Scaling Spike-driven Transformer with Efficient Spike Firing Approximation Training

IEEE Transactions on Pattern Analysis and Machine Intelligence, Год журнала: 2025, Номер unknown, С. 1 - 18

Опубликована: Янв. 1, 2025

The ambition of brain-inspired Spiking Neural Networks (SNNs) is to become a low-power alternative traditional Artificial (ANNs).This work addresses two major challenges in realizing this vision: the performance gap between SNNs and ANNs, high training costs SNNs.We identify intrinsic flaws spiking neurons caused by binary firing mechanisms propose Spike Firing Approximation (SFA) method using integer spike-driven inference.This optimizes spike pattern neurons, enhancing efficient training, reducing power consumption, improving performance, enabling easier scaling, better utilizing neuromorphic chips.We also develop an Transformer architecture spike-masked autoencoder prevent degradation during SNN scaling.On ImageNet-1k, we achieve state-of-the-art top-1 accuracy 78.5%, 79.8%, 84.0%, 86.2% with models containing 10M, 19M, 83M, 173M parameters, respectively.For instance, 10M model outperforms best existing 7.2% on ImageNet, time acceleration inference energy efficiency improved 4.5× 3.9×, respectively.We validate effectiveness proposed across various tasks, including object detection, semantic segmentation, vision tasks.This enables match ANN while maintaining advantage, marking significant step towards as general visual backbone.Code available at Spike-driven V3.

Язык: Английский

---

Msfi: Multi-Timescale Spatio-Temporal Features Integration in Spiking Neural Networks

Опубликована: Янв. 1, 2025

Язык: Английский

---

Efficient Spiking Neural Network for RGB–Event Fusion-Based Object Detection

Electronics, Год журнала: 2025, Номер 14(6), С. 1105 - 1105

Опубликована: Март 11, 2025

Robust object detection in challenging scenarios remains a critical challenge for autonomous driving systems. Inspired by human visual perception, integrating the complementary modalities of RGB frames and event streams presents promising approach to achieving robust detection. However, existing multimodal detectors achieve superior performance at cost significant computational power consumption. To address this challenge, we propose novel spiking RGB–event fusion-based network (SFDNet), fully detector capable both low-power high-performance Specifically, first introduce Leaky Integrate-and-Multi-Fire (LIMF) neuron model, which combines soft hard reset mechanisms enhance feature representation SNNs. We then develop multi-scale hierarchical residual attention lightweight aggregation module efficient dual-modality extraction fusion. Experimental results on two public datasets demonstrate that our SFDNet achieves state-of-the-art with remarkably low The scenarios, such as motion blur low-light conditions, highlights robustness effectiveness SFDNet, significantly advancing applicability SNNs real-world tasks.

Язык: Английский

---

Denoising autoencoder multilayer perceptron spiking neural network for isonicotinic acid yield prediction on real industrial dataset

Advanced Engineering Informatics, Год журнала: 2025, Номер 65, С. 103273 - 103273

Опубликована: Апрель 3, 2025

Язык: Английский

---