The Sparse-Activation-ReLU (SAR) layer is proposed, a single-step alternative that promotes activation sparsity without surrogate-gradient training while remaining compatible with event-based computing and is a step towards energy-efficient virtual sensing.
Abstract
Virtual sensing enables digital twins and safety-critical systems to reconstruct and forecast spatial-temporal physics in real time. However, conventional computational and data-driven methods often face challenges in generalization, latency, and energy efficiency for edge deployment. Neural operators offer a promising alternative but remain reliant on power-intensive hardware. Spiking neurons and neuromorphic computing can improve efficiency, yet surrogate-gradient training and multi-step spiking introduce convergence and latency challenges. We propose the Sparse-Activation-ReLU (SAR) layer, a single-step alternative that promotes activation sparsity without surrogate-gradient training while remaining compatible with event-based computing. Within a trunk-based NOMAD architecture, SAR achieves over a fivefold improvement in the combined Latency-Error-Energy (LEE) metric compared with Variable Spiking Neuron (VSN) and Leaky Integrate-and-Fire (LIF) implementations. We further analyze spiking entropy and feature usage and introduce synthetic knowledge distillation, reducing the LEE score by more than twofold. Finally, we improve VSN through a ReLU-based spiking loss and graph-neighbor thresholding. On the Heat Exchanger dataset, these approaches reduce L2 error by more than twofold and nearly sevenfold, respectively, while reducing spiking and spatial aggregation. Overall, the work presented is a step towards energy-efficient virtual sensing by providing an alternative framework that can be positioned towards neuromorphic or other edge device integration that can be a gold standard to compare latency, energy, and error performance for future efficient designs that are sparsity or brain-inspired spiking based.
Spiking neural networks (SNNs) offer an energy-efficient alternative to conventional deep neural networks by exploiting sparse event-driven computation, but their training remains challenging because the non-differentiable spike function requires surrogate gradients whose fixed shape may be suboptimal across layers and training stages. In this work, we introduce SAGE, an uncertainty-modulated surrogate-gradient mechanism for Transformer-based SNNs. SAGE estimates block-level uncertainty from normalized self-attention entropy and uses this signal to adapt the surrogate-gradient slope during training while leaving the inference model unchanged. By modulating only the training-time surrogate parameter, the proposed method preserves the original architecture and deployment cost while improving optimization flexibility. Experiments on CIFAR-10/100 demonstrate that SAGE achieves improved accuracy over fixed-surrogate baselines, with results up to 1-2\% consistent gains across multiple simulation time steps. These results highlight the potential of attention-derived uncertainty as a lightweight training signal for adaptive surrogate-gradient learning in transformer-based SNNs.
Extensive experiments on static and neuromorphic benchmarks show that lower-bit BASC models match or outperform higher-bit baselines and retain this accuracy advantage after structured pruning, while further reducing model storage and synaptic operations.
Linliang Chen, Yan Zhong, Xin Liu et al.· 0 citations
Low-latency decision-making is a critical requirement for autonomous navigation in dynamic and resource-constrained environments. Conventional deep learning-based navigation systems often suffer from high computational overhead and energy consumption, limiting their deployment in real-time robotic applications. This paper presents a neuromorphic navigation framework based on spiking neural networks (SNNs) that leverages event-driven computation for efficient perception and control. The proposed system integrates biologically inspired neuron models with latency-aware learning mechanisms to enable rapid sensory processing and decision-making. By exploiting temporal information encoded in spike trains, the framework achieves faster response times while maintaining robust navigation performance. Experimental evaluations demonstrate that the neuromorphic SNN-based approach significantly reduces inference latency and energy consumption compared to conventional neural network baselines, making it suitable for real-time autonomous navigation tasks. The results highlight the potential of neuromorphic computing as a scalable and energy-efficient solution for next-generation autonomous systems.
Arvind Kumar Singh, Lakshmi Narayanan· International Journal of Int...· 0 citations
This work analyzes flaws of conventional conversion pipelines from residual membrane potential statistics and proposes a novel conversion strategy combining dynamic initial potential tuning and feature enhancement, which generalizes to ReLU CNNs, ANN Transformers, and multi-threshold SNN variants.
Zirui Chen, Zihan Huang, Tong Bu et al.· 0 citations
This work proposes heterogeneous neural networks that combine spiking neural networks (SNNs) and artificial neural networks (ANNs) at bandwidth-limited regions, such as chip boundaries, where spike-based communication reduces data transfer overhead.
Joshua Nardone, Rui-Jie Zhu, Ruhai Lin et al.· International Conference on...· 0 citations
While neuromorphic systems offer a promising path for processing dynamic, event-based data, current benchmarks often fail to isolate the specific impact of temporal integration on model performance. To address this, our research rigorously investigates how neuromorphic architectures encode and integrate temporal information by conducting a comprehensive ablation study using a hybrid network. We systematically transition from a fully spatial Convolutional Neural Network (CNN) to a fully Spiking Neural Network (SNN) by progressively replacing ReLU activations with spiking neurons across nine distinct model configurations. To challenge these architectures, we introduce the Temporal-ASL Dataset, a neuromorphic benchmark specifically curated with signs that exhibit high spatial isomorphism but distinct temporal signatures. This approach allows us to decouple spatial features from motion dynamics and quantify the marginal contribution of spiking membrane dynamics in resolving ambiguities that remain invisible to frame-isolated models. Our analysis reveals a performance hierarchy that peaks at 72.5% with a hybrid SNN configuration, representing a 6.25% improvement over the spatial CNN baseline. Logit trajectory analysis confirms this boost stems from the spiking layers’ ability to disambiguate spatially similar signs. However, accuracy declines steadily in deeper hierarchies, falling to 23.75% for the fully spiking SNN. Ultimately, these findings demonstrate that shallow neuromorphic integration effectively maximizes the gains from temporal integration while mitigating the information loss inherent in binary spike quantization.
J. Seekings, Peyton S. Chandarana, Arshia Eslami et al.· International Conference on...· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduSep 2, 2026
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
MIT News · Artificial Intelligence· news.mit.eduAug 27, 2026
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
