online synaptic plasticity

This work addresses the challenge of enabling biologically plausible neural networks to perform long-horizon, structured credit assignment under sparse, delayed rewards. Methodologically, it introduces a biologically inspired meta-learning framework that—uniquely—embeds tangent propagation into the inner loop of meta-learning to automatically discover three-factor synaptic plasticity rules. These rules depend solely on local pre- and postsynaptic activity plus a global, slowly varying neuromodulatory signal, thereby achieving temporal scale separation. Critically, they require neither backpropagation-through-time nor non-local information, strictly adhering to neurobiological constraints. Empirically, the learned plasticity rule matches the performance of BPTT on multi-task sequential decision-making benchmarks while substantially outperforming existing local learning approaches. This work establishes a novel paradigm bridging synaptic plasticity mechanisms with credit assignment in reinforcement learning, offering a principled, biologically grounded alternative to gradient-based optimization in recurrent networks.

This work addresses the limitation of mainstream neural network models—reliance on discrete parameter updates and strict separation between inference and learning phases—by proposing a continuous-time neural dynamics framework that unifies biologically plausible learning mechanisms. Methodologically, it employs stochastic differential equations to model neural dynamics, introduces a second-scale synaptic plasticity time window, and establishes temporal overlap between input and error signals as a necessary condition for error-driven learning; stability under noise is ensured via feedback alignment and direct feedback alignment. Key contributions include: (i) the first theoretical and simulation-based demonstration that functional facilitation traces operate on a second-scale timescale; (ii) seamless coupling of inference and learning within continuous time; and (iii) robust learning performance despite temporal misalignment and integration noise, yielding significant improvements in deep network accuracy.

Traditional neural networks rely on biologically implausible global backpropagation and inadequately model neuroplasticity mechanisms. To address this, we propose SAPIN—a novel spiking neural architecture that integrates structural plasticity with active inference for the first time. SAPIN jointly optimizes synaptic weights and neuronal spatial positions on a 2D grid, driven solely by local prediction errors: these errors govern Hebbian-like synaptic updates and activity-dependent neuronal migration, enabling co-evolution of network topology and connection strength. The resulting architecture implements a self-stabilizing, predictive coding–oriented computation grounded in error minimization. Evaluated on the CartPole task, SAPIN achieves an 82% average success rate across 100 trials using fixed parameters, demonstrating robust reinforcement learning capability and policy stability. Our core contribution is a biologically plausible, computationally efficient paradigm for dynamic, synergistic optimization—bridging neuroscience realism with machine learning efficacy.

To address the challenge of applying backpropagation-based reinforcement learning (RL) in resource-constrained and non-differentiable neural network settings, this paper proposes a gradient-free, noise-driven RL method. The approach approximates directional derivatives via stochastic neurons and couples reward prediction errors with eligibility traces to enable purely local, biologically plausible temporal credit assignment. It is the first work to integrate directional derivative theory into reward-modulated Hebbian learning (RMHL), eliminating reliance on global error signals and differentiability assumptions. Empirically, the method significantly outperforms conventional RMHL on standard RL benchmarks and matches the performance of backpropagation-based baselines, while maintaining full compatibility with neuromorphic hardware. This establishes a viable, energy-efficient learning paradigm for edge intelligence applications.

To address the rigidity of controller architectures and reliance on global error signals in deep reinforcement learning, this paper proposes an online adaptive neural network growth framework. The method employs local structural plasticity modules (SPMs), which drive topology evolution—neuron insertion and pruning—solely via local neuronal activation and gradient statistics, eliminating manual hyperparameter tuning. It supports both artificial neural networks (ANNs) and spiking neural networks (SNNs), establishing a novel brain-inspired paradigm for plasticity-based control. On standard control benchmarks, the approach achieves comparable or superior cumulative rewards while significantly reducing policy variance and automatically converging to compact, task-adapted architectures. Ablation studies confirm that structural plasticity is critical for reward stability and robust policy learning.

This work addresses the poor scalability and low sample efficiency of traditional neuroevolution in high-dimensional weight spaces and complex continuous control tasks by proposing NEOL, a dual-timescale neuroevolution framework. In NEOL, an outer loop employs NEAT to evolve network architectures, while an inner loop incorporates reward-modulated online synaptic plasticity to dynamically adjust synaptic weights. The framework provides, for the first time, a sublinear regret bound that theoretically guarantees convergence and superior sample efficiency through the integration of neuroevolution with online learning. Empirical results demonstrate that NEOL significantly outperforms standard NEAT across four benchmark continuous control tasks, achieving higher final fitness and lower variance, and matches the performance of state-of-the-art reinforcement learning algorithms.

This work proposes NeuroPlastic, a novel optimizer that addresses the limitations of conventional deep learning optimizers, which rely solely on local gradients and lack inspiration from biological neural plasticity, often underperforming under data scarcity or noisy conditions. NeuroPlastic is the first to incorporate multi-factor synaptic plasticity mechanisms from neuroscience into the optimization process, integrating gradient, activity-like, and memory-like signals through a lightweight adaptive modulation layer. This layer dynamically adjusts parameter updates without altering standard training pipelines. Compatible with mainstream deep learning frameworks, NeuroPlastic consistently outperforms baseline optimizers on Fashion-MNIST and few-shot image classification tasks and demonstrates strong generalization in CIFAR-10 transfer experiments—all without requiring re-tuning of hyperparameters.

This work addresses the long-standing challenges in spiking neural network (SNN) training research—namely, the lack of a systematic taxonomy and unified evaluation protocols, which have led to poor reproducibility and fragmented progress. To this end, we propose the first fine-grained, systematic classification framework for SNN training algorithms and introduce NeuroTrain, an open-source benchmark built upon snnTorch. NeuroTrain enables, for the first time, modular integration and fair comparison across diverse algorithmic paradigms, including surrogate gradient backpropagation, local and three-factor learning rules, biologically plausible plasticity mechanisms, ANN-to-SNN conversion methods, and unconventional optimization strategies. The framework standardizes the implementation of representative algorithms and supports consistent evaluation across datasets, architectures, and training configurations, thereby significantly enhancing the reproducibility and systematic investigation of SNN training methodologies.

This work addresses the high energy consumption, substantial memory overhead, and limited biological plausibility of conventional artificial neural networks in continual learning, which stem from frequent parameter updates. Inspired by human negativity bias and error-related negativity (ERN) signals observed in neuroscience, the authors propose a novel mechanism termed "memorized error-gated learning," wherein synaptic updates are triggered exclusively upon current or historical classification errors. This approach introduces error gating into continual learning for the first time, requiring no additional hyperparameters. It substantially reduces parameter update frequency by 50%–80% while maintaining competitive model performance, thereby significantly lowering both computational energy demands and reliance on replay buffers. The method exhibits strong biological interpretability and engineering practicality, making it well-suited for both incremental and online continual learning scenarios.

This work proposes a biologically plausible learning framework based on a dual-stream excitatory/inhibitory neural network that strictly adheres to Dale’s law—where each neuron’s synapses are exclusively excitatory or inhibitory—and achieves credit assignment without weight transposition through an error diffusion mechanism. The method incorporates modular error routing for multi-class classification, combined with layer-specific Sigmoid widths, batch-centered class-wise error signals, and asymmetric initialization to effectively alleviate credit assignment bottlenecks across diverse tasks. In supervised learning, the model attains 96.7% and 61.7% accuracy on MNIST and CIFAR-10, respectively. Furthermore, the proposed ED-PPO algorithm demonstrates competitive performance against current backpropagation-free baselines in Brax and Craftax reinforcement learning environments, confirming its scalability and generality.