Search

In this study, we propose a simple and effective quantization-aware neural network pruning method to prune and quantize deep neural networks and then run them on lightweight and cost-effective embedded industrial devices. The proposed pruning method provides a suitable pruning strategy for a given bit precision before performing the subsequent quantization. In particular, the dropout rate is set to be bounded below by the maximum number of quantization levels, which enables the regulation of the tradeoff between sparsity and performance loss. The Kullback-Leibler divergence term of the evidence lower bound is also weighted for adjusting the performance loss. From learned dropout rates, efficient bit allocation is provided to maintain randomness even through the quantization step. In supervised and reinforcement learning applications, the proposed quantization-aware pruning method was proven to be successful in obtaining sparse and lightweight neural networks while maintaining their performance even after quantization. In particular, a real inverted pendulum system was used to determine whether the proposed method works well for actual physical control systems.

Recent advancements in deep reinforcement learning for real control tasks have received interest from many researchers and field engineers in a variety of industrial areas. However, in most cases, optimal policies obtained by deep reinforcement learning are difficult to implement on cost-effective and lightweight platforms such as mobile devices. This can be attributed to their computational complexity and excessive memory usage. For this reason, this article proposes an off-policy deep reinforcement learning algorithm called the sparse variational deterministic policy gradient (SVDPG). SVDPG provides highly efficient policy network compression under the standard reinforcement learning framework. The proposed SVDPG integrates Bayesian pruning, which is known as a state-of-the-art neural network compression technique, with the policy update in an actor–critic architecture for reinforcement learning. It is demonstrated that SVDPG achieves a high compression rate of policy networks for continuous control benchmark tasks while preserving a competitive performance. The superiority of SVDPG in low-computing power devices is proven by comparing the level of compression in terms of the memory requirements and computation time on a commercial microcontroller unit. Finally, it is confirmed that the proposed SVDPG is also reliable in real-world scenarios since it can be applied to the swing-up control of an inverted pendulum system.

Previous research shows that economic sanctions affect three facets of public opinion in target states: support for the policy at issue, support for the target government, and hostility toward the sanctioner. We explore the dynamics between the facets of opinion and link them to consumer behavior. How do supportive opinions of the policy and the target government lead to hostility? How does this hostility affect consumers’ propensity to buy the sanctioner's branded products? We examine a case in which China imposed economic sanctions on South Korea in opposition to South Korea's decision to deploy Terminal High Altitude Area Defense. We collect comments from relevant newspaper articles and conduct attitude analysis using machine learning. We conduct difference-in-differences analyses using barcode-level data regarding monthly beer sales, for which we argue boycotts are more likely to occur. We find that a backlash effect in public opinion occurred with respect to two facets of opinion. However, despite the public antagonism, we observe no significant backlash in consumer behavior. These findings imply that effects of economic sanctions in target states are multidimensional and thus it is too simplistic to assess the effectiveness of economic sanctions only by looking at public opinion in target states.

We previously reported that physical inactivity (PI) induces cognitive decline and depressive states, which were ameliorated by regular exercise. However, the mechanism underlying the preventive effect of exercise remains unelucidated. Irisin has recently been identified as an exercise-inducible myokine that improves cognitive impairment. Plasma irisin levels increase during physical exercise; therefore, PI could lead to a decline in cognitive function by reducing plasma irisin. Therefore, this study aimed to examine whether irisin is associated with cognitive decline and mental deterioration altered by PI and exercise. The mice were housed for eight weeks in the PI cage, whose living space was one-sixth that of a standard cage. Simultaneously, the mice were subjected to regular exercise in the presence or absence of an irisin-neutralizing antibody. PI increased the epididymal fat mass without increasing body weight, muscle mass, or plasma corticosterone levels. Additionally, PI induced anxiety, depressive states, and a decline in working memory. In contrast, regular exercise after PI elevated irisin levels in plasma and increased fibronectin type III domain-containing 5 (FNDC5) and peroxisome proliferator-activated receptor gammacoactivator 1α expression in skeletal muscle. Regular exercise also increased hippocampal brain-derived neurotrophic factor (BDNF) expression and BrdU-positive cells, alleviating cognitive decline and mental deterioration induced by PI. The beneficial effects of exercise were compromised by the administration of an irisin-neutralizing antibody. Moreover, plasma irisin level was positively correlated with working memory, hippocampal BDNF levels, and hippocampal cell proliferation. These findings suggest that exercise-inducible irisin is critical for maintaining cognitive function in the PI state.

