Search for: All records

Generally, deciduous and evergreen trees coexist in subtropical forests, and both types of leaves are attacked by numerous insect herbivores. However, trees respond and defend themselves from herbivores in different ways, and these responses may vary between evergreen and deciduous species. We examined both the percentage of leaf area removed by herbivores as well as the percentage of leaves attacked by herbivores to evaluate leaf herbivore damage across 14 subtropical deciduous and evergreen tree species, and quantified plant defenses to varying intensities of herbivory. We found that there was no significant difference in mean percentage of leaf area removed between deciduous and evergreen species, yet a higher mean percentage of deciduous leaves were damaged compared to evergreen leaves (73.7% versus 60.2%). Although percent leaf area removed was mainly influenced by hemicellulose concentrations, there was some evidence that the ratio of non-structural carbohydrates:lignin and the concentration of tannins contribute to herbivory. We also highlight that leaf defenses to varying intensities of herbivory varied greatly among subtropical plant species and there was a stronger response for deciduous trees to leaf herbivore (e.g., increased nitrogen or lignin) attack than that of evergreen trees. This work elucidates how leaves respond to varying intensities of herbivory, and explores some of the underlying relationships between leaf traits and herbivore attack in subtropical forests.

 

Graph embeddings have been tremendously successful at producing node representations that are discriminative for downstream tasks. In this paper, we study the problem of graph transfer learning: given two graphs and labels in the nodes of the first graph, we wish to predict the labels on the second graph. We propose a tractable, noncombinatorial method for solving the graph transfer learning problem by combining classification and embedding losses with a continuous, convex penalty motivated by tractable graph distances. We demonstrate that our method successfully predicts labels across graphs with almost perfect accuracy; in the same scenarios, training embeddings through standard methods leads to predictions that are no better than random.