self-supervised learning
PulseAugur coverage of self-supervised learning — every cluster mentioning self-supervised learning across labs, papers, and developer communities, ranked by signal.
1 天有情绪数据
-
CurvSSL框架通过流形几何增强自监督学习
研究人员推出了一种新颖的自监督学习框架CurvSSL，该框架将局部流形几何纳入其训练过程。该方法通过添加基于曲率的正则化器来增强标准的SSL技术，该正则化器可以对不同数据增强之间的局部流形弯曲进行对齐和去相关。在MNIST和CIFAR-10数据集上的实验表明，与Barlow Twins和VICReg等现有方法相比，CurvSSL在线性评估中取得了具有竞争力或更优的性能，这表明显式建模局部几何是统计SSL的一个有价值的补充。
-
Massive FOMO260K dataset released to boost AI in brain MRI analysis
Researchers have introduced FOMO260K, a substantial dataset comprising over 260,000 3D brain MRI scans. This dataset is designed to facilitate the advancement of self-supervised learning techniques within the field of m…
-
自监督网络在可比准确率下产生更少的线性区域
一项发表在arXiv上的新研究调查了自监督深度ReLU网络中线性区域的复杂性。研究人员发现，自监督学习方法在达到相似准确率的情况下，与监督方法相比产生的线性区域更少。研究还观察到，对比学习方法会随着时间的推移扩展这些区域，而自蒸馏方法会合并它们，并且这些几何特性可以指示表征质量并检测模型崩溃的早期迹象。
-
ArmSSL framework offers robust black-box watermarking for self-supervised learning encoders
Researchers have introduced ArmSSL, a novel framework designed to protect intellectual property in self-supervised learning (SSL) encoders. This method enables ownership verification even when the stolen encoders are ac…