Unverified50% confidenceFactExact time
对于有限宽度网络，NTK在训练中的变化量与1/√m成正比（m为宽度），这一偏差是特征学习发生的机制
1
Sources
50%
Confidence
Long-term
Relevance
8/22/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedNTK theory approximates the training process of infinitely wide neural networks as a kernel method, thereby rigorously proving the global convergence of infinitely wide networks under gradient descent, but cannot capture feature learning phenomena80% similarUnverified监督训练的ResNet-50内在维度相对CLIP更高，说明对比预训练产生了更紧凑的语义流形63% similarUnverifiedResearch such as Microsoft's BitNet shows that if 1-bit quantization-aware training (QAT) is applied from the pre-training stage, accuracy loss can be controlled, whereas post-training quantization (PTQ) typically incurs greater loss.63% similarUnverifiedDauphin et al.'s 2014 study showed that difficulties in neural network training mostly arise from saddle points and plateau regions rather than local minima, with the number of saddle points growing exponentially with dimension n.62% similarUnverified1-bit/1.58-bit要求从训练阶段就以低精度权重为目标，与训练后量化方案本质不同61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/788393API
curl https://kongchang.com/api/v1/knowledge/claims/788393MCP
get_claim(id=788393)