Loss

最新推荐文章于 2023-04-20 16:25:20 发布

XuMengyaAmy

最新推荐文章于 2023-04-20 16:25:20 发布

阅读量209

点赞数

分类专栏：人工智能文章标签：神经网络深度学习

本文链接：https://blog.youkuaiyun.com/weixin_43332432/article/details/109661254

版权

人工智能专栏收录该内容

10 篇文章

订阅专栏

本文总结了论文《WHAT’S IN A LOSS FUNCTION FOR IMAGE CLASSIFICATION?》的关键发现：不同的损失函数和正则化方法虽然对准确率的影响有限，但足以在某些场景中产生意义；选择的损失函数主要影响网络的最后几层；不同目标导致倒数第二层的表示有显著差异，且此类表示与任务间的迁移能力成反比。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

Summary of Paper “WHAT’S IN A LOSS FUNCTION FOR IMAGE CLASSIFICATION?”

Different losses and regularizers achieve broadly similar accuracies. Although the acuuracy differences are larger enough to be meaningful in some contexts, the largerst difference is still < 1.5%.
The choice of loss function affects representations in only the last few layers of the network, suggesting inherent limitations to what can be achieved by manipulating the loss.
Different objectives (losses/regularizers) lead to substantially different penultimate layer representations. Claa separation is an important factor that distinguishes these different penultimate layer representations, ans show that it is inversely related to transferability of representation to other tasks.

References

Paper “WHAT’S IN A LOSS FUNCTION FOR IMAGE CLASSIFICATION?” https://arxiv.org/abs/2010.16402
https://neptune.ai/blog/pytorch-loss-functions?utm_source=facebook&utm_medium=post-in-group&utm_campaign=blog-pytorch-loss-functions&fbclid=IwAR13rfPMXDJQdTRBboUrycW8W-qIrevFbsvS4ppy00Da40hqJRK0kG9aNaU