Although pretrained Transformers such as BERT achieve high accuracy on indistribution examples, do they generalize to new distributions? We systematically measure out-of-distribution (OOD) generalization for seven NLP datasets by constructing a new robustness benchmark with realistic distribution shifts. We measure the generalization of previous models including bag-of-words models, ConvNets, and LSTMs, and we show that pretrained Transformers' performance declines are substantially smaller. Pretrained transformers are also more effective at detecting anomalous or OOD examples, while many previous models are frequently worse than chance. We examine which factors affect robustness, finding that larger models are not necessarily more robust, distillation can be harmful, and more diverse pretraining data can enhance robustness. Finally, we show where future work can improve OOD robustness.

The high density of heat generated in power electronics and optoelectronic devices is a critical bottleneck in their application. New materials with high thermal conductivity are needed to effectively dissipate heat and thereby enable enhanced performance of power controls, solid-state lighting, communication, and security systems. We report the experimental discovery of high thermal conductivity at room temperature in cubic boron arsenide (BAs) grown through a modified chemical vapor transport technique. The thermal conductivity of BAs, 1000 ± 90 watts per meter per kelvin meter-kelvin, is higher than that of silicon carbide by a factor of 3 and is surpassed only by diamond and the basal-plane value of graphite. This work shows that BAs represents a class of ultrahigh-thermal conductivity materials predicted by a recent theory, and that it may constitute a useful thermal management material for high-power density electronic devices.

Gefitinib, a small molecule inhibitor of the epidermal growth factor receptor tyrosine kinase, has been shown to induce autophagy as well as apoptosis in tumor cells. Yet, how to exploit autophagy and apoptosis to improve therapeutic efficacy of this drug against cancer remains to be explored. We reported here that MK-2206, a potent allosteric Akt inhibitor currently in Phase I trials in patients with solid tumors, could reinforce the cytocidal effect of gefitinib against glioma. We found that co-treatment with gefitinib and MK-2206 increased the cytotoxicity of this growth factor receptor inhibitor in the glioma cells, and the Compusyn synergism/antagonism analysis showed that MK-2206 acted synergistically with gefitinib. The benefit of the combinatorial treatment was also demonstrated in an intracranial glioma mouse model. In the presence of MK-2206, there was a significant increase in apoptosis in glioma cells treated with gefitinib. MK-2206 also augmented the autophagy-inducing effect of gefitinib, as evidenced by increased levels of the autophagy marker, LC3-II. Inhibition of autophagy by silencing of the key autophagy gene, beclin 1 or 3-MA, further increased the cytotoxicity of this combinatorial treatment, suggesting that autophagy induced by these agents plays a cytoprotective role. Notably, at 48 hours following the combinatorial treatment, the level of LC3-II began to decrease but Bim was significantly elevated, suggesting a switch from autophagy to apoptosis. Based on the synergistic effect of MK-2206 on gefitinib observed in this study, the combination of these two drugs may be utilized as a new therapeutic regimen for malignant glioma.