CASIA OpenIR  > 毕业生  > 博士学位论文
面向开放环境模式识别的小样本学习方法研究
王瑞琪
Subtype博士
Thesis Advisor刘成林 ; 张煦尧
2022-05-29
Degree Grantor中国科学院自动化研究所
Place of Conferral中国科学院自动化研究所
Degree Name工学博士
Degree Discipline模式识别与智能系统
Keyword小样本学习 广义小样本学习 特征自适应 数据增强 表示学习
Abstract

得益于深度学习(Deep Learning)方法的不断发展,机器智能已经可以在很多任务上接近甚至超过人类水平。但是,当今阶段人工智能仍有明显的缺陷,主要表现为模型学习需要大量的标记数据,且难以应对开放环境中类别和分布变化所带来的挑战。针对开放环境模式识别,如何构建能够适应类别和分布变化并根据少量样本进行学习的模型,是一个具有重要意义且极富挑战性的问题。本文基于图像分类任务,研究面向开放环境模式识别的小样本学习方法,内容主要包括特征自适应、表示学习和类别拓展的小样本学习等,旨在提高小样本条件下开放环境模式识别的性能。本文的主要创新性工作和成果如下:

一、提出了一种针对未知域小样本识别任务的元原型学习方法。}自适应是面对开放环境中各种挑战的一个必然选择。但是在小样本学习中,极少的样本数量使得其中的自适应问题更具有挑战性。本文通过结合原型分类器和元学习实现了在有剧烈域变化的开放环境中基于少量样本的高效自适应。该方法首先根据开放环境中新类别的标记样本,基于“类内紧致,类间分开”的原则进行特征自适应;在几步自适应之后,根据少量标记样本计算每个新类别的原型(类别中心)并以待分类样本到每个类别原型的距离进行分类。在训练时,本方法对参数的优化涉及到双层优化(bilevel optimization),使得模型的训练变得相对复杂。对此,本文基于梯度反传提出了一种一阶近似梯度更新方法,避免了求解高阶导带来的计算时间和计算空间上的开销,提高了模型的训练效率。实验表明,本方法在跨域的小样本识别任务上能带来明显的性能提升,并且在自适应的稳定性和效率上也明显优于对比方法。

二、提出了一种基于图的小样本子空间学习方法。}基于梯度的特征自适应需要进行多次前向传播和反向传播,这决定了它必然会有较大的计算开销和较低的效率。对此,本文提出了一种基于图的非线性特征子空间学习方法。该方法避免使用矩阵分解来求解特征空间,而是在图上以神经网络前向传播的方式学习子空间的基。具体来说,该方法先在统一的特征空间中依据标记样本初始化一组基向量,然后通过连接基向量和一个小样本任务中所有样本特征建立一个图。针对每个小样本任务,其特征子空间的基通过在图上进行消息传递更新节点特征来最终获得。本方法将有标记样本和待分类样本同时包含于图中用于学习,是一种直推式(transductive)方法。实验结果表明,本方法以更高的效率达到了更优的性能,在多种数据集的不同实验设定上表现出优秀的综合性能。

三、提出了一种针对开放环境的类别合成数据增强方法。}数据增强是一种被广泛使用的可以简单有效地提高模型鲁棒性和表示能力的技术。之前的数据增强方法大都针对闭集问题,旨在通过在训练集合成更多样本来提高模型在与训练集数据独立同分布(independent and identically distributed,i.i.d.)数据上的泛化性。但是在小样本学习这种处理开放环境中域外分布(out-of-distribution, o.o.d.)数据的任务中,这些数据增强方法难以带来明显的性能提升。对此,本文提出了一种合成新类别的数据增强方法以增强模型在开放环境中的特征表示性能。该方法将训练集中的类别拆分为不同视觉组件,并通过重新组合这些视觉组件来合成新的类别以实现在更大类别集上的训练。其最显著的特点在于合成新类别,而不是合成已知类别的新样本。同时,本文还探讨了训练时分类器中softmax函数的温度系数对模型中表示学习部分过拟合的影响。实验表明,本方法既简洁易实现,又能在小样本任务上实现稳定可观的性能提升。同时,本方法还具有较强的鲁棒性,可与多种不同的分类器适配。另外,本方法还在与开放环境相关的其它问题(如连续学习)中展现出其有效性。

四、提出了一种广义小样本学习中基于类分布的分类器学习方法。}目前的广义小样本学习方法都比较复杂,一般需要在训练集上训练两次。本文认为这主要是因为它们使用的判别式分类器不能存储训练集中基础类别(旧类别)数据分布的信息,因此不适合进行广义小样本学习这种增量式的学习。这就导致目前的一些方法大都需要通过一次额外的训练来调整模型。本文提出通过一次训练即得到一个视觉特征提取器和训练集中基础类别的类中心,并基于验证集估计一个所有类别共用的协方差矩阵。测试时,模型基于支持集计算新类别的类别中心,即可简洁地得到训练集中基础类别和开放环境中新类别的统一分类器,比如直接以每个分布的中心作为类别原型构建原型分类器。更进一步,基于估计的类分布模型也可以构建二次分类器来完成分类任务。实验表明,该方法在显著简化了广义小样本学习算法流程的同时达到了较好的分类性能。

