SmartDeal: Remodeling Deep Network Weights for Efficient Inference and Training.

The record-breaking performance of deep neural networks (DNNs) comes with heavy parameter budgets, which leads to external dynamic random access memory (DRAM) for storage. The prohibitive energy of DRAM accesses makes it nontrivial for DNN deployment on resource-constrained devices, calling for minimizing the movements of weights and data in order to improve the energy efficiency. Driven by this critical bottleneck, we present SmartDeal, a hardware-friendly algorithm framework to trade higher-cost memory storage/access for lower-cost computation, in order to aggressively boost the storage and energy efficiency, for both DNN inference and training. The core technique of SmartDeal is a novel DNN weight matrix decomposition framework with respective structural constraints on each matrix factor, carefully crafted to unleash the hardware-aware efficiency potential. Specifically, we decompose each weight tensor as the product of a small basis matrix and a large structurally sparse coefficient matrix whose nonzero elements are readily quantized to the power-of-2. The resulting sparse and readily quantized DNNs enjoy greatly reduced energy consumption in data movement as well as weight storage, while incurring minimal overhead to recover the original weights thanks to the required sparse bit-operations and cost-favorable computations. Beyond inference, we take another leap to embrace energy-efficient training, by introducing several customized techniques to address the unique roadblocks arising in training while preserving the SmartDeal structures. We also design a dedicated hardware accelerator to fully utilize the new weight structure to improve the real energy efficiency and latency performance. We conduct experiments on both vision and language tasks, with nine models, four datasets, and three settings (inference-only, adaptation, and fine-tuning). Our extensive results show that 1) being applied to inference, SmartDeal achieves up to 2.44× improvement in energy efficiency as evaluated using real hardware implementations and 2) being applied to training, SmartDeal can lead to 10.56× and 4.48× reduction in the storage and the training energy cost, respectively, with usually negligible accuracy loss, compared to state-of-the-art training baselines. Our source codes are available at: https://github.com/VITA-Group/SmartDeal.

Biosynthesis of lactones from diols mediated by an artificial flavin.

BACKGROUND: Lactones are important compounds in the field of medicine, material and chemical industry. One of the promising accesses to these flexible scaffolds is NAD(P)+-dependent alcohol dehydrogenases-catalyzed oxidative lactonization of diols, which relies on the construction of an efficient NAD(P)+ regeneration system. RESULTS: In this study, a novel system combining horse liver alcohol dehydrogenase (HLADH) with the synthetic bridged flavin cofactor was established for biosynthesis of lactones. The reaction conditions of this system were optimized and a variety of lactones including chiral lactones were efficiently obtained from various diols. Compared to the previously reported NAD(P)+-regeneration systems, this system showed better regeneration efficiency and product yield. A two-phase system was further applied to solve the problem of product inhibition, and 80% yield was obtained at the condition of 300 mM substrate. CONCLUSIONS: This study provides an efficient method to synthesis of lactones from diols under mild conditions. We believe this system will be a promising alternative to promote the synthesis of other valuable compounds.

Gambogic acid inhibits spinal cord injury and inflammation through suppressing the p38 and Akt signaling pathways.

Gamboge is the dry resin secreted by Garcinia hanburyi Hook.f, with the function of promoting blood circulation, detoxification, hemostasis and killing insects, used for the treatment of cancer, brain edema and other diseases. Gambogic acid is the main effective constituent of Gamboge. The present study investigated the protective effects of gambogic acid on spinal cord injury (SCI) and its anti­inflammatory mechanism in an SCI model in vivo. Basso, Beattie and Bresnahan (BBB) testing was used to detect the protective effects of gambogic acid on nerve function of SCI rats. The water content of the spinal cord was used to analyze the protective effects of gambogic acid on the damage of SCI. Treatment with gambogic acid effectively improved BBB scores and inhibited water content of the spinal cord in SCI rats. Also, gambogic acid significantly reduced inflammatory cytokines levels of [tumor necrosis factor­α, interleukin (IL)­6, IL­12 and IL­1ß] and oxidative stress (malondialdehyde, superoxide dismutase, glutathione and glutathione­peroxidase) factors, and suppressed receptor activator of nuclear factor κB ligand, phosphorylated p38 protein expression and toll­like receptor 4/nuclear factor­κB pathway activation, and increased phosphatidylinositol 3­kinase/protein kinase B (Akt) pathway activation in SCI rats. These results provide evidence that gambogic acid inhibits SCI and inflammation through suppressing the p38 and Akt signaling pathways.

Photon-Induced Light Emission from Foamed Gold with Micro/Nanohollow Sphere Structures.

We present a study on photon-induced light emission at room temperature from macroscale foamed gold with micro/nanoscale hollow spheres synthesized by seed-mediated growth method. Samples with a fixed sphere diameter but different Au densities are examined. It is demonstrated that strong and characteristic light emission from these samples can be achieved under optical excitation. In a short excitation wavelength regime (280-470 nm), the peak position in the photoemission spectrum increases almost linearly with excitation wavelength. In a relatively long-wavelength excitation regime (478-520 nm), photoluminescence (PL) can be observed where the peak position in the PL spectrum depends very little on excitation wavelength and two peaks can be seen in the PL emission spectrum. These effects do not change significantly with varying sample density, although it is found that the intensity of the light emission increases with sample density. We find that the features of the PL emission from foamed gold with micro/nanoscale hollow spheres differ significantly from those observed for Au nanoparticles. This study is relevant to the application of Au micro/nanostructures as advanced optoelectronic materials and devices.