Unverified50% confidenceFactExact time
Collapse模型的状态更新公式为 h ← h − strength · (1 − cos(h, W)) · normalize(h − W)
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
No MLP or Attention Required: How Point-Attractor Dynamics Learn Word Embeddings
redditr/learnmachinelearning7/10/2026
Related Claims
UnverifiedBatchNorm在推理时可以直接折叠进卷积的权重和偏置中，折叠后新权重 W' = γW/√(σ²+ε)，新偏置 b' = γ(b-μ)/√(σ²+ε) + β60% similarUnverifiedCollapse模型中每个token向量同时扮演词表示、点吸引子和几何读出三重角色，无额外解码器或输出矩阵60% similarUnverifiedThe forward diffusion process of a diffusion model can be described using a Markov chain: q(x_t|x_{t-1}) = N(x_t; √(1-β_t)x_{t-1}, β_tI).59% similarUnverified模式坍缩是生成模型领域的经典难题，指模型对同一开放性问题反复给出高度雷同的答案58% similarUnverifiedCollapse/noun-collapse模型完全不使用MLP、Transformer、注意力层或独立输出矩阵，仅用点吸引子动力学系统学习词嵌入54% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/491358API
curl https://kongchang.com/api/v1/knowledge/claims/491358MCP
get_claim(id=491358)