System alignment supports cross-domain learning and zero-shot generalisation

Recent findings suggest conceptual relationships hold across modalities. For instance, if two concepts occur in similar linguistic contexts, they also likely occur in similar visual contexts. These similarity structures may provide a valuable signal for system alignment when learning to map between domains, such as when learning the names of objects. To assess this possibility, we conducted a paired-associate learning experiment in which participants mapped objects that varied on two visual features to locations that varied along two spatial dimensions. We manipulated whether the featural and spatial systems were \textit{aligned} or \textit{misaligned}. Although system alignment was not required to complete this supervised learning task, we found that participants learned more efficiently when systems aligned and that aligned systems facilitated zero-shot generalisation. We fit a variety of models to individuals' responses and found that models which included an offline unsupervised alignment mechanism best accounted for human performance. Our results provide empirical evidence that people align entire representation systems to accelerate learning, even when learning seemingly arbitrary associations between two domains.

Aligned natural inflation in the Large Volume Scenario

We embed natural inflation in an explict string theory model and derive observables in cosmology. We achieve this by compactifying the type IIB string on a Calabi-Yau orientifold, stabilizing moduli via the Large Volume Scenario, and configuring axions using D7-brane stacks. In order to obtain a large effective decay constant, we employ the Kim-Nilles-Peloso alignment mechanism, with the required multiple axions arising naturally from generically anisotropic bulk geometries. The bulk volumes, and hence the axion decay constants, are stabilized by generalized one-loop corrections and subject to various conditions: the Kähler cone condition on the string geometry; the convex hull condition of the weak gravity conjecture; and the constraint from the power spectrum of scalar perturbations. We find that all constraints can be satisfied in a geometry with relatively small volume and thus heavy bulk axion mass. We also covariantize the convex hull condition for the axion-dilaton-instanton system and verify the normalization of the extremal bound.

A virtual joint method of analytical form for parallel manipulators

Oriented towards the stiffness optimization of parallel manipulators, a stiffness modeling method based on subspace analysis was proposed. The paper revealed the “shielding effect” of passive joints, i.e. shielding the partial local stiffness of a mechanism on Cartesian space, and accordingly brought forward the concept of Effective Stiffness Tensor (EST), which was then obtained by introducing the orthogonal projector. The method allows a stiffness model of analytical form as well as high computation efficiency, which makes it suitable for applications in the stiffness optimization of parallel manipulators. Proposed method had been applied to the parallel alignment mechanism in China’s large space telescope and verified by dynamic experiments.

A statistical analysis of dust polarization properties in ALMA observations of Class 0 protostellar cores

Context. Recent observational progress has challenged the dust grain-alignment theories used to explain the polarized dust emission routinely observed in star-forming cores. Aims. In an effort to improve our understanding of the dust grain alignment mechanism(s), we have gathered a dozen ALMA maps of (sub)millimeter-wavelength polarized dust emission from Class 0 protostars and carried out a comprehensive statistical analysis of dust polarization quantities. Methods. We analyze the statistical properties of the polarization fraction Pfrac and the dispersion of polarization position angles S. More specifically, we investigate the relationship between S and Pfrac as well as the evolution of the product S × Pfrac as a function of the column density of the gas in the protostellar envelopes. We compare the observed trends with those found in polarization observations of dust in the interstellar medium and in synthetic observations of non-ideal magneto-hydrodynamic (MHD) simulations of protostellar cores. Results. We find a significant S ∝ Pfrac−0.79 correlation in the polarized dust emission from protostellar envelopes seen with ALMA; the power-law index significantly differs from the one observed by Planck in star-forming clouds. The product S × Pfrac, which is sensitive to the dust grain alignment efficiency, is approximately constant across three orders of magnitude in envelope column density (from NH2 = 1022 cm−2 to NH2 = 1025 cm−2), with a mean value of 0.36−0.17+0.10. This suggests that the grain alignment mechanism producing the bulk of the polarized dust emission in star-forming cores may not systematically depend on the local conditions such as the local gas density. However, in the lowest-luminosity sources in our sample, we find a hint of less efficient dust grain alignment with increasing column density. Our observations and their comparison with synthetic observations of MHD models suggest that the total intensity versus the polarized dust are distributed at different intrinsic spatial scales, which can affect the statistics from the ALMA observations, for example, by producing artificially high Pfrac. Finally, synthetic observations of MHD models implementing radiative alignment torques (RATs) show that the statistical estimator S × Pfrac is sensitive to the strength of the radiation field in the core. Moreover, we find that the simulations with a uniform perfect alignment (PA) of dust grains yield, on average, much higher S × Pfrac values than those implementing RATs; the ALMA values lie among those predicted by PA, and they are significantly higher than the ones obtained with RATs, especially at large column densities. Conclusions. Ultimately, our results suggest that dust alignment mechanism(s) are efficient at producing dust polarized emission in the various local conditions typical of Class 0 protostars. The grain alignment efficiency found in these objects seems to be higher than the efficiency produced by the standard RAT alignment of paramagnetic grains. Further studies will be needed to understand how more efficient grain alignment via, for example, different irradiation conditions, dust grain characteristics, or additional grain alignment mechanisms can reproduce the observations.

