in biological navigation, offering a biologically grounded framework for path integration modeling and contributing to the development of brain-inspired navigation algorithms.

The neocortex is composed of spiking neurons interconnected in a sparse, recurrent network. Spiking activity within these networks underlies the computations that transform sensory inputs into appropriate behavioral responses. In this study, we train recurrent spiking neural network (SNN) models constrained by neocortical connectivity statistics and investigate the architectural changes that enable task-relevant, spike-based computations. We employ a binary state change detection task—an experimental paradigm used in animal behavioral studies. Our SNNs consist of interconnected excitatory and inhibitory units with connection probabilities and strengths modeled after the mouse neocortex and maintained throughout training and evaluation. Following training, we find that SNNs selectively modulate firing rates based on the binary input state, and that excitatory and inhibitory connectivity within and between input and recurrent layers adjusts accordingly. Notably, inhibitory neurons in the recurrent layer that positively modulate firing rates in response to one input state strengthen their connections to recurrent units with the opposite modulation. This push-pull connectivity—where excitation and inhibition are dynamically balanced in an opponent fashion—emerges as a key computational strategy and is reminiscent of connectivity observed in primary visual cortex. Using a one-hot output encoding yields identical firing rates to both input states, yet the push-pull inhibitory motif still arises. Importantly, this motif fails to emerge when Dale’s principle is not enforced during training, and task performance also declines.Furthermore, disrupting spike timing by a few milliseconds significantly impairs task performance, highlighting the importance of precise spike time coordination for computation in sparse networks like neocortex. The emergence of push-pull inhibition through task training in spiking models underscores the crucial role of interneurons and structured inhibition in shaping neural dynamics and spike-based information processing.

Spiking Neural Networks compute using sparse communication and are attracting increased attention as a more energy-efficient alternative to traditional Artificial Neural Networks. While standard Artificial Neural Networks are stateless, spiking neurons are stateful and hence intrinsically recurrent, making them well-suited for spatio-temporal tasks. However, the duration of this intrinsic memory is limited by synaptic and membrane time constants. Delays are a powerful additional mechanism and, in this paper, we propose an event-based training method for Spiking Neural Networks with delays, grounded in the EventProp formalism, which enables the calculation of exact gradients with respect to weights and delays. Our method supports multiple spikes per neuron and introduces a delay learning algorithm that can, in contrast to previous methods, also be applied to recurrent Spiking Neural Networks. We evaluate our method on a simple sequence detection task, as well as the Yin-Yang, Spiking Heidelberg Digits, Spiking Speech Commands and Braille letter reading datasets, demonstrating that our algorithm can optimise delays from suboptimal initial conditions and enhance classification accuracy compared to architectures without delays. We also find that recurrent delays are particularly beneficial in small networks. Finally, we show that our approach uses less than half the memory of the current state-of-the-art delay-learning method and is up to 26 × faster.

ABSTRACT Electroencephalography (EEG) is a crucial tool for diagnosing neurological disorders like epilepsy. While Artificial Neural Networks (ANNs) have shown strong performance, their large parameter counts and high power consumption limit their practical application. Spiking Neural Networks (SNNs), with their inherent sparsity and parallelism, offer a promising solution; yet most existing SNN models for epilepsy detection are confined to binary classification and fail to fully exploit the rich spatiotemporal dependencies within EEG data. To address these limitations, this study proposes a lightweight Bidirectional Spiking Recurrent Neural Network (Bi‐SRNN) for advanced seizure stage classification. We employ Step‐Forward (SF) encoding to mitigate information loss from high‐frequency EEG oscillations and introduce the Bi‐SRNN architecture, based on the Adaptive Leaky Integrate‐and‐Fire (ALIF) model, to specifically enhance multi‐class classification performance and capture long‐term temporal features. Our model achieved accuracies of 100% and 99.00% in binary and ternary classification tasks on the public Bonn dataset through five‐fold cross‐validation, also achieving strong results on the New Delhi dataset. Furthermore, in transfer learning experiments where the model pre‐trained on the Bonn dataset was applied to new datasets, it demonstrated good generalization performance, also achieving strong results on the New Delhi dataset. With superior performance in both accuracy and model efficiency, the proposed method is well‐suited for deployment on edge devices, offering a more effective tool to assist in clinical diagnosis and treatment.

The rising rate of oral cancers (OC) which is (oral squamous cell carcinoma [OSCC]) more troubling still, the increasing rate of OSCC in younger populations adds significance to the need for valid and consistent prediction methods. More typical research methods for predicting OSCC and cancer in general are usually extremely computationally intensive and have extremely low prediction accuracy. An automated OC classification pipeline that employs deep recurrent spiking neural networks (DRSNN) has been developed. To develop this pipeline, the input was pre-processed by applying a median filter (MF), unsharp masking, and contrast-limited adaptive histogram equalization (CLAHE) to filter some noise and enhance the images. Then features were extracted using the Gabor Wavelet Transform (GWT). After feature extraction, a newly developed Hybrid Bowerbird-Grey Wolf Optimization (HBGWO) was implemented to determine the most representative features. Finally, the images were classified as normal or abnormal using the DRSNN classifier, hence outlining a quicker and efficient prediction methodology for OC detection. The framework was validated using the Kaggle Oral OSCC (lips and tongue images) dataset that was simultaneously augmented to 1,310 samples. According to the experimental data, the suggested method outperforms the F-measure, accuracy, specificity, and precision of the predictive models that have been in use for the past year and achieves an overall classification accuracy of 98.92% when compared favorably to current relevant predictive algorithms, such as ANN, DNN, DCNN-LSTM, and SVM.

Abstract Biologically plausible learning rules for neural networks, such as e-prop (eligibility propagation), are essential both for advancing neuromorphic computing and for understanding fundamental mechanisms of learning in animal brains. However, their behavior under different network conditions remains unclear. Here, we investigate the performance of the e-prop learning algorithm in recurrent spiking neural networks (RSNNs) across different levels of recurrent connectivity and input noise using a complex temporal credit assignment task, a supervised classification problem known to be solvable by rodents. We show that increased sparsity in the recurrent layer significantly enhances learning performance by promoting the generation of more diverse activation patterns. Analysis of the network’s evolution further reveals that the e-prop-trained input layer evolves to route distinct inputs to different regions of the recurrent layer while suppressing the contribution of noise. This partially resembles signal routing functions attributed to the thalamus in mammalian sensory systems, providing additional support for the biological plausibility of e-prop. These findings offer promising insights for efficiency and advantages of biologically inspired training in RSNNs.

Current machine learning systems consume vastly more energy than biological brains. Neuromorphic systems aim to overcome this difference by mimicking the brain’s information coding via discrete voltage spikes. However, it remains unclear how both artificial and natural networks of spiking neurons can learn energy-efficient information processing strategies. Here we propose Predictive Coding Light (PCL), a recurrent hierarchical spiking neural network for unsupervised representation learning. In contrast to previous predictive coding approaches, PCL does not transmit prediction errors to higher processing stages. Instead it suppresses the most predictable spikes and transmits a compressed representation of the input. Using only biologically plausible spike-timing based learning rules, PCL reproduces a wealth of findings on information processing in visual cortex and permits strong performance in downstream classification tasks. Overall, PCL offers a new approach to predictive coding and its implementation in natural and artificial spiking neural networks.

Recurrent spiking neural networks as models of the entorhinal–hippocampal system for path integration: Grid cells and beyond

Benchmarking reinforcement learning and accurate modeling of ground source heat pump systems: Intelligent strategy using spiking recurrent neural network combined with spider WASP inspired optimization algorithm

We present an experimental imaging flow cytometer using a 1 µs temporal resolution event-based complementary metal-oxide semiconductor (CMOS) camera, with data processed by adaptive feedforward and recurrent spiking neural networks. Our study classifies polymethyl methacrylate (PMMA) particles (12, 16, 20 µm) flowing at 0.7 m/s in a microfluidic channel. Processing of experimental data highlighted that spiking recurrent networks, including long short-term memory (LSTM) and gated recurrent unit (GRU) models, achieved 98.4% accuracy by leveraging temporal dependencies. Additionally, adaptation mechanisms in lightweight feedforward spiking networks improved accuracy by 4.3%. This work outlines a technological roadmap for neuromorphic-assisted biomedical applications, enhancing classification performance while maintaining low latency and sparsity.

