ImageNet-100
PulseAugur coverage of ImageNet-100 — every cluster mentioning ImageNet-100 across labs, papers, and developer communities, ranked by signal.
3 天有情绪数据
-
研究探讨稀疏分配如何影响剪枝后神经网络的恢复能力
一篇新研究论文调查了神经网络中稀疏分配的分配方式如何影响其在剪枝后恢复精度的能力，尤其是在没有标记的重新训练数据的情况下。该研究比较了ERK和LAMP等不同的稀疏分配方法在各种数据集和架构上的表现，发现分配方式的选择显著影响剪枝后修复的精度。研究人员确定了一个关键的过渡区域，在此区域标准修复方法开始失效，这凸显了联合考虑剪枝分配和修复策略的必要性。
-
新的EIHF方法提升了视觉模型中的OOD检测能力
研究人员开发了一种名为早期高频注入（EIHF）的新方法，以改进计算机视觉模型中的分布外（OOD）检测。EIHF通过在输入数据被第一个卷积层处理之前注入高频信息来实现，而无需改变训练目标。这种方法通过重塑特征几何形状和减少分数重叠，增强了模型区分分布内和分布外数据的能力，尤其是在几何敏感的任务中。在CIFAR-100和ImageNet-100数据集上的实验显示了有希望的结果，包括假阳性率和受试者工作特征曲线下面积的提高。
-
HamJEPA advances JEPAs with Hamiltonian geometry and symplectic prediction
Researchers have introduced HamJEPA, a novel approach to Joint Embedding Predictive Architectures (JEPAs) that moves beyond isotropic regularization. This new method encodes views as phase-space states and uses a learne…
-
New Covariance-Aware Goodness method boosts Forward-Forward learning performance
Researchers have developed a new method called Covariance-Aware Goodness (BiCovG) to improve the performance of the Forward-Forward (FF) learning algorithm, particularly in convolutional neural networks. This approach a…
-
Checkerboard attack offers efficient, learning-free backdoor for deep learning models
Researchers have developed a new method called Checkerboard for launching clean-label backdoor attacks on deep learning models. This learning-free technique uses a closed-form checkerboard trigger derived from linear se…
-
Physics-inspired graph ensembles achieve high accuracy in image classification
Researchers have developed a novel physics-inspired approach for natural image classification, moving away from computationally expensive high-dimensional CNN features. Their method interprets frozen MobileNetV2 feature…