neurons

Spiking neural networks (SNNs) suffer from limited interpretability and suboptimal energy–accuracy trade-offs due to insufficient understanding of neuron-level dynamical differences and parameter sensitivity. Method: This work systematically investigates the dynamical disparities between leaky integrate-and-fire (LIF) and resonant adaptive firing (RAF) neurons via differential equation modeling, phase-plane analysis, parameter sensitivity scanning, and spike-statistics characterization. Contribution/Results: We quantitatively uncover fundamental distinctions in dynamic response properties, frequency selectivity, and noise robustness. We propose the first interpretable hyperparameter tuning framework tailored to LIF/RAF neurons, explicitly linking input encoding schemes and excitatory–inhibitory population configurations to emergent dynamics. Furthermore, we introduce a lightweight, hardware-friendly parameterization guideline that significantly improves the accuracy–energy trade-off under low-latency constraints. Our framework provides both theoretical foundations and practical design principles for deployable SNNs.

To address the limited temporal representation capability of traditional leaky integrate-and-fire (LIF) neurons—hindering time-sensitive computation—this paper proposes a second-order spiking neuron model grounded in damped-driven pendulum dynamics. By leveraging intrinsic nonlinear oscillatory behavior, the model naturally supports phase coding and periodic spiking, thereby significantly enhancing temporal structure modeling. Integrated with spike-timing-dependent plasticity (STDP), it forms multilayer spiking neural networks, implemented and validated via Brian2 simulations. A lightweight deployment strategy tailored for neuromorphic chips is also devised. Compared to LIF neurons, the proposed model achieves superior biological plausibility, reduced energy consumption, and improved performance on sequence processing and symbolic learning tasks—demonstrating enhanced temporal information encoding and processing. This work establishes a novel, temporally aware neuron primitive for neuromorphic computing.

Traditional biophysical neuron models often fail to accurately reproduce real neural activity due to inadequate characterization of ion channel dynamics or excessive structural simplification. This work proposes a plug-and-play hybrid modeling paradigm that embeds neural ordinary differential equations into conductance-based models, enabling data-driven learning of unknown or misspecified channel dynamics directly from voltage recordings while preserving mechanistic interpretability. By introducing voltage-dependent parameterizations of steady-state and time-constant functions, the method efficiently recovers gating variable dynamics without requiring predefined functional forms and supports single-compartment approximations of complex multi-compartmental models. Experiments demonstrate that the framework accurately fits 2,400 distinct ion channel models from a single recording, exhibits strong generalization under out-of-distribution stimuli and parameter misspecification, and reduces computational cost by an order of magnitude.

Ultra-low-power CMOS analog spiking neurons suffer from unreliable excitability in event-driven neuromorphic computing, particularly under low-stimulus or noise-dominated regimes. Method: We propose an intrinsic excitability criterion independent of external input stimuli—relying solely on membrane potential threshold crossing—and integrate SPICE circuit simulations (using industrial-grade compact transistor models) with nonlinear dynamical modeling to quantitatively characterize excitability. Contribution/Results: Our framework establishes a precise, quantitative excitability decision rule, elucidates the parametric influence of key circuit elements—such as leakage conductance, capacitance, and threshold voltage—on action potential generation, and systematically uncovers how intrinsic thermal and shot noise modulate neuronal dynamics. This work provides a general theoretical foundation and practical design guidelines for robust, energy-efficient neuron implementations in brain-inspired chips.

Addressing the challenge of simultaneously achieving biological plausibility and scalability in neuromorphic hardware, this paper proposes a function-driven, minimalist resonant spiking neuron circuit design paradigm. We first discover that the persistent sodium current (I_Na,p) exhibits an N-shaped negative differential resistance (NDR) characteristic, enabling the construction of three NDR-based resonant neuron circuits. Integrating simplified I_Na,p + I_K dynamics with analog VLSI implementation, we design eleven minimal spiking neuron circuits—including three resonant variants. Compared to the Hodgkin–Huxley model, the proposed circuits reduce area and power consumption by over 70%, significantly enhancing hardware efficiency while preserving essential biophysical behaviors such as subthreshold resonance and spiking regularity. This work establishes a novel pathway toward low-overhead, interpretable neuromorphic chips.

This work addresses a critical limitation in existing neuron-level concept explanation methods, which often assume that all neurons possess clear functional roles, thereby overlooking redundant or misleading neurons that can distort interpretations of model decision-making. To overcome this, the authors propose the Select-Hypothesize-Verify (SHV) framework: it first selects the most representative samples based on activation distributions, then generates natural language concept hypotheses, and finally validates these hypotheses through a neuron activation verification mechanism. SHV introduces, for the first time, a systematic pipeline for concept validation, effectively identifying and focusing on neurons with genuine semantic meaning. Experimental results demonstrate that concepts produced by SHV activate target neurons at 1.5 times the rate of state-of-the-art methods, substantially improving the accuracy and reliability of model interpretations.

本文探讨了神经形态工程的核心原则，提出应基于生物系统中的计算原理来设计人工机器，而非简单模仿生物组件。

This study addresses the lack of verifiable ground truth in neuroscience for evaluating mechanistic models by introducing a whole-brain neuromechanical simulation platform based on larval zebrafish, which provides a transparent and controllable benchmark environment. The work innovatively integrates large language model (LLM)-guided symbolic regression with neural architectural priors to automatically discover predictive and interpretable models of neural mechanisms through tree search. The resulting models significantly outperform conventional baselines, demonstrate strong out-of-distribution generalization, and accurately recapitulate key functional architectures observed in real neural circuits.

This study investigates the mechanisms by which neural networks sustain self-sustained activity in the absence of continuous external input. The authors construct a recurrent network of 200 Hodgkin–Huxley neurons incorporating multiple biologically plausible features, including spike-timing-dependent plasticity (STDP), probabilistic vesicular release, random synaptic connectivity, receptor heterogeneity, and voltage-dependent inhibition. Following a brief initial stimulus, the network maintains sparse, irregular spiking at an average rate of 1.13 Hz for up to 1800 seconds, with 67% of neurons firing below 1 Hz and the population Fano factor stably ranging between 1 and 2. This work represents the first demonstration of long-lasting self-sustained activity in a fully spiking, biologically realistic model and reveals spontaneous reorganization of collective firing patterns over time.

This study addresses the challenge of effectively modeling the influence of inhibitory and excitatory connections on neuronal spiking dynamics within the framework of transmission neural networks. To this end, it introduces—in the first such incorporation—both inhibitory synapses and populations of neurotransmitters into this paradigm, yielding a stochastic network model that features binary neuronal states coupled with dynamic transmission mechanisms. The authors derive an analytical expression for neuronal firing probability and demonstrate through theoretical analysis that the model is equivalent to a two-dimensional continuous-state system. In the limit as the number of neurotransmitters tends to infinity, they formulate a limiting model and establish sufficient conditions for its stability and contractivity. This work thus provides novel modeling and analytical tools for understanding the role of inhibition in spiking neural networks.