Inductive Link Prediction for Nodes Having Only Attribute Information

Predicting the link between two nodes is a fundamental problem for graph data analytics. In attributed graphs, both the structure and attribute information can be utilized for link prediction. Most existing studies focus on transductive link prediction where both nodes are already in the graph. However, many real-world applications require inductive prediction for new nodes having only attribute information. It is more challenging since the new nodes do not have structure information and cannot be seen during the model training. To solve this problem, we propose a model called DEAL, which consists of three components: two node embedding encoders and one alignment mechanism. The two encoders aim to output the attribute-oriented node embedding and the structure-oriented node embedding, and the alignment mechanism aligns the two types of embeddings to build the connections between the attributes and links. Our model DEAL is versatile in the sense that it works for both inductive and transductive link prediction. Extensive experiments on several benchmark datasets show that our proposed model significantly outperforms existing inductive link prediction methods, and also outperforms the state-of-the-art methods on transductive link prediction.

MOOC opinion mining based on attention alignment

Purpose The purpose of this paper is to propose an attention alignment method for opinion mining of massive open online course (MOOC) comments. Opinion mining is essential for MOOC applications. In this study, the authors analyze some of bidirectional encoder representations from transformers (BERT’s) attention heads and explore how to use these attention heads to extract opinions from MOOC comments. Design/methodology/approach The approach proposed is based on an attention alignment mechanism with the following three stages: first, extracting original opinions from MOOC comments with dependency parsing. Second, constructing frequent sets and using the frequent sets to prune the opinions. Third, pruning the opinions and discovering new opinions with the attention alignment mechanism. Findings The experiments on the MOOC comments data sets suggest that the opinion mining approach based on an attention alignment mechanism can obtain a better F1 score. Moreover, the attention alignment mechanism can discover some of the opinions filtered incorrectly by the frequent sets, which means the attention alignment mechanism can overcome the shortcomings of dependency analysis and frequent sets. Originality/value To take full advantage of pretrained language models, the authors propose an attention alignment method for opinion mining and combine this method with dependency analysis and frequent sets to improve the effectiveness. Furthermore, the authors conduct extensive experiments on different combinations of methods. The results show that the attention alignment method can effectively overcome the shortcomings of dependency analysis and frequent sets.

Voting at the dawn of a global pandemic

What is the impact of a global health crisis on political behavior? We study the effect of the COVID-19 pandemic on electoral choice based on the case of Germany, one of the countries most heavily affected by the crisis. Our data come from the German state of Bavaria, where local elections were held right at the beginning of the pandemic. The elections took place early during the outbreak when there was still substantial variation in the extent to which individual counties and municipalities were affected by the outbreak. This variation provides a unique opportunity to study the causal impact of an event that would shortly after grow into an all-encompassing epidemic. We provide evidence that shows that the disease spread across the state in a mostly haphazard fashion. This lack of a discernible pattern coupled with within-county estimation of effects and a difference-in-differences strategy allow us to causally asses the effect of the spreading of the virus on electoral outcomes. Our results show that the crisis strongly and consistently benefited the dominant regional party, the CSU, and its candidates. For 3 known cases per 100,000 inhabitants, vote shares increased by about 4 percent. We explain our findings with a strategic-alignment mechanism, whereby voters vote into power candidates that they deem most likely to be able to solicit support from higher levels of government. Our findings emphasize the merit of forward-looking theories of voting and provide insights on the functioning of democracy during times of crisis.