Researchers have introduced Vision SmolMamba, a novel energy-efficient spiking state-space architecture designed for visual modeling. This architecture integrates spike-driven dynamics with linear-time selective recurrence, utilizing a Spike-Guided Spatio-Temporal Token Pruner (SST-TP) to estimate token importance based on spike activation and latency. By progressively removing redundant tokens, Vision SmolMamba preserves crucial spatio-temporal information, enabling efficient scaling and improved accuracy-efficiency trade-offs. Experiments on various benchmarks show it reduces energy costs by at least 1.5x compared to previous spiking Transformer and Mamba variants. AI
影响 Introduces a more energy-efficient approach to spiking neural networks for vision tasks, potentially reducing computational costs.
排序理由 Academic paper introducing a new model architecture and pruning technique.
- CIFAR10
- CIFAR100
- CIFAR10-DVS
- DVS128 Gesture
- Spiking Transformers
- SST-TP
- Vision SmolMamba
- ImageNet-1K
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
