Abstract Passive Acoustic Monitoring (PAM) is widely used for marine mammal observation, but conventional monitoring systems typically require continuous high-power signal acquisition and processing, limiting long-term autonomous deployment. In this work, we present an Always-On Sperm-Whale Detector (AO-SWD) based on a neuromorphic and ultra-low-power processing pipeline combining a bio-inspired artificial cochlea with FPGA-based spiking neural network (SNN) inference. The proposed system uses a 16-channel ultra-low-power (ULP) cochlea implemented in 65 nm CMOS technology operating in the subthreshold regime. The cochlea converts underwater acoustic signals directly into asynchronous spike trains while consuming only 4 μW total power (250 nW/channel). The generated spike-based representation is processed using lightweight feedforward SNN architectures trained with snnTorch and optimized for FPGA deployment using the ModNEF framework. Several network architectures and temporal resolutions were evaluated to study the trade-off between detection accuracy, latency, and hardware complexity. The best software configuration, using a network architecture of 16 input channels, 128 hidden neurons, and 1 output neuron with 20 time steps, achieved 93.30% classification accuracy for sperm whale click detection. The corresponding FPGA implementation achieved 91.15% accuracy after hardware-aware optimization while maintaining real-time operation. The proposed architecture demonstrates the feasibility of combining ultra-low-power neuromorphic sensing with embedded spiking inference for energy-efficient underwater acoustic monitoring. These results highlight the potential of event-driven bio-inspired systems for long-duration autonomous ecological monitoring applications.
Event cameras provide sparse, asynchronous measurements triggered by changes in scene brightness, offering high temporal resolution but also making learning-based optical flow sensitive to illumination-induced noise. We propose pixel-wise state node (SN) encoding, a biologically inspired state-space representation in which each pixel maintains a persistent state that evolves via leaky temporal integration with continuous-time dynamics. By embedding temporal information directly into state evolution, SN introduces a strong inductive bias that yields compact and stable spatio-temporal features while naturally suppressing spurious events arising from sensor noise. Integrated into RAFT-style optical flow models, SN encoding consistently improves over time surfaces on MVSEC and DSEC-Flow, and achieves competitive performance relative to voxel-grid-based approaches while avoiding explicit temporal binning and additional temporal aggregation mechanisms required to process multiple bins. As a result, SN improves accuracy and robustness under event noise and offers a promising pathway toward low-latency, near-sensor edge-device implementations.
Drug-resistant epilepsy affects millions of people worldwide and calls for real-time, low-power seizure detection to enable effective responsive neurostimulation (RNS). This work presents a multichannel seizure detection system based on reservoir computing implemented with two-dimensional MoS _2 charge-trap memory (CTM) devices. A low-complexity decision-tree classifier supports real-time operation while minimising computational overhead. The proposed architecture processes multiple intracranial electroencephalography channels in parallel, a capability that remains uncommon in neuromorphic seizure detection systems. The system achieves competitive detection accuracy, sensitivity, and early detection performance, evaluated with respect to both electrographic and clinical seizure onset. In addition, the architecture offers the potential for ultra-low-power implementation. These results highlight the potential of 2D material–based neuromorphic hardware for next-generation implantable RNS systems.
While Spiking Neural Networks (SNNs) offer a promising path toward energy-efficient edge intelligence, conventional Binary representations often suffer from high inference latency and discretization errors. Multi-level (graded) spike models address these limitations by increasing information density per pulse, yet they typically necessitate power-hungry Multiply-Accumulate units that nullify the energy advantages of neuromorphic hardware. In this work, we propose a hardware-algorithm co-design framework that leverages multi-level Integrate-and-Fire neurons coupled with a specialized multiplier-less architecture. By decomposing multi-level interactions into a cascaded shift-and-add logic, our hardware preserves the event-driven sparsity of SNNs while supporting high-precision activations. We evaluate our proposed accelerator on the Xilinx Zynq MPSoC ZCU102 platform using the Google Speech Commands (GSCs), CIFAR-10, and DVS-Gesture datasets. Experimental results demonstrate that the graded spike paradigm achieves up to a $6.7\times$ reduction in total spike activity through temporal information compression. Compared with the binary baseline benchmark on the DVS Gesture dataset, the proposed architecture reduces memory accesses by 68.59%, improves steady-state throughput from 33.08 FPS to 251 FPS, and increases energy efficiency from 9.01 to 54.96 GOPS W $^ {-1}$ . Furthermore, by minimizing pipeline stalls and workload imbalance, our design achieves an Efficiency-per-Processing-Element of up to 3.05 GOPS W $^ {-1}$ PE $^ {-1}$ , thereby providing a competitive PE-level energy efficiency compared with representative SNN accelerators. This research provides a scalable pathway for deploying high-performance, ‘always-on’ neuromorphic systems in resource-constrained environments.
Spiking neural networks (SNNs) offer a highly energy-efficient, neuromorphic alternative to standard deep learning, but their real-world deployment necessitates trustworthy and explainable outputs. Class Activation Mapping (CAM) methods are a robust and popular way of implementing this explainability in continuous artificial neural networks, but translating these methods to the discrete, temporal domain of SNNs introduces distinct theoretical failure modes. In this work, we formalize these limitations, establishing a fundamental Sensitivity-Faithfulness Tradeoff in gradient-based attributions and expose systematic binary threshold effects in perturbation-based approaches. Building on this, we identify a severe domain mismatch when applying continuous attribution smoothing to sparse neuromorphic data. To correct the resulting out-of-distribution hallucinations, we introduce Poisson SmoothGrad , a physically-consistent noise injection method that safely marginalizes threshold volatility to yield the strongest gradient-based attributions we measure on native event data. Benchmarking six attribution methods on a spiking VGG-11 across CIFAR-10, CIFAR-10-DVS and DVSGesture, we find that no single method dominates: domain-native smoothing leads the gradient family on sparse event streams, while the activation-based spike activation map is the strongest overall. Ultimately, we measure the network-level dynamics that allow CAMs to keep working despite these neuron-level mathematical flaws, and synthesize our empirical benchmarks into a Practitioner’s Guide that provides actionable heuristics for deploying robust, domain-aware visual explanations in SNNs.
