Unverified50% confidenceFactExact time
BERT分类头线性层的输出维度严格等于类别数，一旦num_classes固定，模型输出空间被锁死，无法在不修改网络结构的前提下直接预测新增类别
1
Sources
50%
Confidence
Long-term
Relevance
7/13/2026
First Seen
Sources
BERT新增分类类别：增量微调的三种策略与灾难性遗忘解法
Related Claims
Unverified分类头的维度固定，训练好的10分类模型输出维度是10，无法在不修改网络结构的前提下直接预测新增类别70% similarUnverified自回归大模型推理瓶颈并非计算量而是内存带宽，推测解码通过拒绝采样保证输出分布与原始大模型完全一致65% similarUnverified词表大小直接决定模型嵌入层和输出层的维度，分词器变更意味着模型无法简单从前代继承权重64% similarUnverifiedGrammar-constrained decoding does not alter the model's weights or training process; it only acts as a safety net during the inference stage, and the relative probability ranking among legal tokens remains unchanged.63% similarUnverified大模型中的参数在数学上是神经网络每条连接边上的权重数值和偏置值，训练前随机初始化，通过反向传播算法和梯度下降不断调整61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/502635API
curl https://kongchang.com/api/v1/knowledge/claims/502635MCP
get_claim(id=502635)