Unverified50% confidenceBenchmarkExact time
Towards Monosemanticity研究在单层MLP上训练SAE，从约512维激活空间中提炼出超过4000个可解释特征
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
Unverified乘积量化的核心思想是将高维向量空间分解为若干低维子空间，在每个子空间内独立进行码本学习与编码65% similarUnverifiedDeepSeek-V2引入Multi-Head Latent Attention（MLA）机制和DeepSeekMoE架构，其每百万token推理成本约为同期GPT-4级别模型的百分之一61% similarUnverified乘积量化（PQ）最早由Jégou等人于2011年提出，其核心思想是将高维向量空间分解为若干低维子空间并在每个子空间内独立进行码本学习与编码60% similarVerifiedCommon image batch data in deep learning is a 4-dimensional tensor with dimensions of batch size × number of channels × height × width.60% similarUnverifiedGoogle DeepMind released Gemma 4 quantization-aware training weights, reducing the memory footprint of the 12B parameter model to approximately 1GB through targeted 2-bit compression technology.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/533683API
curl https://kongchang.com/api/v1/knowledge/claims/533683MCP
get_claim(id=533683)