人工智能培训

搜索

人工智能论文:元辅助学习的自我监督泛化(Self-Supervised Generalisation with Meta Auxiliary Learning)

[复制链接]
hjxbakl 发表于 2019-1-28 12:03:50 | 显示全部楼层 |阅读模式
hjxbakl 2019-1-28 12:03:50 458 0 显示全部楼层
人工智能论文:元辅助学习的自我监督泛化(Self-Supervised Generalisation with Meta Auxiliary Learning)已经证明,学习辅助任务可以改善主要任务的概括。然而,这是以手动标记附加任务为代价的,这些任务可能对主要任务有用,也可能没有用。我们提出了一种新方法,可以自动学习辅助任务的标签,这样就可以改进任何监督学习任务,而无需访问其他数据。该方法是训练两个神经网络:用于预测辅助标签的标签生成网络,以及用于训练辅助任务的主要任务的多任务网络。标签生成网络的损失包含了多任务网络的性能,因此两个网络之间的这种交互可以被视为元学习的一种形式。我们表明,我们提出的方法Meta AuXiliary Learning(MAXL)在7个图像数据集上的表现优于单一任务,而不需要额外的辅助标签。我们还展示了MAXLout执行其他几个基线来生成辅助标签,并且与人工定义的辅助标签相比,它们具有竞争力。我们方法的这种监督性质导致了一种朝着自动化概括的有希望的新方向。源代码位于\ url {此https网址}。
Learning with auxiliary tasks has been shown to improve the generalisation ofa primary task.However, this comes at the cost of manually-labellingadditional tasks which may, or may not, be useful for the primary task.Wepropose a new method which automatically learns labels for an auxiliary task,such that any supervised learning task can be improved without requiring accessto additional data.The approach is to train two neural networks: alabel-generation network to predict the auxiliary labels, and a multi-tasknetwork to train the primary task alongside the auxiliary task.The loss forthe label-generation network incorporates the multi-task network's performance,and so this interaction between the two networks can be seen as a form of metalearning.We show that our proposed method, Meta AuXiliary Learning (MAXL),outperforms single-task learning on 7 image datasets by a significant margin,without requiring additional auxiliary labels.We also show that MAXLoutperforms several other baselines for generating auxiliary labels, and iseven competitive when compared with human-defined auxiliary labels.Theself-supervised nature of our method leads to a promising new direction towardsautomated generalisation.The source code is available at\url{this https URL}.人工智能论文:元辅助学习的自我监督泛化(Self-Supervised Generalisation with Meta Auxiliary Learning) pUTm6fOEDX7yfx7F.jpg
URL地址:https://arxiv.org/abs/1901.08933     ----pdf下载地址:https://arxiv.org/pdf/1901.08933    ----人工智能论文:元辅助学习的自我监督泛化(Self-Supervised Generalisation with Meta Auxiliary Learning)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

hjxbakl当前离线
新手上路

查看:458 | 回复:0

快速回复 返回顶部 返回列表