Unverified50% confidenceFactExact time
Linear attention replaces the softmax kernel with a factorizable similarity function φ(q)·φ(k) and uses the associativity of matrix multiplication to rewrite O(L²) sequence modeling into a linear-complexity form.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSAGA supports L1 regularization through the proximal operator technique while preserving the variance-reduction property.68% similarUnverifiedThe forward computation of linear attention can be written in closed-form matrix multiplication, thereby allowing rigorous asymptotic analysis using tools such as bias-variance decomposition and random matrix theory.68% similarUnverifiedRecurrentGemma adopts a Linear Recurrent Unit (LRU) architecture, reducing the computational complexity of attention from quadratic growth with sequence length to linear complexity.65% similarUnverifiedThe paper employs a linear-softmax policy model and derives an exact decomposition that splits parameter updates into semantic components and style components.64% similarUnverifiedBefore clustering, L2 normalization can be applied, or the Spherical K-Means variant using cosine distance as the metric can be adopted to address the magnitude-dominance problem.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130840API
curl https://kongchang.com/api/v1/knowledge/claims/1130840MCP
get_claim(id=1130840)