A circular ferroelectric tunnel junction (C-FTJ) is proposed for a larger memory window and better endurance than a conventional planar FTJ (P-FTJ). The electrical characteristics of the proposed C-FTJ are evaluated compared with a conventional P-FTJ by using device simulation. It is confirmed that C-FTJs have more excellent ferroelectric switching than P-FTJs because the electric field becomes more concentrated across the ferroelectric layer. Also, C-FTJs show better endurance because the electric field applied to the interfacial layer is alleviated.

Acute physical exercise increases reactive oxygen species in skeletal muscle, leading to tissue damage and fatigue. Molecular hydrogen (H2) acts as a therapeutic antioxidant directly or indirectly by inducing antioxidative enzymes. Here, we examined the effects of drinking H2 water (H2-infused water) on psychometric fatigue and endurance capacity in a randomized, double-blind, placebo-controlled fashion. In Experiment 1, all participants drank only placebo water in the first cycle ergometer exercise session, and for comparison they drank either H2 water or placebo water 30 min before exercise in the second examination. In these healthy non-trained participants (n = 99), psychometric fatigue judged by visual analogue scales was significantly decreased in the H2 group after mild exercise. When each group was divided into 2 subgroups, the subgroup with higher visual analogue scale values was more sensitive to the effect of H2. In Experiment 2, trained participants (n = 60) were subjected to moderate exercise by cycle ergometer in a similar way as in Experiment 1, but exercise was performed 10 min after drinking H2 water. Endurance and fatigue were significantly improved in the H2 group as judged by maximal oxygen consumption and Borg’s scale, respectively. Taken together, drinking H2 water just before exercise exhibited anti-fatigue and endurance effects.

Bone grafting is necessary before dental implant treatment in patients with jaw bone defects. Currently, autologous bone grafting is a major burden on the patient. However, it is impossible to form a sufficient foundation for the implant with a bone-filling agent alone. It is, therefore, necessary to prepare hybrid artificial bone tissue containing osteoblasts and osteoclasts. In this study, mouse MC3T3-E1 pre-osteoblast cells and human embryonic-derived osteoblastic cell line hFOB1.19 were cultured in radial-flow bioreactors (RFB) to form three-dimensional artificial bone filled with porous beads of β-tricalcium phosphate (β-TCP) or hydroxyapatite (HA)—which are clinically used as bone-filling agents—as cell culture carriers. When circulation culturing was performed in the growth medium for the first 10–12 days, glucose consumption was increased in the cultures with HA beads in comparison to the cultures with β-TCP beads. When cultured in the differentiation culture medium during the second half of the culture period, the glucose consumption decreased in the culture with HA beads. A DNA microarray analysis suggested that osteogenesis progressed fast in three-dimensional culture filled with HA beads and that partly differentiation into osteoblasts was prominent in cultures with β-TCP beads. In the growth process of MC3T3-E1 cells, the vitamin A metabolism was also activated, the synthesis and degradation of retinoic acid was enhanced, and the metabolism of the same process decreased at the end of differentiation in three-dimensional cultures. Three-dimensional circulation culture in RFB is considered to be useful for the formation of hybrid bio-artificial bone tissue. Electronic supplementary material The online version of this article (10.1007/s13577-018-0218-x) contains supplementary material, which is available to authorized users.

Vitamin A is a fat-soluble vitamin required for many physiological functions. It is known that polar bears (Ursus maritimus) store a large amount of vitamin A within hepatic stellate cells (HSCs). However, the molecular mechanisms for high-capacity storage of vitamin A by polar bear HSCs are unknown. We biochemically and histochemically analyzed lecithin:retinol acyltransferase (LRAT), an enzyme responsible for conversion of retinol to its storage form, retinylesters, in the liver. Surprisingly, LRAT activities of the liver microsome fractions from polar bears weighing 100–500 kg were considerably lower than those obtained from rats (Rattus norvegicus). Although there is a possibility that the quality of polar bear liver specimens was reduced during transport, intrinsic LRAT activity of polar bears might actually be lower than those of other mammals. Fluorescent immunohistochemistry revealed that LRAT and cellular retinol-binding protein (CRBP) I were well co-localized within the liver lobules of polar bears. Taking into account that retinol bound to CRBP I is prevented from degradation and becomes a good substrate of LRAT, we hypothesized that CRBP I effectively conveys retinol to the LRAT, thereby circumventing the low enzymatic activity of LRAT in polar bear livers. As the top predator in the Arctic, polar bears have access to large amounts of vitamin A via the food web. We suggest that polar bears acquired an efficient mechanism of esterification of vitamin A within the liver during the course of their evolution.

