人工智能培训

搜索

机器学习论文:通过对齐变分自动编码器进行广义零点和少量学习(Generalized Zero- and Few-Shot Learning via Aligne

[复制链接]
baiselong 发表于 2018-12-6 11:54:53 | 显示全部楼层 |阅读模式
baiselong 2018-12-6 11:54:53 406 0 显示全部楼层
机器学习论文:通过对齐变分自动编码器进行广义零点和少量学习(Generalized Zero- and Few-Shot Learning via Aligned Variational  Autoencoders)广义零射击学习中的许多方法依赖于图像特征空间和类嵌入空间之间的跨模态映射。由于标记图像很少见,因此一个方向是通过生成图像或图像特征来增加数据集。然而,前者错过了细粒度的细节,后者需要学习与类嵌入相关的映射。在这项工作中,我们将特征生成更进一步,并提出了一个模型,其中图像特征和类嵌入的共享潜在空间是由模态特定的对齐变分自动编码器学习的。这使我们得到关于潜在特征中的图像和类的所需的判别信息,我们在其上训练softmax分类器。我们的方法的关键是我们将从图像和侧面信息中学习的分布对齐,以构建包含与看不见的类相关的基本多模态信息的潜在特征。我们在几个基准数据集上评估我们学到的潜在特征,即CUB,SUN,AWA1和AWA2,并建立了关于广义零射击以及少数射击学习的最新技术。此外,我们在ImageNet上的结果各种零点分割表明,我们的特征在大规模设置中得到了很好的推广。
Many approaches in generalized zero-shot learning rely on cross-modal mappingbetween the image feature space and the class embedding space.As labeledimages are rare, one direction is to augment the dataset by generating eitherimages or image features.However, the former misses fine-grained details andthe latter requires learning a mapping associated with class embeddings.Inthis work, we take feature generation one step further and propose a modelwhere a shared latent space of image features and class embeddings is learnedby modality-specific aligned variational autoencoders.This leaves us with therequired discriminative information about the image and classes in the latentfeatures, on which we train a softmax classifier.The key to our approach isthat we align the distributions learned from images and from side-informationto construct latent features that contain the essential multi-modal informationassociated with unseen classes.We evaluate our learned latent features onseveral benchmark datasets, ie CUB, SUN, AWA1 and AWA2, and establish a newstate-of-the-art on generalized zero-shot as well as on few-shot learning.Moreover, our results on ImageNet withvarious zero-shot splits show that ourlatent features generalize well in large-scale settings.机器学习论文:通过对齐变分自动编码器进行广义零点和少量学习(Generalized Zero- and Few-Shot Learning via Aligned Variational  Autoencoders) UlFz4fzG64YN6ftl.jpg
URL地址:https://arxiv.org/abs/1812.01784     ----pdf下载地址:https://arxiv.org/pdf/1812.01784    ----机器学习论文:通过对齐变分自动编码器进行广义零点和少量学习(Generalized Zero- and Few-Shot Learning via Aligned Variational  Autoencoders)
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

baiselong当前离线
新手上路

查看:406 | 回复:0

快速回复 返回顶部 返回列表