Mechanical Enhancement of Graded Nanoporous Structure

Inspired by the development of strong and ductile composite and gradient materials over the past decade, here we report the investigation of a graded nanoporous (NP) structure utilizing molecular dynamics simulations. The structure is generated by inducing a gradient scaling parameter in a Gaussian random field model. It has a large ligament/pore size toward the core and a small ligament/pore size toward the surface. The redistribution of stress and strain under tensile loading is then scrutinized and compared between the functional graded NP structure and two conventional NP structures with identical relative density but constant ligament size. During loading, the thick ligaments in the gradient structure yield at high stress, leading to the entire structure's high mechanical strength. The thin ligaments help the structure accommodate significant plastic strain by promoting uniform deformation. Both parts of the gradient structure worked collectively and resulted in the structure exhibiting a synergy of excellent strength and good deformability.

Reducing Underflow in Mixed Precision Training by Gradient Scaling

By leveraging the half-precision floating-point format (FP16) well supported by recent GPUs, mixed precision training (MPT) enables us to train larger models under the same or even smaller budget. However, due to the limited representation range of FP16, gradients can often experience severe underflow problems that hinder backpropagation and degrade model accuracy. MPT adopts loss scaling, which scales up the loss value just before backpropagation starts, to mitigate underflow by enlarging the magnitude of gradients. Unfortunately, scaling once is insufficient: gradients from distinct layers can each have different data distributions and require non-uniform scaling. Heuristics and hyperparameter tuning are needed to minimize these side-effects on loss scaling. We propose gradient scaling, a novel method that analytically calculates the appropriate scale for each gradient on-the-fly. It addresses underflow effectively without numerical problems like overflow and the need for tedious hyperparameter tuning. Experiments on a variety of networks and tasks show that gradient scaling can improve accuracy and reduce overall training effort compared with the state-of-the-art MPT.

Feedback control of morphogen gradient scale

SUMMARYGradients of the morphogen decapentaplegic (Dpp) pattern Drosophila wing imaginal discs, establishing gene expression boundaries at specific locations. As discs grow, Dpp gradients expand, keeping relative boundary positions approximately stationary. Such scaling fails in mutants for Pentagone (pent), a gene repressed by Dpp that encodes a diffusible protein that expands Dpp gradients. Although these properties fit a recent mathematical model of automatic gradient scaling, we show here that Pent lacks a property essential to that model—the ability to spread with minimal loss throughout the morphogen field. Instead, Pent’s actions appear confined to within a few cell diameters of its site of synthesis, and can be phenocopied by manipulating non-diffusible targets of Pent strictly within the Pent expression domain. Through genetic manipulation and mathematical modeling we develop an alternative model of scaling, driven by feedback down-regulation of Dpp receptors and co-receptors. Among the model’s predictions is a size limit beyond which scaling fails—something we observe directly in wing discs.

Scale invariance of BMP signaling gradients in zebrafish

In both vertebrates and invertebrates, spatial patterning along the Dorsal-ventral (DV) embryonic axis depends on a morphogen gradient of Bone Morphogenetic Protein signaling. Scale invariance of DV patterning by BMPs has been found in both vertebrates and invertebrates, however the mechanisms that regulate gradient scaling remain controversial. To obtain quantitative data that can be used to address core questions of scaling, we introduce a method to tune the size of zebrafish embryos by reducing varying amounts of vegetal yolk. We quantified the BMP signaling gradient in wild-type and perturbed embryos and found that the system scales for reductions in cross-sectional perimeter of up to 30%. Furthermore, we found that the degree of scaling for intraspecies scaling within zebrafish is greater than that between Danioninae species.

LSTM-Based Hierarchical Denoising Network for Android Malware Detection

Mobile security is an important issue on Android platform. Most malware detection methods based on machine learning models heavily rely on expert knowledge for manual feature engineering, which are still difficult to fully describe malwares. In this paper, we present LSTM-based hierarchical denoise network (HDN), a novel static Android malware detection method which uses LSTM to directly learn from the raw opcode sequences extracted from decompiled Android files. However, most opcode sequences are too long for LSTM to train due to the gradient vanishing problem. Hence, HDN uses a hierarchical structure, whose first-level LSTM parallelly computes on opcode subsequences (we called them method blocks) to learn the dense representations; then the second-level LSTM can learn and detect malware through method block sequences. Considering that malicious behavior only appears in partial sequence segments, HDN uses method block denoise module (MBDM) for data denoising by adaptive gradient scaling strategy based on loss cache. We evaluate and compare HDN with the latest mainstream researches on three datasets. The results show that HDN outperforms these Android malware detection methods,and it is able to capture longer sequence features and has better detection efficiency than N-gram-based malware detection which is similar to our method.