研究人员发现了一种称为 Fisher Rank Inflation 的现象，它作为一种谱信号，表明深度学习模型何时从学习真实模式过渡到记忆有噪声或损坏的数据。这种过渡以每个示例的最后一层梯度的有效秩暂时扩张然后收缩为标志。研究表明，这种秩膨胀与标签噪声的严重程度直接相关，甚至可能先于测试性能的可观察到的下降，从而提供了一种检测和理解模型记忆的新方法。 AI
影响 提供了一种新的谱信号来检测和理解模型对带噪声数据的记忆，有可能提高模型的鲁棒性和可靠性。
排序理由 该集群包含一篇详细介绍机器学习新研究发现的学术论文。[lever_c_demoted from research: ic=1 ai=1.0]
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
arXiv:2607.12438v1 Announce Type: cross Abstract: Deep networks trained with label noise often learn clean structure before memorizing corrupted labels. We show that this transition leaves a spectral signature in the centered scatter of per-example last-layer gradients. Its effec…
Deep networks trained with label noise often learn clean structure before memorizing corrupted labels. We show that this transition leaves a spectral signature in the centered scatter of per-example last-layer gradients. Its effective rank transiently expands during memorization …