Unverified50% confidenceFactExact time
2015年Koch等人提出的孪生网络（Siamese Network）通过学习样本间的相似度度量来实现少样本分类
1
Sources
50%
Confidence
Long-term
Relevance
8/21/2026
First Seen
Sources
Related Claims
UnverifiedThe Siamese Network was originally proposed by Bromley et al. in 1993 for signature verification, with the core idea of measuring the similarity or dissimilarity between two inputs through contrastive learning.72% similarUnverified研究表明DINOv2等自监督模型的中间层特征仅通过简单的线性探针或最近邻匹配就能在多种密集预测基准上取得有竞争力的结果61% similarUnverified相关性检测通常采用基于预训练语言模型（如BERT、Sentence-BERT）的嵌入向量余弦相似度计算61% similarUnverified作者建议的组合策略是用至少5个随机种子分别训练基线和提出方法，报告准确率均值和标准差配合配对检验，并对代表性种子额外运行 McNemar 检验展示样本级预测分歧60% similarUnverified当比较两个具体训练模型的预测行为差异时应固定种子使用同一测试集运行McNemar检验，当评估方法整体优越性时应用多种子重复实验配合配对t检验或非参数检验60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/782263API
curl https://kongchang.com/api/v1/knowledge/claims/782263MCP
get_claim(id=782263)