Unverified50% confidenceOpinionExact time
If the depth and segmentation branches participate in the forward pass during training and are suddenly removed at inference, the gradient distribution the encoder receives changes, and some feature channels that depend on the auxiliary paths may become 'idle,' wasting representational capacity.
1
Sources
50%
Confidence
Long-term
Relevance
11/8/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedDuring cross-domain fine-tuning, an excessively large learning rate will destroy pre-trained weights, while an excessively small one will result in almost no learning, requiring a careful trade-off.77% similarUnverifiedThe causes of the lost-in-the-middle phenomenon are usually attributed to the decay effect of positional encoding, the fact that beginning and end content in training data is more often associated with answers, and the dilution of the attention softmax over long sequences.76% similarUnverifiedThe mechanism of model collapse is distribution drift caused by recursive training, a learning-level problem; whereas the spread of false claims through document copying is content-level contamination, and the two have different intervention methods.75% similarUnverified当目标数据域与源数据域差异悬殊时,预训练权重可能产生负迁移(Negative Transfer),反而干扰目标域的学习75% similarUnverified深度神经网络天然倾向于过拟合训练分布,对分布外样本的处理能力往往大幅下降73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1147512API
curl https://kongchang.com/api/v1/knowledge/claims/1147512MCP
get_claim(id=1147512)