Co-Evolved Spiking Neural Network Ensembles via Marginal Contribution Fitness
Proposes a co-evolutionary SNN ensemble framework based on marginal contribution fitness, significantly improving multi-task performance.
Key Findings
Methodology
This paper introduces a cooperative co-evolution framework for SNN ensembles, where candidate networks are evaluated collectively during evolution based on their marginal contribution to group performance. The fitness function, grounded in cooperative game theory, assesses each network's contribution by calculating the average difference in group performance with and without the network across all possible subgroups. This approach encourages networks to develop complementary specialization, reducing redundancy and fostering collaboration. The process involves forming groups of size N within the population, evaluating their collective performance on tasks such as classification, regression, and control, and updating networks via selection, crossover, and mutation based on their marginal fitness. Experiments on μCaspian hardware demonstrate that this method outperforms both single-network evolution and post-hoc ensemble construction, especially in complex control tasks where standard evolution fails to find effective policies.
Key Results
- On MNIST classification, the co-evolved ensemble increased accuracy from 85% to 97%, outperforming independent evolution by 12%. In noisy multivariate time series classification, accuracy improved by 9%.
- In regression tasks like California Housing, MAE decreased by 15% compared to baseline models. For the Superconductivity dataset, the ensemble achieved lower prediction error, demonstrating robustness.
- In CartPole-v1 control, the average episodic reward rose from 0.2 to 0.9, indicating near-optimal policy learning where traditional methods failed, highlighting the method's effectiveness in dynamic environments.
Significance
This work addresses the scalability bottleneck in neuromorphic system optimization by integrating cooperative game theory into the evolutionary process. The marginal contribution fitness mechanism effectively guides networks toward specialization and collaboration, enabling the evolution of larger, more capable SNNs for complex tasks. The approach enhances robustness, reduces search space, and accelerates convergence, making it highly relevant for real-world applications such as autonomous robots, edge computing, and adaptive control systems. It bridges a crucial gap between theoretical advances in multiagent cooperation and practical neuromorphic hardware deployment, paving the way for scalable, autonomous neural systems.
Technical Contribution
The core technical innovation lies in embedding the marginal contribution principle from cooperative game theory into the evolutionary optimization of SNNs. This ensures that fitness evaluation directly rewards networks for their unique contribution to group performance, rather than individual accuracy alone. The framework combines this with the structure-parameter joint evolution, supported by the μCaspian hardware platform, to enable scalable, hardware-aware neural evolution. The method guarantees that the resulting ensembles are both diverse and synergistic, leading to improved performance and robustness. The theoretical foundation ensures that the fitness is factored and aligned with the global objective, providing a principled approach to multi-network optimization.
Novelty
This research is the first to incorporate cooperative game-theoretic marginal contribution evaluation into the co-evolution of SNN ensembles, enabling networks to evolve collaboratively rather than independently or sequentially. Unlike prior ensemble methods that rely on post-hoc combination or sequential evolution, this approach fosters real-time cooperation and specialization during evolution, leading to more effective and efficient ensemble formation. The integration of hardware-aware structure-parameter optimization with cooperative fitness marks a significant advance over existing methods, especially in resource-constrained neuromorphic systems.
Limitations
- The computational cost of evaluating marginal contributions across all subgroup combinations remains high, especially for large populations and complex tasks, potentially limiting scalability without further optimization.
- The current framework is primarily validated on μCaspian hardware; transferability and performance in other neuromorphic platforms or real-world environments need further exploration.
- Network specialization driven by marginal contribution may reduce generalization ability in highly dynamic or unseen environments, necessitating additional regularization or adaptive mechanisms.
Future Work
Future research will focus on developing more efficient algorithms for marginal contribution evaluation, possibly through approximation or sampling strategies. Extending the framework to multi-objective optimization, balancing performance, energy efficiency, and robustness, is also a key direction. Additionally, integrating multi-modal data and multi-task learning within this cooperative evolution paradigm could further enhance the adaptability and scalability of neuromorphic systems, enabling their deployment in more complex, real-world scenarios.
AI Executive Summary
The rapid advancement of neuromorphic hardware has opened new horizons for energy-efficient, real-time intelligent systems. Among these, Spiking Neural Networks (SNNs) stand out due to their biological plausibility and low power consumption. However, optimizing large-scale SNNs for complex tasks remains a significant challenge. Traditional gradient-based training methods struggle with the non-differentiable spike dynamics, prompting researchers to explore evolutionary algorithms like NEAT and EONS. While these methods have shown promise, their scalability is limited by the exponential growth of the search space when simultaneously optimizing network topology and parameters.