Training spiking recurrent neural networks (SRNNs) presents significant challenges compared to standard recurrent neural networks (RNNs) that model neural firing rates more directly. Here, we investigate the origins of these difficulties by training networks of spiking neurons and their parameter-matched instantaneous rate-based RNNs on supervised learning tasks. We applied FORCE training to leaky integrate-and-fire spiking networks and their matched rate-based counterparts across various dynamical tasks, keeping the FORCE hyperparameters identical. We found that at slow learning rates, spiking and rate networks behaved similarly: FORCE training identified highly correlated weight matrix solutions, and both network types exhibited overlapping hyperparameter regions for successful convergence. Remarkably, these weight solutions were largely interchangeable-weights trained in the spiking network could be transferred to the rate network and vice versa while preserving correct dynamical decoding. However, at fast learning rates, the correlation between learned solutions dropped sharply, and the solutions were no longer fully interchangeable. Despite this, rate networks still functioned well when their weight matrices were replaced with those learned from spiking networks. Additionally, the two network types exhibited distinct behaviours across different sizes: faster learning improved performance in rate networks but had little effect in spiking networks, aside from increasing instability. Through analytic derivation, we further show that slower learning rates in FORCE effectively act as a low-pass filter on the principal components of the neural bases, selectively stabilizing the dominant correlated components across spiking and rate networks. Our results indicate that some of the difficulties in training spiking networks stem from the inherent spike-time variability in spiking systems-variability that is not present in rate networks. These challenges can be mitigated in FORCE training by selecting appropriately slow learning rates. Moreover, our findings suggest that the decoding solutions learned by FORCE for spiking networks approximate a cross-trial firing rate-based decoding.

Training spiking recurrent neural networks (SRNNs) presents significant challenges compared to standard recurrent neural networks (RNNs) that model neural firing rates more directly. Here, we investigate the origins of these difficulties by training networks of spiking neurons and their parameter-matched instantaneous rate-based RNNs on supervised learning tasks. We applied FORCE training to leaky integrate-and-fire spiking networks and their matched rate-based counterparts across various dynamical tasks, keeping the FORCE hyperparameters identical. We found that at slow learning rates, spiking and rate networks behaved similarly: FORCE training identified highly correlated weight matrix solutions, and both network types exhibited overlapping hyperparameter regions for successful convergence. Remarkably, these weight solutions were largely interchangeable—weights trained in the spiking network could be transferred to the rate network and vice versa while preserving correct dynamical decoding. However, at fast learning rates, the correlation between learned solutions dropped sharply, and the solutions were no longer fully interchangeable. Despite this, rate networks still functioned well when their weight matrices were replaced with those learned from spiking networks. Additionally, the two network types exhibited distinct behaviours across different sizes: faster learning improved performance in rate networks but had little effect in spiking networks, aside from increasing instability. Through analytic derivation, we further show that slower learning rates in FORCE effectively act as a low-pass filter on the principal components of the neural bases, selectively stabilizing the dominant correlated components across spiking and rate networks. Our results indicate that some of the difficulties in training spiking networks stem from the inherent spike-time variability in spiking systems—variability that is not present in rate networks. These challenges can be mitigated in FORCE training by selecting appropriately slow learning rates. Moreover, our findings suggest that the decoding solutions learned by FORCE for spiking networks approximate a cross-trial firing rate-based decoding.

Abstract Working memory serves as a crucial building block in cognitive systems due to its fundamental role in information processing and decision-making. Acting as a temporary storage and manipulation mechanism, working memory enables individuals to actively hold and manipulate relevant information essential for ongoing tasks. This cognitive function is pivotal for various complex processes, including problem-solving, language comprehension, and decision-making. This paper introduces a novel working memory model designed as a spiking recurrent neural network of excitatory and inhibitory neurons. The proposed model incorporates biological mechanisms that allows attractors to be active phasically, thereby reducing the energy budget associated with maintaining attractor states. The core innovation of the model lies in its ability to leverage phasic attractors for state-dependent computation in a probabilistic manner. By modulating the activity of attractors via synaptic delays, the model demonstrates context-sensitive information processing. The results showcase the efficiency gains achieved by the proposed phasic attractor mechanism and highlight the model’s capacity for flexible and adaptive information processing.