arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1705.03071cs.LG
收藏

深度学习中的优化几何和隐式正则化

Geometry of Optimization and Implicit Regularization in Deep Learning

Toyota Technological Institute at Chicago(芝加哥丰田技术学院);MSR Cambridge(微软研究院剑桥);Carnegie Mellon University(卡内基梅隆大学)

Behnam Neyshabur, Ryota Tomioka, Ruslan Salakhutdinov, Nathan Srebro

摘要:

我们认为,通过隐式正则化,优化在深度学习模型的泛化中起着至关重要的作用。我们这样做是通过证明泛化能力不是由网络大小控制,而是由其他一些隐式控制。然后,我们展示了如何改变经验优化过程可以提高泛化,即使实际的优化质量不受影响。我们通过研究深度网络的参数空间的几何形状来实现这一点,并设计出一种与这种几何形状相协调的优化算法。

英文摘要:

We argue that the optimization plays a crucial role in generalization of deep learning models through implicit regularization. We do this by demonstrating that generalization ability is not controlled by network size but rather by some other implicit control. We then demonstrate how changing the empirical optimization procedure can improve generalization, even if actual optimization quality is not affected. We do so by studying the geometry of the parameter space of deep networks, and devising an optimization algorithm attuned to this geometry.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