Researchers have introduced Gated Slot Attention-2 (GSA2), a novel approach to enhance linear attention models by improving their fixed-size recurrent memory. GSA2 combines a new Gated Oja Rule for key-side correction with a Gated Delta Rule for value-side correction, utilizing shared latent slots. This architecture aims to provide effective memory correction and a natural way to operate on both sides of an association. Experiments show that GSA2 outperforms existing linear-attention baselines on various benchmarks while maintaining linear-time sequence modeling and constant-memory recurrent decoding. AI
IMPACT Introduces a new method to improve memory efficiency in linear attention models, potentially leading to more capable sequence modeling.
RANK_REASON The cluster contains a research paper detailing a new model architecture. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Gated Delta Rule-2
- Gated Oja Rule
- Gated Oja Rule-2
- Gated Slot Attention
- Gated Slot Attention-2
- Hugging Face
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →