持续学习——Continual Unsupervised Representation Learning——NeurIPS2019

最新推荐文章于 2022-11-29 17:12:47 发布

原创最新推荐文章于 2022-11-29 17:12:47 发布 · 654 阅读

2 ·

CC 4.0 BY-SA版权

文章标签：

#计算机视觉 #机器学习

博士科研同时被 3 个专栏收录

28 篇文章

订阅专栏

计算机视觉

17 篇文章

订阅专栏

增量学习

10 篇文章

订阅专栏

本文探讨了无监督持续学习方法，特别是在未知任务标识和边界的情况下学习表示。研究中使用生成模型来捕捉数据分布，并通过混合高斯隐空间进行任务推断，结合动态扩展和混合生成回放来创建新概念，最小化灾难性遗忘。

在这里插入图片描述

Abstract

Unsupervised continual learning (learning representations without any knowledge about task identity)

Introduction

挖坑写法，however, most of these techniques have focused on a sequence of tasks in which both the identity of the task (task label) and boundaries between tasks are provided; moreover, they often focus on the supervised learning setting, where class labels for each data point are given.
Unsupervised, 1) the absence of task labels (or indeed well-defined tasks themselves) 2) the absence of external supervision such as class labels, regression targets, or external rewards.

Method

用了生成模型来刻画数据分布
在这里插入图片描述

Conclusion

The proposed approach, performs task inference via a mixture-of-Gaussians latent space, and uses dynamic expansion and mixture generative replay to instantiate new concepts and minimize catastrophic forgetting.

Key points: 代码开源，paper with code stars多；no knowledge of task labels and boundaries; 实验数据集比较简单，MNIST和Ominiglot；文章写的一般；