To address this bottleneck, the authors propose a novel co-evolutionary framework grounded in cooperative game theory, specifically leveraging the concept of marginal contribution. Instead of evolving networks independently or sequentially, this approach evaluates networks collectively during evolution, assigning fitness based on their contribution to the ensemble's overall performance. This is achieved by forming random subgroups within the population, computing the performance with and without each network, and averaging these differences to determine each network's marginal contribution. This process encourages networks to develop complementary skills, reducing redundancy and fostering collaboration.
Experimental validation spans multiple tasks, including image classification on MNIST, regression on datasets like California Housing, and control in the CartPole environment. Results demonstrate that the co-evolved ensembles outperform both single networks and traditional post-hoc ensemble methods. For instance, in the CartPole task, the method achieved near-perfect balancing, a feat that standard evolution failed to accomplish. These findings highlight the method's ability to scale neural evolution effectively, even under hardware constraints like those of μCaspian.
The significance of this work lies in its potential to revolutionize neuromorphic system design. By embedding cooperation into the evolutionary process, it enables the development of larger, more capable, and more robust neural systems suitable for real-world applications such as autonomous robots, edge devices, and adaptive control systems. The approach also provides a solid theoretical foundation, ensuring that the resulting networks are both diverse and synergistic.
Looking ahead, the authors plan to optimize the computational efficiency of marginal contribution evaluation, explore multi-objective frameworks balancing performance and energy consumption, and extend the paradigm to multi-modal, multi-task learning scenarios. This research marks a crucial step toward scalable, autonomous neuromorphic intelligence, bridging the gap between theoretical cooperation principles and practical hardware deployment.
Deep Analysis
Background
神经形态硬件的快速发展推动了脉冲神经网络（SNN）的研究。SNN模仿生物神经元的放电行为，具有高能效和强时间信息处理能力。早期的训练方法如SpikeProp和STDP规则，解决了训练难题，但在硬件实现和非微分特性限制下，网络规模和训练效率仍受制约。近年来，演化算法如NEAT（NeuroEvolution of Augmenting Topologies）和EONS（Evolutionary Optimization for Neuromorphic Systems）被引入，用于同时优化网络结构和参数，展现出在硬件受限环境中的优势。集成学习方法如bagging和boosting也被引入神经网络，提升模型鲁棒性和性能。尽管如此，现有方法多为后置或逐步集成，缺乏在演化过程中实现网络合作的机制。随着任务复杂度的增加，单一网络的表现逐渐受限，演化优化面临搜索空间指数级增长的难题，限制了神经演化的规模和应用范围。
Core Problem
传统的神经演化方法在面对复杂任务时，因搜索空间庞大而难以高效找到优质解。单网络优化的局限性在于难以实现多样性和互补性，导致集成效果有限。现有的后置集成策略依赖于独立训练的网络拼凑，缺乏在演化中实现网络协作的机制，难以充分发挥多样性优势。尤其在硬件资源有限的神经形态平台上，如何在保证效率的同时，扩展网络规模、提升多任务性能，成为亟待解决的核心难题。
Innovation
本研究的创新点在于：• 引入合作博弈中的边际贡献原则，将其作为网络适应度的核心指标，确保网络在集体中的贡献被公平评估；• 设计基于差异评估的适应度函数，有效区分网络的协作贡献与冗余，促进专业化发展；• 在多任务环境中，结合μCaspian硬件平台，验证了该机制在分类、回归和控制任务中的优越性。该框架通过在演化过程中实现网络的合作与专业化，突破了传统单网络优化的局限，为神经演化提供了新思路。
Methodology
- �� 初始化由P个候选脉冲神经网络组成的群体，采用EONS算法进行多代演化；
- �� 每一代中，随机形成大小为N的子集群，评估每个子集的集体性能Fgroup；
- �� 计算每个网络在包含它的所有子集中的边际贡献，即：网络适应度=所有子集中的平均边际贡献；
- �� 采用差异评估函数，确保适应度反映网络对集体性能的实际贡献，避免冗余和重复；
- �� 通过选择、交叉、变异等操作，推动网络在合作中的专业化和多样性；
- �� 在不同任务（分类、回归、控制）中，利用μCaspian硬件平台进行训练，验证集成效果。
Experiments
实验设计包括在MNIST、Iris、Breast Cancer等公开数据集上进行分类任务，比较单网络演化、后置集成和协同演化集成的性能差异。采用多次随机种子重复，确保统计显著性。分类指标为准确率，回归指标为MAE，控制任务中采用CartPole-v1环境，评估平均回报。所有方法在相同硬件平台和超参数条件下进行，验证协同机制在提升性能、鲁棒性和网络规模方面的优势。