We demonstrate a novel solution-free technique which can be used to print metal oxide patterns without masks by adopting an organic vapor jet printing (VJP) method. To avoid the high deposition temperatures required for metal oxide vaporization, metal-organic compounds are first jet-printed as metal-oxide precursors and converted into metal oxides by a photo-conversion process. As an example, we show that indium oxides can easily be prepared from jet-printed films of indium acetylacetonate upon photo-conversion; the effectiveness of which is supported by X-ray photoelectron spectroscopy (XPS) showing a reduction of the carbon portion associated with the alkyl chains of the precursors. A carrier mobility of 4.5 cm2/V·s and an on/off ratio of 105 are demonstrated in the oxide thin-film transistors based on indium oxides prepared by the proposed method, illustrating the potential usefulness of this process in electronic device fabrication.

While it has recently been the dominant trend in NLP to treat words as flat vectors, we are well aware that words have intelligible structures within. In this thesis, I aim to extract more structured, interpretable representations of words that reflect such inner workings, by means of generative modeling. Such representations are discovered in attempts to model the phenomenon of inflectional morphology, bywhich words are systematically categorized into different grammatical classes, and manifest form changes accordingly to such classes. Approaches to studying internal structures of words can categorize into one of the three families of morphological theories: morpheme-based item-and-arrangement (IA) models, lexeme-based item-and-process (IP) models, and word-based wordand-paradigm (WP) models. Here, I pay attention to the fact there have been relatively few attempts to computationally learn IP morphologies, especially in unsupervisedsettings. As the modern NLP community is equipped with more powerful tools to model—or approximate—complicated distributions, we are now in a good position to attempt extracting morphological models based on functional processes. In this work, I employ two different techniques to approach the problem of learning IP morphologies by generative modeling. First, I implement hidden Markovmodels with factored latent states, which emit continuous word embeddings, as a non-neural baseline. Then, I implement neural autoencoders that extend the factoredHMMs via adversarial training, to lift a number of simplifying assumptions made by the baseline. After a series of extensive evaluation experiments, I find out that although modeling of IP morphologies is an attainable goal when we have access to annotated data, purely unsupervised learning of IP morphologies remains to be a significant challenge. Specifically, it is confirmed that unrestricted training based on the common likelihood-based objective is likely to drift apart from human linguistic knowledge, without any means to regularize the training of highly complicatedmodels.

Although there are differences in the classification of category adopted by each country, small UAS is usually classified as the one less than 25 kg. UAS has been mainly used for military and public purposes, but in recent years, it has spread to the private sector for hobby, media, and so on. Especially, considering the nature of the operating region and applications, it is necessary to improve operating time, noise and vibration in small UAS to ensure the same level of safety with a manned aircraft. This is because the drone can pose health and safety hazard through collision with manned aircraft or crashing into the ground. In this paper, we investigated operational regulations in the United States and European countries. Based on the investigation, a domestic system development plan for small UAS operation is under development. Corresponding Author, E-mail: hjahn@kari.re.kr, TEL: 042-870-3572Ⓒ 2015 The Korean Society of Mechanical Engineers AC : Airworthiness CertificateAGL : Above Ground LevelARC : Aviation Rulemaking CommitteeATC : Air Traffic ControlBVLOS : Beyond Visual Line of SightCOA : Certification of Waiver or AuthorizationC2 : Command and ControlCS : Certification SpecificationsDAA : Detect and AvoidEASA : Europe Aviation Safety AgencyEU : Europe UnionEVLOS : Enhanced Visual Line of SightFAA : Federal Aviation AdministrationICAO : International Civil Aviation OrganizationLOS : Line of SightOC : Operator CertificateRTCA : Radio Technical Commission for AeronauticsTC : Type CertificateUAS : Unmanned Aerial SystemUAV : Unmanned Aerial VehicleVLOS : Visual Line of Sight - 기호설명

