Search

Instruction tuning has been attracting much attention to achieve generalization ability across a wide variety of tasks. Although various types of instructions have been manually created for instruction tuning, it is still unclear what kind of instruction is optimal to obtain cross-task generalization ability. This work presents instruction optimization, which optimizes training instructions with respect to generalization ability. Rather than manually tuning instructions, we introduce learnable instructions and optimize them with gradient descent by leveraging bilevel optimization. Experimental results show that the learned instruction enhances the diversity of instructions and improves the generalization ability compared to using only manually created instructions., Comment: 14pages, 6 figures, accepted for Findings of ACL2023

Understanding the long-term impact of scientific findings requires understanding the dynamic process of new research fields' formation. In new research fields, slow-cited papers (SCP) and the awakeners (AW) are more likely to exist, indicating explorers revisited underrated but significant past papers relocating the findings in the new paradigm. This study acquired SCP-AW pairs located in the integrated point of two different research fields using the inheritance of clusters. We found that research field integration, including SCP-AW pairs, was diverse but followed a similar pattern throughout history, generating an equal mix of SCP and AW fields. The recent trend toward more AW-centric disciplinary combinations supports the belief that field integration will become increasingly technology-driven in the coming years.

Nesfatin-1 is an anorectic peptide expressed in both peripheral tissues and brain areas involved in the regulation of feeding, emotion and emesis. The aim of the present study is to characterize the distribution of NUCB2/nesfatin-1 in Suncus murinus and to investigate the actions of nesfatin-1 to affect gastrointestinal contractility, emesis, food and water intake, and locomotor activity. The deduced amino acid sequence of S. murinus nesfatin-1 using in silico cloning showed high homology with humans and rodents. NUCB2 mRNA was detected throughout the entire brain and in the gastrointestinal tract, including the stomach and gut. Western blot analysis and immunohistochemistry confirmed the expression of nesfatin-1 protein in these regions. The NUCB2 mRNA levels in the hypothalamus, hippocampus and brainstem were significantly decreased, whereas that in the striatum were increased after 24 h starvation compared to ad libitum-fed animals (p < 0.05). In in vitro studies, nesfatin-1 (0.3–1,000 pM) failed to contract or relax the isolated gastric antrum and intestinal segments. In conscious, freely moving animals, intracerebroventricular administration of nesfatin-1 (1–50 pmol) induced emesis (p < 0.05) and suppressed 6-h cumulative food intake (p < 0.05), without affecting the latency to feeding. Nesfatin-1 (25 pmol, i.c.v.) decreased 24-h cumulative food and water intake by 28.3 and 35.4%, respectively (p < 0.01). No significant differences in locomotor activity were observed. In conclusion, NUCB2/nesfatin-1 might be a potent regulator of feeding and emesis in S. murinus. Further studies are required to elucidate the mechanism of actions of this peptide as a mediator linking the brainstem NUCB2/nesfatin-1 to forebrain system.

In recent years, there has been a rapid increase in viral spreading social issues that emerge in the public consciousness through the fast spread of information online. With the advent of social media, they sometimes yield unexpected side effects such as product boycotts. Therefore, it is important to recognize them as the earliest and take preventive measures against them. Existing researches on social issue extraction have mainly focused on news channels and newspapers as the primary information sources. However, such approaches cannot be applied for the early extraction of viral spreading social issues because their epicenter is the online public’s opinion. In this study, we propose a constructive method inspired by a social issues research approach, called constructivism, for the early extraction of viral spreading social issues. It is characteristic that our method identifies the keywords related to social issues using information obtained from the claims-making activities on Twitter and Twitter-user clustering. We conducted experiments on a large Twitter dataset comprising tens of billions of tweets and the proposed method successfully extracted six out of the seven viral spreading social issues earlier than their first TV news coverage. Furthermore, the proposed method could identify such cases approximately two weeks earlier, on average, than the first national TV news coverage.

It has been observed that the origination of innovation tends to concentrate in areas in which environmental conditions are formed as clusters. It is networks with "small-world" architecture that underpin the high level of creative power for innovation that clusters possess. If it were possible, through policy efforts, to form networks of a highly small-world character that provide an environment that can facilitate a good balance between short-distance interaction and long-distance interaction, that would lead to the enhancement of a region's creative power for innovation. The cluster policy currently being implemented has completed its first phase and has reached a watershed at which it moves into the second phase. In this paper we attempt to formulate a method of objectively grasping and evaluating network architecture from the standpoint of these two types of interaction. Specifically, with regard to the architecture of major networks formed domestically, we use network analysis methods to conduct a comparative analysis of the 12 regions and fields before (in 2000) and after (in 2005) the implementation of cluster policy. As a result of this, we have clarified the following points. (1) Networks are expanding in all regions, (2) the networks that excel in long-distance interaction also excel in short-distance interaction, (3) with a small number of exceptions, the bigger the network is, the more both of these characteristics are enhanced, (4) during this five-year period there have been no major changes with regard to the comparative advantage of each region or field, (5) there is some degree of correlation between the independence of modules and the extent of their small-world character, and (6) disparities between regions are greater than disparities between industry sectors. Through the use of this analytical method it was possible to obtain a quantitative grasp of network architecture in a form that made possible comparisons with other regions and fields, and from this to extract objective information necessary for policy-making. We hope that efficient networking activity will be carried out in the future on the basis of the results of analysis of this kind.

The diversity of industry-university-government networks formed within regional clusters may dictate the degree of vigor of cutting-edge industries within a zone. After gaining a multidimensional grasp of the structural characteristics of the networks, we discuss their relationship with the location of concentrations of cutting-edge industries and with their growth. Specifically, we introduced certain network analysis methods and conducted quantitative analysis of health care-related industries in the Kinki (Osaka, Kyoto, Kobe) economic zone and the semiconductor industry in the northern Kyushu economic zone. As a result, we verified first that the networks in both regions and fields have the characteristics of 'small-world' networks; second, that within both networks there exist numerous modules in groupings of varying sizes, and that these have a loosely bound structure; third, that the groups of firms in same business fields within the health care-related industries in the Kinki region form tight horizontally linked modules, while among the system LSI and other semiconductor industries in northern Kyushu there is a mixture of modules formed by individual core manufacturers' vertically segmented keiretsu and horizontal collaboration, and differences in architecture between the two types were perceived; fourth, that there is a high degree of unity in the network within each discrete block-area economic zone; and fifth, that by means of microanalysis in which we observed the major nodes, we established that the core firms, research universities, trading companies, etc., in industrial fields act as connector hubs of the networks. In these two regions characterized by the remarkable concentration and growth of industries in cutting-edge fields, wide-area networks have been formed in a manner adapted for the rapid exchange and melding of information and knowledge, joint business activity, and industry-university collaboration. This indicates the possibility that excellent networks contribute to the nurturing of cutting-edge industries. In addition, in similar large-scale, cutting-edge industries, network structures differ according to the core industrial fields and regional characteristics. Policy efforts aimed at extending networks could be more effective if they were based on these structural characteristics.