少样本学习是目前机器学习研究领域的热点和难点.针对现有的少样本学习模型不能有效捕捉数据特征与数据标签之间的联系,造成分类模型泛化能力弱的问题,提出一种基于元学习的原型空间图卷积网络少样本学习模型FSL-GCNPS(Few-Shot Learnin...少样本学习是目前机器学习研究领域的热点和难点.针对现有的少样本学习模型不能有效捕捉数据特征与数据标签之间的联系,造成分类模型泛化能力弱的问题,提出一种基于元学习的原型空间图卷积网络少样本学习模型FSL-GCNPS(Few-Shot Learning of Graph Convolutional Network on Prototype Space).首先,利用卷积神经网络提取多任务数据的特征向量;其次,为了将特征向量映射到原型空间中,根据元学习的训练策略得到特征向量的类原型表达;然后,通过类原型向量和类向量之间的嵌入表示,构建图结构数据,并进行图卷积网络训练、推理.实验结果表明,相较于经典少样本学习方法,FSL-GCNPS模型拥有更好的分类准确率和分类稳定性.同时,在医学图像领域数据集上实验表明,FSL-GCNPS具有很好的跨域适应性.展开更多
深度神经网络是一种需要大量的数据来进行有效训练的模型。军事装备类数据普遍存在数据量较少,无法满足深度神经网络的训练需求,容易出现过拟合的问题。针对该问题,本文引入迁移学习技术,通过构建多类型样本训练集,微调预训练模型,构建...深度神经网络是一种需要大量的数据来进行有效训练的模型。军事装备类数据普遍存在数据量较少,无法满足深度神经网络的训练需求,容易出现过拟合的问题。针对该问题,本文引入迁移学习技术,通过构建多类型样本训练集,微调预训练模型,构建军事装备类集成分类器。实践证明迁移学习在少样本分类任务中的应用节省了模型训练时间,解决了模型过拟合及对数据标签依赖性强的问题,能有效提高基于深度学习的军事装备类小样本图像分类的准确性。A large amount of data is indispensable for effective training of deep neural networks. Military equipment data generally suffers from insufficient quantities, which fails to meet the training requirements of deep neural networks and easily leads to over fitting. To address this issue, this paper introduces transfer learning technology by constructing a multi-type sample training set and fine-tuning pre-trained models, and an ensemble classifier for military equipment is built. Experimental results have confirmed that transfer learning saves training time on small samples tasks, resolves issues of model over fitting and strong dependence on data labels simultaneously, and can effectively improve the accuracy of small sample image classification of military equipment based on deep learning.展开更多
文摘少样本学习是目前机器学习研究领域的热点和难点.针对现有的少样本学习模型不能有效捕捉数据特征与数据标签之间的联系,造成分类模型泛化能力弱的问题,提出一种基于元学习的原型空间图卷积网络少样本学习模型FSL-GCNPS(Few-Shot Learning of Graph Convolutional Network on Prototype Space).首先,利用卷积神经网络提取多任务数据的特征向量;其次,为了将特征向量映射到原型空间中,根据元学习的训练策略得到特征向量的类原型表达;然后,通过类原型向量和类向量之间的嵌入表示,构建图结构数据,并进行图卷积网络训练、推理.实验结果表明,相较于经典少样本学习方法,FSL-GCNPS模型拥有更好的分类准确率和分类稳定性.同时,在医学图像领域数据集上实验表明,FSL-GCNPS具有很好的跨域适应性.
文摘深度神经网络是一种需要大量的数据来进行有效训练的模型。军事装备类数据普遍存在数据量较少,无法满足深度神经网络的训练需求,容易出现过拟合的问题。针对该问题,本文引入迁移学习技术,通过构建多类型样本训练集,微调预训练模型,构建军事装备类集成分类器。实践证明迁移学习在少样本分类任务中的应用节省了模型训练时间,解决了模型过拟合及对数据标签依赖性强的问题,能有效提高基于深度学习的军事装备类小样本图像分类的准确性。A large amount of data is indispensable for effective training of deep neural networks. Military equipment data generally suffers from insufficient quantities, which fails to meet the training requirements of deep neural networks and easily leads to over fitting. To address this issue, this paper introduces transfer learning technology by constructing a multi-type sample training set and fine-tuning pre-trained models, and an ensemble classifier for military equipment is built. Experimental results have confirmed that transfer learning saves training time on small samples tasks, resolves issues of model over fitting and strong dependence on data labels simultaneously, and can effectively improve the accuracy of small sample image classification of military equipment based on deep learning.
