人工智能培训

搜索

深度学习论文:通过查询训练学习非定向模型(Learning undirected models via query training)

[复制链接]
犀牛 发表于 2019-12-9 14:24:33 | 显示全部楼层 |阅读模式
犀牛 2019-12-9 14:24:33 258 0 显示全部楼层
深度学习论文:通过查询训练学习非定向模型(Learning undirected models via query training)变分自动编码器中的典型摊销推论专门用于单个概率查询。在这里,我们提出了一种推理网络体系结构,该体系结构可以推广到看不见的概率查询。代替编码器/解码器对,我们可以使用不仅在样本上而且在查询上都是随机的成本函数,直接从数据训练单个推理网络。我们可以使用该网络执行与具有隐藏变量的无向图形模型中相同的推理任务,而不必处理棘手的分区函数。结果可以映射到实际无向模型的学习,这是一个众所周知的难题。我们的网络还根据需要边缘化讨厌的变量。我们证明了我们的方法可以针对看不见的测试数据概括出看不见的概率查询,从而提供快速灵活的推断。实验表明,该方法在9个基准数据集上的性能优于或匹配PCD和AdVIL。
Typical amortized inference in variational autoencoders is specialized for asingle probabilistic query.Here we propose an inference network architecturethat generalizes to unseen probabilistic queries.Instead of an encoder-decoderpair, we can train a single inference network directly from data, using a costfunction that is stochastic not only over samples, but also over queries.Wecan use this network to perform the same inference tasks as we would in anundirected graphical model with hidden variables, without having to deal withthe intractable partition function.The results can be mapped to the learningof an actual undirected model, which is a notoriously hard problem.Our networkalso marginalizes nuisance variables as required.We show that our approachgeneralizes to unseen probabilistic queries on also unseen test data, providingfast and flexible inference.Experiments show that this approach outperforms ormatches PCD and AdVIL on 9 benchmark datasets.深度学习论文:通过查询训练学习非定向模型(Learning undirected models via query training)
URL地址:https://arxiv.org/abs/1912.02893     ----pdf下载地址:https://arxiv.org/pdf/1912.02893    ----深度学习论文:通过查询训练学习非定向模型(Learning undirected models via query training)
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

犀牛当前离线
新手上路

查看:258 | 回复:0

快速回复 返回顶部 返回列表