In this paper, we investigated the IMD (Inter-metal Dielectric) TDDB model where the BEOL interconnect physical spacing is reduced to less than 10nm as scaling progresses. IMD TDDB experiments were performed on 14 and 10 nm process technologies which uses double patterning technique having various interconnect line spacing ranging from the minimum design rule to wider spacings. The structures were examined for metal to metal (Mx-Mx) and metal to via (Vx-Mx) structures. Experimental results showed that the voltage and field acceleration parameters increased more than the model predicted below the specific spacing region due to more than expected current reduction during lower bias stress. As a result, it was found that the voltage/field acceleration parameters and the TDDB leakage depended upon the stress voltage below the certain physical spacing. This implied that the field-based TDDB model with constant field acceleration factor did not work in the narrow space region, which is about 18nm physically. Our long-term (up to 8 months of package level) TDDB results on 14nm and 10nm minimum Mx-Vx design rule structures and 1000hrs product HTOL data showed that square root E model is conservative. Power-law model was more appropriate enabling the further technology scaling will be detailed in the paper.

Background and aims We developed a bio-artificial liver (BAL) using a radial-flow bioreactor and rescued mini-pig models with lethal acute liver failure (ALF). The point of the rescue is the recovery from hepatic encephalopathy (HE). HE on ALF has sometimes resulted in brain death following brain edema with astrocyte swelling. Several factors, including ammonia and glutamine, have been reported to induce astrocyte swelling and injury. However, many clinicians believe that there are any other factors involved in the development of HE. Therefore, the aim of this study was to identify novel HE-inducible factors, particularly those inducing astrocyte dysfunction. Methods Mini-pig plasma samples were collected at three time points: before the administration of toxins (α-amanitin and LPS), when HE occurred after the administration of toxins, and after treatment with extracorporeal circulation (EC) by the BAL. To identify the causative factors of HE, each plasma sample was subjected to a comparative proteome analysis with two-dimensional gel electrophoresis and mass spectrometry. To assess the direct effects of candidate factors on the astrocyte function and injury, in vitro experiments with human astrocytes were performed. Results Using a proteome analysis, we identified alpha-1 antichymotrypsin (ACT), which was increased in plasma samples from mini-pigs with HE and decreased in those after treatment with EC by BAL. In in vitro experiments with human astrocytes, ACT showed growth-inhibitory and cytotoxic effects on astrocytes. In addition, the expression of water channel protein aquaporin-4, which is induced in injured astrocytes, was increased following ACT treatment. Interestingly, these effects of ACT were additively enhanced by adding arginine-vasopressin (AVP) and were canceled by adding an AVP receptor antagonist. Conclusions These results suggest that ACT is involved in astrocyte injury and dysfunction in concert with AVP during the development of acute HE.

Advanced process technology allows high memory density. However, the memory cell shrinkage introduces more memory defects and this causes a memory yield problem. To overcome the issue, memory ECC has become a critical solution. This paper proposes a hardware architecture to support memory ECC utilizing memory spares. Overheads imposed by the proposed architecture are analyzed and compared against conventional non-ECC memory architecture.

The simulation software for predicting the performance of a wind turbine generator system (WTGS) is validated using the field measured data obtained from the idling test run of a dual rotor wind turbine recently developed and installed in Korea. Both steady-state and transient responses at low and high wind conditions are compared with the theoretically predicted ones from the simulation software WINSIM.

Additional file 10: Table S10. Tukey HSD for mRNA 2 (Experiment 6).

Recently, touchable displays are widely used in most embedded systems. Although these devices naturally requires interactive user interfaces, most of interactive graphical user interface implementations usually need lots of system resources. We have presented Text-based User Interfaces (TUI) are more suitable for low-tier embedded systems. In this paper, we present an object-oriented implementation of the TUI windowing systems, also with object-oriented application programming interfaces. Our system is very simple and intuitive, with few amount of computing resource. It supports multiple windows, context menus, message boxes, and other user interface components, with graceful object-oriented application interfaces.

In this paper, we propose a high performance and low power unified 8 times 8 transform and 4 times 4 transform architecture for the H.264/AVC codec. The proposed architecture integrates 4 times 4 transforms and 8 times 8 transforms in the common architecture with reusability, which reduces complexity significantly. Also, we simplify the transform architecture and data flow to get the high performance merit without decomposition. Furthermore 8times8 transforms can be processed in 5 cycles by parallel processing. This way the transform architecture can process full-HD (1080p) video images at 27 MHz clock frequency, therefore we get the low power merit. As a result, the proposed architecture trades-off high performance and low power against chip area.