Question

一段英文求翻译

We also run both methods with noise introduced using the technique mentioned in subsection 6.4. As seen in Figure 6, the DP algorithm performs signifi- cantly better than the supervised learning method, achiev- ing nearly 25% lower regret when there is no noise for RCV1. This is particularly encouraging given that the amount of feedback the supervised algorithm receives is vastly supe- rior in informativeness to that of the online learning method: While the supervised algorithm receives the relevance labels of each document for each of the user’s intent, the DP al- gorithm only receives a single preference (which has atmost 5 documents) in each iteration. Even for the η = 0.2 case, the DP algorithm is able to achieve lower regret eventually, indicating that the trend holds even under noisy conditions. Finally, note that the (per-iteration) training times of the supervised batch method are vastly larger than those of the DP algorithm (∼ 1000s vs. 0.1s). This is because the supervised method solves a more complex optimization problem (the structural SVM objective), while training the Diversifying Perceptron involves just a single update step. Consequently, this makes the DP algorithm especially use- ful in problem settings where we would like to continu- ously improve the learned model over time, something that would be prohibitively expensive with the supervised learn- ing method.

Answer 1

我们还有两种方法与噪声引入6.4款提到的技术。如图6所示，DP算法的性能明显fi明显优于监督学习方法，实现了近25%的下后悔的时候，没有噪声RCV1。这是特别令人鼓舞的，反馈监督算法接收量远远超优越在资讯，在线学习的方法：在有监督的算法接收为每个用户的意图，每个文档的相关性的标签，铝gorithm只接收一个单一的偏好的DP（其中有最多5个文件在每一次迭代）。即使对于η= 0.2的情况下，DP算法能够实现较低的遗憾，最终，表明趋势认为，即使在嘈杂的环境。最后，请注意，（每次迭代）的监督分批法训练次数大大高于DP算法较大（∼1000 VS 0.1s）。这是因为监督的方法解决了一个更复杂的优化问题（结构SVM的目标），而培训的多样化的感知器只涉及一个单一的更新步骤。因此，这使得DP算法特别适用的问题设置，我们要不断地改善模型随着时间的推移，这将是昂贵的监督学习方 希望对你有帮助，望采纳，谢谢