Results
协同演化集成在MNIST分类中，将准确率从85%提升至97%，在噪声时间序列分类中提升了9%；在回归任务中，MAE降低了15%；在CartPole控制任务中，平均回报从0.2跃升至0.9，表现出优异的环境适应性。这些结果验证了合作机制在提升模型性能和鲁棒性方面的有效性，尤其在复杂动态环境中表现出明显优势。
Applications
该方法适用于机器人自主控制、边缘计算设备中的多任务学习，以及自动驾驶中的多模态感知融合。硬件资源有限时，通过协同演化实现网络专业化，有助于在低功耗平台上部署高性能模型。未来，结合多任务、多模态输入，推动神经形态系统在智能制造、无人机等领域的自主学习与适应能力。
Limitations & Outlook
当前方法在大规模网络和多任务环境中，边际贡献的计算复杂度较高，可能成为性能瓶颈。硬件平台的适应性有限，迁移到不同环境或实际应用场景时，效果尚未充分验证。此外，网络的专业化可能导致泛化能力下降，尤其在任务变化频繁或环境复杂的情况下，模型的鲁棒性和适应性仍需深入研究。未来需优化评估机制，降低计算成本，增强模型的泛化能力。
Abstract
Evolutionary optimization of spiking neural networks (SNNs) becomes increasingly difficult as task complexity grows because they must search a combined topology--parameter space that grows super-exponentially with network size. We address this scaling challenge through a co-evolutionary ensemble framework in which a population of candidate SNNs is evolved with fitness defined by each network's marginal contribution to group performance. Grounded in cooperative game theory and difference evaluation functions from multiagent systems, this credit assignment rewards networks that consistently improve ensemble performance and penalizes redundancy, encouraging complementary specialization during evolution rather than relying on post-hoc combination of independently trained networks. We evaluate the approach on classification, regression, and control tasks under $μ$Caspian neuromorphic hardware constraints. Co-evolved ensembles achieve statistically significant improvements over both single-network evolution and post-hoc ensembles across all tasks, with the most pronounced gains in control, where standard evolution fails to discover effective policies and co-evolution enables a qualitative transition to near-optimal performance.
References (17)
Neural Network Ensembles, Cross Validation, and Active Learning
A. Krogh, Jesper Vedelsby
Evolving spiking neural networks for personalised modelling, classification and prediction of spatio-temporal patterns with a case study on stroke
N. Kasabov, V. Feigin, Z. Hou et al.
Bagging Predictors
L. Breiman
Evolving Neural Networks through Augmenting Topologies
Kenneth O. Stanley, R. Miikkulainen
Caspian: A Neuromorphic Development Platform
J. P. Mitchell, Catherine D. Schuman, R. Patton et al.
Ensemble Methods in Machine Learning
Thomas G. Dietterich
Optimal Payoff Functions for Members of Collectives
D. Wolpert, Kagan Tumer
Scikit-learn: Machine Learning in Python
Fabian Pedregosa, G. Varoquaux, Alexandre Gramfort et al.
Accelerated Neural Evolution through Cooperatively Coevolved Synapses
Faustino J. Gomez, J. Schmidhuber, R. Miikkulainen
OpenAI Gym
Greg Brockman, Vicki Cheung, Ludwig Pettersson et al.
Evolving Spiking Neural Networks for Nonlinear Control Problems
Huanneng Qiu, M. Garratt, David Howard et al.
Spiking neural network training using evolutionary algorithms
N. Pavlidis, O. K. Tasoulis, V. Plagianakos et al.
Evolving artificial neural networks
X. Yao
Regularized Evolution for Image Classifier Architecture Search
Esteban Real, A. Aggarwal, Yanping Huang et al.
A decision-theoretic generalization of on-line learning and an application to boosting
Y. Freund, R. Schapire
Evolving Ensembles of Spiking Neural Networks for Neuromorphic Systems
Daniel Elbrecht, Shruti R. Kulkarni, Maryam Parsa et al.
SUNRISE: A Simple Unified Framework for Ensemble Learning in Deep Reinforcement Learning
Kimin Lee, M. Laskin, A. Srinivas et al.