深度学习劝退文

深度强化学习劝退:

[1709.06560] Deep Reinforcement Learning that Matters

John Schulman 在自己的PPO那篇文章里用的TRPO,跟他自己在TRPO里用的TRPO的performance都差了一截。
同样的算法同样的参数只换seed得出的结果:
同样的算法换下activation function:
同样的算法换下网络结构:
那几个都声称自己是STOA的算法没有一个consistently outperform the others:

GAN劝退:

[1711.10337] Are GANs Created Equal? A Large-Scale Study


深度学习全面劝退:

Do Convolutional Neural Networks act as Compositional Nearest Neighbors?

CNN费了半天功夫就学了个Nearest Neighbors:

另一篇全面劝退:

[1707.05589] On the State of the Art of Evaluation in Neural Language Models

这篇文章对neural language models的分析基本可以得出:(Frank大神): Since methods were probably not tuned well before, the results were not as good as they could have been, and it's less clear how important the architecture search was. So, all in all, ... could just find results that could've been found with other methods.
另言之:在那些architecture上加几层减几层带来的几个百分点的提升可能根本没有任何意义,因为就算一个最最简陋的architecture,做好hyper parameter tuning,用对optimizer,就也可以达到那样的效果。

结论:

全都是noise.

全都是Nearest Neighbors.

编辑于 2017-11-30