大多数现有的视觉语言预训练方法侧重于理解任务,并在训练时使用类似于BERT的损失函数(掩码语言建模和图像文本匹配).尽管它们在许多理解类型的下游任务中表现良好,例如视觉问答、图像文本检索和视觉蕴涵,但它们不具备生成信息的能力....大多数现有的视觉语言预训练方法侧重于理解任务,并在训练时使用类似于BERT的损失函数(掩码语言建模和图像文本匹配).尽管它们在许多理解类型的下游任务中表现良好,例如视觉问答、图像文本检索和视觉蕴涵,但它们不具备生成信息的能力.为了解决这个问题,提出了视觉语言理解和生成的统一多模态预训练(unified multimodal pre-training for vision-language understanding and generation,UniVL).UniVL能够处理理解任务和生成任务,并扩展了现有的预训练范式,同时使用随机掩码和因果掩码,因果掩码即掩盖未来标记的三角形掩码,这样预训练的模型可以具有自回归生成的能力.将几种视觉语言理解任务规范为文本生成任务,并使用基于模版提示的方法对不同的下游任务进行微调.实验表明,在使用同一个模型时,理解任务和生成任务之间存在权衡,而提升这两个任务的可行方法是使用更多的数据.UniVL框架在理解任务和生成任务方面的性能与最近的视觉语言预训练方法相当.此外,实验还证明了基于模版提示的生成方法更有效,甚至在少数场景中它优于判别方法.展开更多
Correlation among physicochemical properties, which include amylose content, alkali spreading values, gel consistency, water absorption, expansion rate, solid content of rice-water, protein content and fat content, an...Correlation among physicochemical properties, which include amylose content, alkali spreading values, gel consistency, water absorption, expansion rate, solid content of rice-water, protein content and fat content, and eating qualities of six kinds of rice samples planted in Heilongjiang Province were studied. Correlation analysis showed that amylose content, water absorption and expansion rate were negatively correlated with eating qualities, yet gel consistency, alkali spreading values, solid content of rice-water and fat content were positively correlated with eating qualities. Among them, eating quality had an obvious correlation with amylose content and gel consistency, but no significant correlation with protein content. The regression equation, which described the relationship of the eating quality scores and physicochemical indexes, was Y=94.439–12.711X1–23.721X2–0.701X3＋0.570X4＋186.938X5（X1, X2, X3, X4 and X5 represented amylose content, water absorption, expansion rate gel, consistency and fat content）. The single factor analysis of variance showed that six kinds of rice existed significant differences in quality category.展开更多
文摘大多数现有的视觉语言预训练方法侧重于理解任务,并在训练时使用类似于BERT的损失函数(掩码语言建模和图像文本匹配).尽管它们在许多理解类型的下游任务中表现良好,例如视觉问答、图像文本检索和视觉蕴涵,但它们不具备生成信息的能力.为了解决这个问题,提出了视觉语言理解和生成的统一多模态预训练(unified multimodal pre-training for vision-language understanding and generation,UniVL).UniVL能够处理理解任务和生成任务,并扩展了现有的预训练范式,同时使用随机掩码和因果掩码,因果掩码即掩盖未来标记的三角形掩码,这样预训练的模型可以具有自回归生成的能力.将几种视觉语言理解任务规范为文本生成任务,并使用基于模版提示的方法对不同的下游任务进行微调.实验表明,在使用同一个模型时,理解任务和生成任务之间存在权衡,而提升这两个任务的可行方法是使用更多的数据.UniVL框架在理解任务和生成任务方面的性能与最近的视觉语言预训练方法相当.此外,实验还证明了基于模版提示的生成方法更有效,甚至在少数场景中它优于判别方法.
基金Supported by Program for Technological Innovation Research Team in University of Heilongjiang Province(2010td04)
文摘Correlation among physicochemical properties, which include amylose content, alkali spreading values, gel consistency, water absorption, expansion rate, solid content of rice-water, protein content and fat content, and eating qualities of six kinds of rice samples planted in Heilongjiang Province were studied. Correlation analysis showed that amylose content, water absorption and expansion rate were negatively correlated with eating qualities, yet gel consistency, alkali spreading values, solid content of rice-water and fat content were positively correlated with eating qualities. Among them, eating quality had an obvious correlation with amylose content and gel consistency, but no significant correlation with protein content. The regression equation, which described the relationship of the eating quality scores and physicochemical indexes, was Y=94.439–12.711X1–23.721X2–0.701X3＋0.570X4＋186.938X5（X1, X2, X3, X4 and X5 represented amylose content, water absorption, expansion rate gel, consistency and fat content）. The single factor analysis of variance showed that six kinds of rice existed significant differences in quality category.