Other Abstract

As the Deep Learning methods develop, machine intelligence has achieved competitive or even superior performance on many tasks. However, current artificial intelligence still has obvious defects. These defects mainly show as that the learning process requires a large amount of labeled data, and the learned models suffer from the challenges of class changes and domain shifts in the open world. Thus, it is a challenging and important problem that how to construct models that can adapt to the class changes and domain shifts and learn novel knowledge based on a few samples. This dissertation studies few-shot learning for open world pattern recognition, including feature adaptation, representation learning, and class-incremental few-shot learning, aiming to boost the performance on few-shot learning for open world pattern recognition. The major contributions are as follows:

1. A meta-prototypical learning method for domain-agnostic few-shot recognition is proposed. Adaptation is a necessary technique to handle various challenges in the open world. However, in few-shot learning, the limited samples make the problem even more challenging. This dissertation achieves efficient adaptation with only a few labeled samples in the open world by integrating prototype classifier and meta-learning. It firstly performs feature adaptation based on labeled samples from novel classes in the open world with the principle ``compact within classes and separate between classes''. After a few steps of adaptation, prototypes (centers) of novel classes are computed based on labeled samples and the queries are classified based on their distances to these prototypes. During training, the optimization of proposed method involves bilevel optimization, which makes the training process relatively complicated. Based on gradient back-propagation, a first-order approximate gradient update algorithm is introduced to avoid the expense on computing time and computing space for deriving high-order gradients, which improves the training efficiency of the proposed method. Experimental results prove that the proposed method brings obvious performance improvement on cross-domain few-shot learning tasks. And it shows superiority on stability and efficiency when compared with related methods.

2. A graph-based few-shot subspace learning method is proposed. Gradient-based feature adaptation requires multiple forward propagations and backward propagations, which bring large computing expenses and low efficiency. This dissertation proposes a graph-based non-linear feature subspace learning method that avoids learning subspaces by matrix decomposition. It learns the subspaces by the forward-propagation on a graph. Technically, the proposed method firstly initializes a group of bases based on labeled samples. Then, a graph is constructed by connecting the bases and sample features of a few-shot task. For each few-shot task, the bases of the subspace is learned by message propagation and vertex update on the graph. The proposed method includes samples with or without labels for learning, which makes it a transductive method. Experimental results prove that the proposed method achieves better performance with higher efficiency, outperforming current methods on different datasets with various experiment settings.

3. A data augmentation method that synthesizes new classes for open world tasks is proposed. Data augmentation is a widely-used technique that can simply but effectively boost the robustness of a model. Previous data augmentation methods are mostly for close-set problems, which aims to boost the generalization ability on independent and identically distributed (i.i.d.) data by synthesizing samples on the training set. However, in few-shot learning, where the model processes tasks in the open world, these methods can hardly bring noticeable performance improvement. Thus, this dissertation proposes a novel data augmentation method that synthesizes new classes. The proposed method divides classes in the training set into visual attributes and recombines these visual attributes to synthesize forged classes, with which the model can be trained with more classes. The most significant characteristic of the proposed method is that it synthesizes new classes other than new samples of known classes. Meanwhile, this dissertation also explores the effect of the temperature parameter in the softmax function during training. Experimental results prove that the proposed method is concise and effective, bringing stable and appreciable performance improvement in few-shot learning. Meanwhile, the proposed method has good robustness, which enables it to be combined with different classifiers. Further, the proposed method also shows its effects on other problems that are related to the open world, such as continual learning.

4. A distribution based classifeir learning method is proposed for generalzied few-shot learning. Most current methods for generalized few-shot learning are complicated, which include dual training processes on the training set. This dissertation argues that the reason is that they use discriminant classifiers that cannot store the information of class distributions, which makes them inappropriate for generalized few-shot learning that requires learning incrementally. Thus, most current methods require an extra training process to adjust the model. This dissertation proposes to learn the visual feature extractor and centers of training classes within one training process. Further, the validation set is used to compute a covariance matrix that is shared by all classes. During testing, based on centers of novel classes computed on support set, the model can generate a unified classifier for all base classes and novel classes. For example, a prototype classifier can be derived by directly using the centers of the distributions as prototypes. Further, based on class distributions, the model can also generate quadratic classifiers. Experimental results prove that the proposed method achieves good classification performance with a significantly concise procedure for generalized few-shot learning.

Pages138
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/48708
Collection毕业生_博士学位论文
模式识别国家重点实验室_模式分析与学习
Recommended Citation
GB/T 7714
王瑞琪. 面向开放环境模式识别的小样本学习方法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2022.
Files in This Item:
File Name/Size DocType Version Access License
王瑞琪-学位论文-signed.pdf(6050KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王瑞琪]'s Articles
Baidu academic
Similar articles in Baidu academic
[王瑞琪]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王瑞琪]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.