arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2006.08937cs.LGcs.CVstat.ML
收藏

Few-Shot分类中带遗忘更新模块的信道关系预测

Channel Relationship Prediction with Forget-Update Module for Few-shot Classification

Nanjing University(南京大学)

Minglei Yuan, Cunhao Cai, Tong Lu

摘要:

本文提出了一种支持集内类间关系的推理流水线,并给出了一个使用遗忘更新模块的查询示例。我们首先提出了一个新的架构模块称为“通道向量序列构造模块”,它通过收集所有支持样本和查询样本的整体信息来提高基于序列预测模型的Few-Shot分类方法的性能。该模块生成的通道向量序列以这样的方式组织,即该序列的每个时间步包含来自所有支持样本和待推断的查询样本的对应通道的信息。通过卷积神经网络和全连接网络得到信道向量序列,拼接后的信道向量序列是支持样本和查询样本对应的信道向量按原信道顺序拼接而成。此外,我们提出了一个遗忘更新模块组成的堆叠遗忘更新块。遗忘块用学习的权重修改原始信息,更新块为模型建立密集连接。该流水线由通道向量序列构造模块和遗忘更新模块组成,能够在Few-Shot分类场景下推断查询样本和支持样本之间的关系。实验结果表明,该流水线在miniImagenet、CUB数据集和跨域场景上均能取得较好的效果。

英文摘要:

In this paper, we proposed a pipeline for inferring the relationship of each class in support set and a query sample using forget-update module. We first propose a novel architectural module called "channel vector sequence construction module", which boosts the performance of sequence-prediction-model-based few-shot classification methods by collecting the overall information of all support samples and a query sample. The channel vector sequence generated by this module is organized in a way that each time step of the sequence contains the information from the corresponding channel of all support samples and the query sample to be inferred. Channel vector sequence is obtained by a convolutional neural network and a fully connected network, and the spliced channel vector sequence is spliced of the corresponding channel vectors of support samples and a query sample in the original channel order. Also, we propose a forget-update module consisting of stacked forget-update blocks. The forget block modify the original information with the learned weights and the update block establishes a dense connection for the model. The proposed pipeline, which consists of channel vector sequence construction module and forget-update module, can infer the relationship between the query sample and support samples in few-shot classification scenario. Experimental results show that the pipeline can achieve state-of-the-art results on miniImagenet, CUB dataset, and cross-domain scenario.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