New sLoTh framework enables energy-efficient continual learning for sparse vision transformers

Researchers have introduced sLoTh, a novel framework designed for parameter-efficient continual learning in sparse event-based vision transformers. This approach freezes the backbone of the model and focuses plasticity on low-rank attention updates and neuronal threshold modulation, requiring updates to less than 1% of parameters. Experiments on various datasets like CIFAR-100 and ImageNet variants demonstrated competitive performance in class-incremental and online continual learning scenarios without the need for replay buffers. Notably, sLoTh achieved approximately 6.5 times lower energy consumption compared to traditional dense vision transformers. AI

IMPACT This research offers a more energy-efficient approach to continual learning for vision transformers, potentially enabling wider deployment in resource-constrained robotic and edge intelligence systems.

RANK_REASON The cluster describes a new research paper published on arXiv detailing a novel framework for continual learning in computer vision. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• CIFAR-100
• Hugging Face
• ImageNet-100
• ImageNet-R
• seLoRA
• sLoTh
• Tiny-ImageNet
• Vaishnavi Nagabhushana

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →