Search

Reduced precision computation is a key enabling factor for energy-efficient acceleration of deep learning (DL) applications. This article presents a 7-nm four-core mixed-precision artificial intelligence (AI) chip that supports four compute precisions--FP16, Hybrid-FP8 (HFP8), INT4, and INT2--to support diverse application demands for training and inference. The chip leverages cutting-edge algorithmic advances to demonstrate leading-edge power efficiency for 8-bit floating-point (FP8) training and INT4 inference without model accuracy degradation. A new HFP8 format combined with separation of the floating- and fixed-point pipelines and aggressive circuit/architecture optimization enables performance improvements while maintaining high compute utilization. A high-bandwidth ring protocol enables efficient data communication, while power management using workload-aware clock throttling maximizes performance within a given power budget. The AI chip demonstrates 3.58-TFLOPS/W peak energy efficiency and 26.2-TFLOPS peak performance for HFP8 iso-accuracy training, and 16.9-TOPS/W peak energy efficiency and 104.9-TOPS peak performance for INT4 iso-accuracy inference.

BACKGROUND The aim of this study was to examine osseous changes following lateral bone augmentation using a novel Ossix Volumax (OV) scaffold alone and compare it to combination therapy using freeze-dried bone allograft (FDBA) and resorbable collagen membrane (FDBA/CM). METHODS Thirty patients completed this 9-months prospective two-center cohort clinical trial. Before surgery and 9-months re-entry, linear measurements were performed, and impressions taken. Cone-beam computed tomography (CBCT) were done at baseline and 9 months. DICOM slice data were converted into volumetric images using 3D Slicer. Following 3D volumetric image construction, pre- and post-op Standard Triangle Language files were superimposed and volumetric data were extracted for a 10-mm region of interest. Linear measurements were compared similarly. RESULTS Baseline clinical parameters were similar in both groups (4.22 and 4.53 mm for OV and FDBA/CM at -2 mm, respectively). Following treatment, vertical distance from the stent had changed minimally (-0.36 and -0.12 mm, respectively). Similarly, lateral bone gain ranged from 0 to 0.4 mm, for both groups. To the contrary, the CBCT measurements showed a significantly greater increase in horizontal width in the control at -2 mm (0.95 ± 0.2 mm) compared with -0.62 mm for the OV (P = 0.000). Similar changes were observed at -5 mm (0.63 and -0.41 mm, respectively, P = 0.01). Sites volume had increased from 266 ± 149 mm3 to 360 ± 138 mm3 (P = 0.001) for FDBA/CM with negligible changes for OV (from 334 to 335 mm3 , P = 0.952). these between-group changes being statistically significant (P = 0.023). CONCLUSION FDBA/CM yielded better albeit moderate increase in the volume of the edentulous ridge, while OV scaffolds failed to produce similar results.

This investigation was designed to evaluate the reestablishment of bone-to-implant contact on infected dental implant surfaces following decontamination with an erbium, chromium:yttrium-scandium-gallium-garnet (Er,Cr:YSGG) laser and reconstructive therapy. Three patients presenting with at least one failing implant each were enrolled and consented to treatment with the Er,Cr:YSGG laser surface decontamination and reconstruction with a bone replacement allograft and a collagen membrane. The laser treatment was carried out at a setting of 1.5 W, air/water of 40%/50%, and pulse rate of 30 Hz. At 6 months, all three patients returned for the study. En bloc biopsy samples of four implants were obtained and analyzed. Two patients had excellent clinical outcomes, while one patient with two adjacent failing implants experienced an early implant exposure during the follow-up period. There was histologic evidence of new bone formation with two implant specimens and less bone gain with the others. Despite the small sample size, these were optimistic findings that suggested a positive role of Er,Cr:YSGG laser in debridement of a titanium implant surface to facilitate subsequent regenerative treatment. This investigation provides histologic evidence as well as encouraging clinical results that use of the Er,Cr:YSGG laser can be beneficial for treatment of peri-implantitis, but further long-term clinical studies are needed to investigate the treatment outcome obtained.

This investigation was designed to evaluate the healing response of 9.3-μm CO2 laser-assisted periodontal therapy. Five patients presenting with moderate to severe periodontitis, with an initial pocket depth (PD) ≥ 5 mm and with teeth predetermined to be surgically extracted, were enrolled and consented to treatment with full-mouth CO2 laser-assisted therapy. The laser treatment was carried out in the Ultraguide Mode at a setting of 0.25-mm spot size, with an average power of 0.65 to 1.15 watts and 20% mist. The laser tip was passed from the gingival margin and down apically to the base of the pocket with a sweeping L motion. The teeth were intensely scaled with piezoultrasonic instrumentation afterwards. A second pass of the laser tip was performed for the study teeth. At 9 months, all patients were clinically reevaluated. For sites with an initial pocket depth of ≥ 7 mm, a mean PD reduction of 3.97 ± 1.36 mm and a mean clinical attachment level gain of 3.54 ± 1.54 mm were achieved, resulting in a mean PD of 3.91 ± 0.77 mm. En bloc biopsy samples of four teeth were obtained and analyzed; two demonstrated histologic evidence of new bone formation while the other two healed with a long junctional epithelium with minimal inflammatory infiltrate. Further long-term clinical studies are needed to investigate the treatment stability obtained with a 9.3-μm CO2 laser compared to conventional surgical therapy. Nevertheless, the encouraging clinical results indicated that adjunctive use of the 9.3-μm CO2 laser-assisted periodontal therapy can be beneficial for treatment of periodontally compromised patients.

Sheet metal-based products serve as a major portion of the furniture market and maintain higher quality standards by being competitive. During industrial processes, while converting a sheet metal to an end product, new defects are observed and thus need to be identified carefully. Recent studies have shown scratches, bumps, and pollution/dust are identified, but orange peel defects present overall a new challenge. So our model identifies scratches, bumps, and dust by using computer vision algorithms, whereas orange peel defect detection with deep learning have a better performance. The goal of this paper was to resolve artificial intelligence (AI) as an AI landing challenge faced in identifying various kinds of sheet metal-based product defects by ALDB-DL process automation. Therefore, our system model consists of multiple cameras from two different angles to capture the defects of the sheet metal-based drawer box. The aim of this paper was to solve multiple defects detection as design and implementation of Industrial process integration with AI by Automated Optical Inspection (AOI) for sheet metal-based drawer box defect detection, stated as AI Landing for sheet metal-based Drawer Box defect detection using Deep Learning (ALDB-DL). Therefore, the scope was given as achieving higher accuracy using multi-camera-based image feature extraction using computer vision and deep learning algorithm for defect classification in AOI. We used SHapley Additive exPlanations (SHAP) values for pre-processing, LeNet with a (1 × 1) convolution filter, and a Global Average Pooling (GAP) Convolutional Neural Network (CNN) algorithm to achieve the best results. It has applications for sheet metal-based product industries with improvised quality control for edge and surface detection. The results were competitive as the precision, recall, and area under the curve were 1.00, 0.99, and 0.98, respectively. Successively, the discussion section presents a detailed insight view about the industrial functioning with ALDB-DL experience sharing.

Background: Antinuclear antibody pattern recognition is vital for autoimmune disease diagnosis but labor-intensive for manual interpretation. To develop an automated pattern recognition system, we established machine learning models based on the International Consensus on Antinuclear Antibody Patterns (ICAP) at a competent level, mixed patterns recognition, and evaluated their consistency with human reading. Methods: 51,694 human epithelial cells (HEp-2) cell images with patterns assigned by experienced medical technologists collected in a medical center were used to train six machine learning algorithms and were compared by their performance. Next, we choose the best performing model to test the consistency with five experienced readers and two beginners. Results: The mean F1 score in each classification of the best performing model was 0.86 evaluated by Testing Data 1. For the inter-observer agreement test on Testing Data 2, the average agreement was 0.849 (κ) among five experienced readers, 0.844 between the best performing model and experienced readers, 0.528 between experienced readers and beginners. The results indicate that the proposed model outperformed beginners and achieved an excellent agreement with experienced readers. Conclusions: This study demonstrated that the developed model could reach an excellent agreement with experienced human readers using machine learning methods.

This survey paper summarizes recent progress of deep learning circuits and systems technologies and contains four topics: hardware-centric deep learning algorithms, digital architectures, analog architectures, and system demonstrations. We present an overview of these four areas and introduce key contributions of papers in this special issue.

Low-precision computation is the key enabling factor to achieve high compute densities (T0PS/W and T0PS/mm2) in AI hardware accelerators across cloud and edge platforms. However, robust deep learning (DL) model accuracy equivalent to high-precision computation must be maintained. Improvements in bandwidth, architecture, and power management are also required to harness the benefit of reduced precision by feeding and supporting more parallel engines to achieve high sustained utilization and optimize performance within a given product power envelope. In this work, we present a 4-core AI chip in 7nm EUV technology that exploits cutting-edge algorithmic advances for iso-accurate models in low-precision training and inference [1, 2] and aggressive circuit/architecture optimization to achieve leading-edge power-performance. The chip supports fp16 (DLFIoat16 [8]) and hybrid-fp8(hfp8) [1] formats for training and inference of DL models, as well as int4 and int2 formats for highly scaled inference.

This work proposes a multi-criteria artificial bee colony (MABC) algorithm to optimize the energy consumption problem in wireless sensor networks. The approach uses the artificial bee colony (ABC) algorithm to discover sensor nodes in a network as a cluster header combination. Different nodes are dynamically selected according to their current status in the network. The purpose is to cluster sensor nodes in the network in such a way that nodes can transmit packets to their cluster header, and then identify the most energy efficient packet routing from the cluster headers to the Internet of Things (IoT) base station. The routing strategy takes into account nodes’ residual energy and energy consumption, routing distance, number of hops, and frequency, in order to assign decision scores to help the algorithm discover a better solution. The use case shows that the MABC algorithm provides energy-efficient packet routing, and thus extends the wireless sensor network lifespan, which is confirmed by the multi-criteria analysis evaluation of the candidate routing. The contribution of this research is its use of swarm intelligence algorithms in wireless sensor network routing, with a multi-criteria artificial bee colony algorithm used in a wireless sensor network to address the problem of fast convergence of the algorithm.

The goal of this multicenter randomized controlled study was to evaluate the effectiveness of a newly developed ionic-sonic electric toothbrush in terms of plaque removal and reduction of gingival inflammation. A total of 78 subjects from three dental centers were invited to join the study. They were randomized to receive either a manual toothbrush (control group) or an ionic-sonic electric brush (test group). Full-mouth prophylaxis and oral hygiene instructions based on the stationary bristle technique were provided 1 week prior to the baseline visit. At baseline and at each follow-up appointment, Plaque Index (PI) and Gingival Index (GI) were recorded. In addition, probing depth (PD) and bleeding on probing were recorded at baseline and at the last appointment (week 5). At completion of the study, subjects in the test group were given a questionnaire regarding their satisfaction with the toothbrush. Sixty-four subjects completed the study (control: 28; test: 36). The mean age of the subjects was 36.90 ± 12.19 years. No significant difference between the baseline and 5-week PD was found. Plaque removal efficacy and reduction in gingival inflammation were more significant for the test group at week 2. Both the control and test groups showed statistically significant improvement in PI and GI from baseline to week 5. The ionic-sonic toothbrush was more effective than manual toothbrush after a 1-week application.

The goal of the present study was to evaluate human histologic healing of dental implants with a unique triangular neck design that is narrower than the implant body. Four patients in need of full-mouth reconstruction were recruited and received several implants to support a full-arch prosthesis. In each patient, two additional customized reduced-diameter implants were placed, designated to be harvested after 6 months of submerged healing. The eight harvested implants were all placed in healed edentulous maxillary or mandibular ridges. These implants were O 3.5 × 8 mm in size, and the final osteotomy drill allowed for the creation of a gap up to 0.2 mm in size between the coronal aspect of the triangular implant neck and the surrounding bone. At the end of the healing period, the implants were retrieved with the surrounding bone. Microcomputed tomography (μCT) was performed before processing the biopsy samples for undecalcified histologic exampination. Bone-to-implant contact (BIC) was measured from the μCT data and from buccolingual/buccopalatal and mesiodistal central histologic sections. All implant gaps were filled by mature remodeled bone. The mean BICs of the BL/BP and MD sections were 64.45% ± 6.86% and 65.39% ± 10.44%, respectively, with no statistically significant difference. The mean 360-degree 3D BIC measured all over the implant surface was 68.58% ± 3.76%. The difference between the BIC measured on the μCT and on the histologic sections was not statistically significant. The positive histologic results of the study confirmed the efficacy of this uniquely designed dental implant.

Introduction: The gut microbiota is suspected to play a role in hypertension and hypertensive end organ damage. In the present study, we used germ-free mice to demonstrate that microbial colonization modulates the response to a hypertensive stimulus. Methods: Four-week-old male germ-free C57BL6/J littermates were randomized to remain germ-free (GF) or to receive microbiota transfer from SPF donor mice to achieve full colonization status (COL). At 12 weeks, Angiotensin (Ang) II was infused s.c. for 14 days (1.44mg/kg/d, osmotic minipumps) and 1% NaCl added to the drinking water; sham-treated mice served as control. After 14 days of AngII we assessed inflammation and organ damage. Results: Fecal bacterial load in COL mice was similar to SPF donor mice (qPCR). Shotgun metagenomic sequencing of fecal samples revealed hypertension-induced alterations in microbiome composition confirming previous reports. Serum metabolome analysis ( Biocrates MxP Quant 500) confirmed the absence of microbiota-dependent metabolites in GF. Interestingly, microbiota-dependent metabolites relevant for cardiovascular risk (TMAO, indoxyl sulfate) were elevated in hypertensive COL mice compared to sham-treated. Hypertensive kidney damage was aggravated in GF mice. However, marker genes for tubular damage ( Lcn2 ), inflammation ( Ccl2 ), and fibrosis ( Col1a3 ) showed a stronger increase in GF mice (fold changes [fc] COL vs. GF: 7.5 vs 11.0, 1.2 vs 3.3, 1.3 vs 2.2, respectively). Albuminuria (fc 2 vs 25) and histology for kidney fibrosis (fc 1.1 vs 1.4) confirmed the aggravated kidney damage in GF mice. Similarly, we observed an aggravated cardiac damage in GF mice. Flow cytometry of splenic lymphocytes showed that the adaptive immune response to AngII + 1% NaCl, as evidenced by Th17 (fc 1.4 vs 2) and CD8+ central memory cells, was intensified in GF mice. In vitro , naïve T cells isolated from GF mice more readily polarized into Th17 (26 ± 5%) compared to T cells from SPF mice (19 ± 1%). Conclusion: The bacterial colonization status has potent effects on the phenotypic response to a hypertensive stimulus, evident to varying degrees in hearts and kidneys. The inflammatory response and the end organ damage in GF compared to COL mice demonstrates the importance of the gut microbiota in hypertension.

A processor core is presented for AI training and inference products. Leading-edge compute efficiency is achieved for robust fp16 training via efficient heterogeneous 2-D systolic array-SIMD compute engines leveraging compact DLFloat16 FPUs. Architectural flexibility is maintained for very high compute utilization across neural network topologies. A modular dual-corelet architecture with a shared scratchpad and a software-controlled network/memory interface enables scalability to many-core SoCs and large-scale systems. The 14nm AI core achieves fp16 peak performance of 3.0 TFLOPS at 0.62V and 1.4 TFLOPS/W at 0.54V.

Although it is possible to quickly make travel route planning through many existing services, such as online map services or GPS, the shortest travel distance or the optimized path recommended by using only a single criterion is not satisfied the traveler’s needs. How to effectively find a reasonable path between point-of-interests and satisfy the requirements of users with different travel preferences will be a research topic worthy to explore. This study uses location-based social network and skyline query method to recommend travel routes to meet user preferences. We establish a traffic route network for various attractions based on the social networking data provided on Foursquare and Instagram and city bus open data. Then skyline query and top-k method are combined to recommend travel routes that cover different landscape categories and meet user preferences. The main contribution of this research is to provide travelers with a recommendation service that integrates social networks and traffic open data to provide suitable travel routes.

The wireless sensor network is the core of the Internet of Things. However, wireless sensors have some limitations and challenges, such as limited power and computing power, data storage, and network bandwidth, especially power requirements. How to find a way to program more flexible and faster according to the state of each sensing node in the network becomes an important issue. The software-defined network separates the control functions from the hardware devices, such as switches or routers, so that these hardware devices only have the data forwarding function, and the control software dynamically controls the flow of the network and data packets according to the network state and application requirements. In order to provide flexibility and adaptability, software-defined networks require a dynamic approach to solving and optimizing routing planning problems. This study will use the artificial bee colony algorithm to monitor the state of the sensor nodes in the software-defined network through the controller and take the best decision dynamically. Artificial bee colony algorithms are used to optimize wireless sensor networks and improve sensor node energy usage and data packets routing issues. The contribution of this research is to dynamically find the optimal routing path for the sensing nodes through the artificial bee colony algorithm, and improve the overall practicability and reliability of the wireless sensor network.

In this paper, the analysis and implementation of active bleeder for TRIAC dimmable LED driver is presented. The insufficient holding current problem occurs when LED driver is connected to TRIAC dimmer, causing TRIAC dimmer misfire and LED flicker. Therefore, a bleeder circuit is parallel-connected to provide TRIAC enough holding current. In the bleeder, the resistor is usually used to control the amount of current through the bleeder, and cause high power loss on the resistor. A bleeder circuit with an inductor, a capacitor and switches is used to provide TRIAC dimmer extra current to reduce the power dissipation by using a resistor. Finally, the hardware type of the bleeder circuit with input voltage of 90~135V is implemented, and able to applied on the TRIAC dimmable 7.2W LED driver. The bleeder can provide enough holding current, widening the conducting angle of TRIAC dimmer with no flicker and misfire occur.

Gingival hyperpigmentation in the anterior region in people with a high smile line is an esthetic concern. Gingival depigmentation can be achieved through various procedures. Most techniques have shown successful short-term results; however, recurrence is observed in 50% of patients after 2 to 4 years. A 46-year-old woman presenting with a large black zone of gingival hyperpigmentation in the anterior maxilla was treated with scalpel blade excision. At 25 years after the depigmentation procedure, the patient displayed no recurrence.

BACKGROUND The aim of this systematic review and meta-analysis was to compare the clinical efficacy of the early dental implant placement protocol with immediate and delayed dental implant placement protocols. METHODS An electronic and manual search of literature was made to identify clinical studies comparing early implant placement with immediate or delayed placement. Data from the included studies were pooled and quantitative analyses were performed for the implant outcomes reported as the number of failed implants (primary outcome variable) and for changes in peri-implant marginal bone level, peri-implant probing depth, and peri-implant soft tissue level (secondary outcome variables). RESULTS Twelve studies met the inclusion criteria. Significant difference in risk of implant failure was found neither between the early and immediate placement protocols (risk difference = -0.018; 95% confidence interval [CI] = -0.06, 0.025; P = 0.416) nor between early and delayed placement protocols (risk difference = -0.008; 95% CI = -0.044, 0.028; P = 0.670). Pooled data of changes in peri-implant marginal bone level demonstrated significantly less marginal bone loss for implants placed using the early placement protocol compared with those placed in fresh extraction sockets (P = 0.001; weighted mean difference = -0.14 mm; 95% CI = -0.22, -0.05). No significant differences were found between the protocols for the other variables. CONCLUSIONS The available evidence supports the clinical efficacy of the early implant placement protocol. Present findings indicate that the early implant placement protocol results in implant outcomes similar to immediate and delayed placement protocols and a superior stability of peri-implant hard tissue compared with immediate implant placement.

To increase disk I/O performance, Linux performs prefetching to read more data that is expected to be used in the near future. Nevertheless, the current Linux prefetching architecture is designed specifically for hard disk drives (HDDs). In recent decades, flash-based solid-state drives (SSDs) have been widely deployed in the storage systems, especially as caches forming a multi-tiered storage system, to boost I/O performance. Since the access characteristics of HDDs and SSDs are drastically different, the data prefetching scheme of Linux must be re-designed. Thus, in this paper, we proposed a new data prefetching architecture, called Cross-Layer Adaptive Caching (CLAP), for multi-tiered storage systems. CLAP adaptively determines the prefetching depth by analyzing the degree of data access sequentiality and the characteristics of the accessed storage devices. From the experimental results, CLAP can successfully improve the I/O access performance in multi-tiered storage systems.

Background In pancreatic cancer, toxicities associated with current chemotherapeutic regimens remain concerning. A modified combination of gemcitabine, S-1, and leucovorin (GSL) was used as the first-line treatment for newly diagnosed locally advanced or metastatic pancreatic adenocarcinoma patients. Methods GSL was administered every 2 weeks—intravenous gemcitabine 800 mg/m2 at a fixed-dose rate of 10 mg/m2/min on day 1 and oral S-1 (80–120 mg/day) plus leucovorin 30 mg twice daily on days 1–7. We retrospectively analyzed the feasibility of GSL and patient outcomes in three medical centers in Taiwan. Results Overall, 49 patients received GSL with a median follow-up of 24.9 months from May 2015 to March 2019. The median patient age was 68 years (range, 47–83 years), with a marginally higher number of females (57.1%). Among the 44 patients who underwent image evaluation, 13 demonstrated a partial response (29.5%) and 17 presented with stable disease (38.6%). The partial response rate and stable disease rate was 26.5% and 34.7%, respectively, in the intent-to-treat analysis. The median time-to-treatment failure was 5.79 months (95% C.I., 2.63–8.94), progression-free survival was 6.94 months (95% C.I., 5.55–8.33), and overall survival time was 11.53 months (95% C.I., 9.94–13.13). For GSL treatment, the most common grade 3 or worse toxicities were anemia (18.3%), neutropenia (6.1%), nausea (4.1%), and mucositis (4.1%). Treatment discontinuation was mostly due to disease progression (65.3%). Conclusions The modified GSL therapy can be a promising and affordable treatment for patients with advanced and metastatic pancreatic cancer in Taiwan. A prospective trial of modified GSL for elderly patients is currently ongoing in Taiwan.

Background: W e investigated the combined impact of abdominal obesity and low skeletal muscle mass on cardiovascular and total mortality in an elderly Asian population. Methods: A total of 1,485 elderly individuals (≥65 years) from Elderly Nutrition and Health Survey in Taiwan (1999–2000) were enrolled, and their survival status was followed using data from the National Death Registry. Skeletal muscle mass index (SMMI) was calculated by dividing skeletal muscle mass (kg) by height squared (m 2 ). Low skeletal muscle mass was defined as the first quartile of SMMI. Abdominal obesity (high triglycerides plus waist circumference [HTGWC]) was defined as triglycerides ≥150 mg/dL and w aist circumference ≥90 cm (men) and ≥80 cm (women). The Cox proportional hazard model was used to evaluate the combined impact of abdominal obesity and low SMMI on cardiovascular and total mortality. Results: During follow -up (median 9.2 years), one third (n = 493) of subjects died from any cause, of which 34% (n = 168) were cardiovascular-related. Total and cardiovascular mortality were 4.2 and 1.4 per 100 person-years, respectively. Low SMMI and HTGWC were independently associated with total mortality in men, but only low SMMI was significantly associated in women. Those with both HTGWC and low SMMI had the highest mortality risk, with the cardiovascular mortality risk increased by >6.8-fold and 3.2-fold in men and women, respectively, compared with controls having normal SMMI and TGWC. Conclusions: Elderly indi viduals with abdominal obesity and low skeletal muscle mass have higher all-cause and cardiovascular mortality risk.

Background Chylous leakage is a well-recognized but rare complication of head and neck surgery, affecting approximately 1% to 2.5% of head and neck dissections. It is a potentially life-threatening condition characterized by electrolyte imbalance, immunosuppression, delayed wound healing, risk of infection, and generalized sepsis. Management can be problematic and prolonged. Methods We present a case of refractory cervical chylous leakage after neck dissection treated with ultrasound-guided intranodal lymphangiography. Results Ultrasound-guided intranodal lymphangiography alone resulted in rapid and complete resolution of chylous leakage with minimal morbidity. Conclusion Based on our clinical experience and after a thorough literature review, we propose that ultrasound-guided intranodal lymphangiography with contrast agent could be considered a viable therapeutic option for persistent chylous leakages in selected patients. © 2015 Wiley Periodicals, Inc. Head Neck 38: E54–E60, 2016

The combination of growth in compute capabilities and availability of large datasets has led to a re-birth of deep learning. Deep Neural Networks (DNNs) have become state-of-the-art in a variety of machine learning tasks spanning domains across vision, speech, and machine translation. Deep Learning (DL) achieves high accuracy in these tasks at the expense of 100s of ExaOps of computation; posing significant challenges to efficient large-scale deployment in both resource-constrained environments and data centers. One of the key enablers to improve operational efficiency of DNNs is the observation that when extracting deep insight from vast quantities of structured and unstructured data the exactness imposed by traditional computing is not required. Relaxing the "exactness" constraint enables exploiting opportunities for approximate computing across all layers of the system stack. In this talk we present a multi-TOPS AI core [3] for acceleration of deep learning training and inference in systems from edge devices to data centers. We demonstrate that to derive high sustained utilization and energy efficiency from the AI core requires ground-up re-thinking to exploit approximate computing across the stack including algorithms, architecture, programmability, and hardware. Model accuracy is the fundamental measure of deep learning quality. The compute engine precision in our AI core is carefully calibrated to realize significant reduction in area and power while not compromising numerical accuracy. Our research at the DL algorithms/applications-level [2] shows that it is possible to carefully tune the precision of both weights and activations to as low as 2-bits for inference and was used to guide the choices of compute precision supported in the architecture and hardware for both training and inference. Similarly, distributed DL training's scalability is impacted by the communication overhead to exchange gradients and weights after each mini-batch. Our research on gradient compression [1] shows by selectively sending gradients larger than a threshold, and by further choosing the threshold based on the importance of the gradient we achieve achieve compression ratio of 40X for convolutional layers, and up to 200X for fully-connected layers of the network without losing model accuracy. These results guide the choice of interconnection network topology exploration for a system of accelerators built using the AI core. Overall, our work shows how the benefits from exploiting approximation using algorithm/application's robustness to tolerate reduced precision, and compressed data communication can be combined effectively with the architecture and hardware of the accelerator designed to support these reduced-precision computation and compressed data communication. Our results demonstate improved end-to-end efficiency of the DL accelerator across different metrics such as high sustained TOPs, high TOPs/watt and TOPs/mm2 catering to different operating environments for both training and inference.

A multi-TOPS AI core is presented for acceleration of deep learning training and inference in systems from edge devices to data centers. With a programmable architecture and custom ISA, this engine achieves >90% sustained utilization across the range of neural network topologies by employing a dataflow architecture and an on-chip scratchpad hierarchy. Compute precision is optimized at 16b floating point (fp 16) for high model accuracy in training and inference as well as 1b/2b (bi-nary/ternary) integer for aggressive inference performance. At 1.5 GHz, the AI core prototype achieves 1.5 TFLOPS fp 16, 12 TOPS ternary, or 24 TOPS binary peak performance in 14nm CMOS.

Background The effect of different surgical techniques for ridge preservation on soft tissue parameters has seldom been investigated. The objective of this study was to compare the effect of two different ridge preservation techniques on soft and hard tissue dimensions. Methods Thirty patients requiring tooth extraction were randomly allocated to either control group C (allograft covered with a non-crosslinked collagen membrane with primary closure) or experimental group E (allograft covered with cross-linked collagen membrane left exposed). Sites were surgically re-entered at 6 months. Soft and hard tissue measurements, cone beam computed tomography (CBCT), and cast measurements were taken at baseline and 6 months. Results Twenty-eight patients were included in this analysis. When the two treatment groups were compared, the width of the buccal keratinized tissue in the E group showed an increase of 0.43 ± 0.42 mm compared to net loss of 1.57 ± 0.51 mm for the C (P = 0.006). Similarly, buccal tissue thickness has increased in the E group 0.46 ± 0.22 mm compared to a loss of 0.15 ± 0.23 mm in the C group (P = 0.068). Volumetric assessment of the changes in the alveolar ridge for the E group showed a slight decrease (68.3 ± 17 mm3) whereas the C group has experienced almost double this loss (107.5 ± 11 mm3; P = 0.07). Crestal width, measured on the CBCT scan, has shown significant reduction in the C group (4.18 ± 0.56 mm) compared to only 1.74 ± 0.4 mm in the E group (P = 0.003). Conclusion Crosslinked collagen membrane with allograft placed intentionally non-submerged resulted in better preservation of the keratinized tissues (width and thickness) with similar and at times better osseous preservation following extraction.

Deep Neural Networks (DNNs) have emerged as a powerful and versatile set of techniques to address challenging artificial intelligence (AI) problems. Applications in domains such as image/video processing, natural language processing, speech synthesis and recognition, genomics and many others have embraced deep learning as the foundational technique. DNNs achieve superior accuracy for these applications using very large models which require 100s of MBs of data storage, ExaOps of computation and high bandwidth for data movement. Despite advances in computing systems, training state-of-the-art DNNs on large datasets takes several days/weeks, directly limiting the pace of innovation and adoption. In this paper, we discuss how these challenges can be addressed via approximate computing. Based on our earlier studies demonstrating that DNNs are resilient to numerical errors from approximate computing, we present techniques to reduce communication overhead of distributed deep learning training via adaptive residual gradient compression (AdaComp), and computation cost for deep learning inference via Prameterized clipping ACTivation (PACT) based network quantization. Experimental evaluation demonstrates order of magnitude savings in communication overhead for training and computational cost for inference while not compromising application accuracy.

In this work, the effects of nitrous oxide plasma treatment on the negative bias temperature stress of indium tin zinc oxide (ITZO) and indium gallium zinc oxide (IGZO) thin film transistors (TFTs) were reported. ITZO TFTs were more suitable for the back channel etched-type device structure because they could withstand both Al- and Cu-acid damage. The initial threshold voltage range could be controlled to within 1 V. The root cause of poor negative bias temperature stress for ITZO was likely due to a higher mobility (∼3.3 times) and more carbon related contamination bonds (∼5.9 times) relative to IGZO. Finally, 65″ active-matrix organic light-emitting diode televisions using the ITZO and IGZO TFTs were fabricated.

We have investigated the role of temperature in the degradation of GaN High-Electron-Mobility-Transistors (HEMTs) under high-power DC stress. We have identified two degradation mechanisms that take place in a sequential manner: the gate leakage current increases first, followed by a decrease in the drain current. Building on this observation, we demonstrate a new scheme to extract the activation energy (Ea) of device degradation from step-temperature measurements on a single device. The Ea’s we obtain closely agree with those extracted from conventional accelerated life test experiments on a similar device technology.

Background and aim Atrial fibrillation (AF) is an important cardiovascular disease in the elderly. The association between hyperuricemia and AF is unclear. Therefore, we aimed to investigate the prospective relationship between uric acid and development of AF in a nationally representative cohort of elderly people. Methods and results A total of 1485 elderly people (age ≥ 65 yrs) from the Elderly Nutrition and Health Survey in Taiwan (1999–2000) were without AF on “electrocardiography” at baseline. Incident AF events (International Classification of Diseases, Ninth Revision, Clinical Modification, ICD-9-CM: 427.31) were identified using data from the National Health Insurance Dataset. Hyperuricemia was defined as levels of uric acid >7.0 mg/dL in men and 6.0 mg/dL in women. A Cox proportional hazards model was used to evaluate the association between hyperuricemia and incident AF. The follow-up period was from 1999 to 2000 to 2008. During the follow-up period (median: 9.16 yrs), 90 AF events occurred (44 in men and 46 in women). Older age, elevated systolic blood pressure, being an ex-smoker, and high uric acid were positively associated with incident AF. Hyperuricemia was positively associated with incident AF in normotensive (age-adjusted hazard ratio (HR): 2.65 and 95% confidence intervals: 1.05–6.69), but not in (1.20:0.74–1.94) hypertensive individuals (systolic blood pressure ≥130 or diastolic blood pressure ≥85 or using hypertensive medicine). A significant association between hyperuricemia and AF (3.78; 1.24–11.59) remained after adjusting for other potential confounders among normotensive older persons. Conclusion Hyperuricemia is associated with the development of AF in elderly people with normal blood pressure.

Deep Neural Networks (DNNs) have emerged as a powerful and versatile set of techniques showing successes on challenging artificial intelligence (AI) problems. Applications in domains such as image/video processing, autonomous cars, natural language processing, speech synthesis and recognition, genomics and many others have embraced deep learning as the foundation. DNNs achieve superior accuracy for these applications with high computational complexity using very large models which require 100s of MBs of data storage, exaops of computation and high bandwidth for data movement. In spite of these impressive advances, it still takes days to weeks to train state of the art Deep Networks on large datasets - which directly limits the pace of innovation and adoption. In this paper, we present a multi-pronged approach to address the challenges in meeting both the throughput and the energy efficiency goals for DNN training.

Background and aim: Body mass index (BMI) has a U-shaped relationship with mortal- ity among the elderly, in contrast to the general adult population. Skeletal muscle mass may be more appropriate than BMI for classifying mortality risk among the elderly. We investigated the relationship between skeletal muscle mass and mortality among elderly Chinese persons. Method and results: A total of 1512 elderly from the Nutrition and Health Survey in Taiwanese Elderly (1999e2000) was enrolled, and the survival status was followed using data from the Na- tional Death Registry. The skeletal muscle mass index (SMMI) was calculated by dividing skeletal muscle mass by height in meters squared. The Cox proportional hazard model was used to esti- mate the association between SMMI and mortality. During the follow-up (average time: 7.9 years), one-third elderly died (n Z 506) by any cause and 25% of them was cardiovascular mor- tality (ICD-9-CM: between 390 and 459). The total mortality and cardiovascular mortality were 4.23 and 1.07 per 100 person-years. Elderly participants with the lowest SMMI had the highest total mortality and cardiovascular mortality among the four quartiles (6.72, 3.76, 3.25 and 3.50 per 100 PY for total mortality; 1.81, 0.76, 0.87, 0.93 for cardiovascular mortality). Those with a low (1st quartile) SMMI had a 2-fold increase in total mortality (1.96; 1.63e2.35) and cardiovas- cular mortality (2.16; 1.51e3.08) risk compared to those with a normal (2nd, 3rd, or 4th quartile) SMMI. Conclusions: The threshold relationship between SMMI and mortality is contrast to the reverse J- shaped relationship between BMI and total mortality. Therefore, skeletal muscle mass measure- ment may be considered with a high priority to identify elderly individuals with a high mortality risk.

We reported amorphous Indium-Gallium-Zinc-Oxide (a-IGZO) TFT OLED TV panel development in AUO. The Generation 6 (G6) threshold voltage uniformity can be lower than 0.8V. The RGB side-by-side OLED patterning was realized by fine metal mask. By the application of a source follower connection pixel compensation circuit, the panels driven by AUO engine show excellent characteristics. For future product performance requirement, amorphous Indium-Tin-Zinc-Oxide (a-ITZO) TFT was developing, high mobility as 33.2 cm2/Vs was achieved. Metal-organic chemical vapor deposition (MOCVD) system was also used to improve the mobility of IGZO TFT. To improve manufacturing efficiency of OLED process, inkjet printing technology is also under development in AUO.

We reported a 65 inches amorphous oxide thin film transistors active-matrix organic light-emitting diode television panel. The thin film transistor shows an excellent characteristic. The long range threshold voltage uniformity is 0.34V. The side by side organic light-emitting diode device is realized by fine metal mask. The dam and fill encapsulation method shows a simple process procedure and high stability.

This paper investigates degradation behavior induced by the self-heating effect for InGaZnO (IGZO) thin-film transistors (TFTs). Both the surrounding oxide and other thermal insulating materials, as well as the low thermal conductivity of the InGaZnO layer itself, cause the self-heating effect in InGaZnO TFTs. The heated channel layer enhances the threshold voltage shift, and the evolution of threshold voltage shift is found to be dominated by charge-trapping effects. Moreover, a nonuniform distribution of channel carrier concentration leads to an uneven temperature distribution throughout the IGZO active layer, which results in the asymmetrical degradation behavior after the self-heating operation. Further verifications indicate that the degree of the threshold voltage shift is only dependent on stress power, regardless of stress Vg, Vd, and channel length. Further, two-stage dependence of the threshold voltage shift on dynamic stress frequency is found.

STEAP1 (six transmembrane epithelial antigen of the prostate 1) is a transmembrane protein that functions as a potential channel or transporter protein. It is overexpressed in certain cancers and is viewed as a promising therapeutic target. However, the prognostic role of STEAP1 is still controversial, and no role for STEAP1 has yet been indicated in colorectal cancer. The aim of this study was to investigate the possible association of STEAP1 expression with colorectal cancer prognosis. STEAP1 expression was analyzed by immunohistochemical staining of a tissue array of 165 cancer specimens from primary colorectal cancer patients. The mean and medium follow-up times after surgery were 5.1 and 3.9 years, respectively. A total of 139 patients died during the 13 years of follow-up in the survey period. The prognostic value of STEAP1 with respect to overall survival was analyzed by Kaplan-Meier analysis and Cox proportional hazard models. In total, 164 samples displayed detectable STEAP1 expression in the cytoplasm and membrane. Low STEAP1 expression was correlated with poor overall survival (five-year survival: 33.7% vs. 57.0%, low expression vs. high expression, p = 0.020). Accordingly, multivariate analysis identified low STEAP1 expression as an independent risk factor (hazard ratio = 1.500, p = 0.018), especially in elderly patients or those with late stage cancers, late T values, and early N values. We suggest that analysis of STEAP1 expression by immunohistochemical staining could serve as an independent prognostic marker for colorectal patients. This finding should be validated by other investigative groups.

ThispaperinvestigatestheeffectofDCandACbias-stressinduceddegradationbehaviorinamorphousInGaZnOthin-filmtransistors (TFTs)underlightillumination.Draincurrent-gatevoltage(ID-VG)aswellascapacitance-voltage(C-V)measurementsareemployed to analyze the degradation mechanism. Illuminated DC stress leads to not only a negative parallel shift but also a C-V curve distortion at the off-state. This can be attributed to barrier-lowering near the drain side due to the asymmetrical hole-trapping effect. To further verify the origin of the degradation behavior, AC bias with identical stress voltage is imposed on either the gate terminal or the drain terminal. It is deduced that the hole-trapping phenomenon near the drain side is dominated by the voltage across the gate and drain, and is responsible for the degradation characteristic after stress carried out under light illumination.

Over the past two decades, studies have widely examined the effectiveness of pay-for-performance (P4P) programs by conducting biochemical tests and assessing complications; however, the reported effectiveness of such programs among participants selected through purposeful sampling is controversial. Therefore, the objective of the current study was to analyze the effectiveness of a P4P program on patients' prognoses, including hospitalization for chronic diabetic complications, and all-cause mortality during specific follow-up years by using a nationwide population-based database in Taiwan. Based on 125,315 newly diagnosed type 2 diabetes patient cohort during 2002-2006, two control sets were designed by propensity-score-matching strategy according to participation of P4P program and followed up to 2012. The results indicated that full participants demonstrated the lowest risks of developing complications and all-cause mortality compared with nonparticipants. These findings confirm the long-term effect of P4P programs on full participants and reveal that this effect is not due to confounding variables. The results indicate the importance of performance management and adherence to interventions for patients with chronic diseases in a long-term observation. Comprehensive and continuous care is suggested to improve patient prognosis and quality of care.

Background: Impaired cognition increases mortality in the aged. It is unclear how dietary quality might affect this relationship. Objective: To examine how dietary diversity and cognition might interact to determine survival. Design: In a Nutrition and Health Survey in Taiwan (NAHSIT 1999-2000), 1,839 representative elderly were followed for mortality up to 10 years. The dietary quality measure was a dietary diversity score (DDS, range: 0-6) to present six food groups (dairy, meat, rice and grains, fruit, vegetable,fat and oil) derived from a 24-h dietary recall. Cognitive function was evaluated by the validated Short Portable Mental Status Questionnaire (SPMSQ). Results: Those with cognitive impairment (SPMSQ≥3 errors) had 2.56 (95% confidence intervals (CI), 1.99- 3.28) times the all-cause-mortality risk of those with intact cognition. After control for potential confounders, the adjusted hazard ratio (HR) remained significant (1.46, 95% CI: 1.06-2.02). Significant interactions for DDS and cognition were found (p

The plasma-process-induced damage (PPID) of low-temperature polycrystalline silicon (poly-Si) thin-film transistors (TFTs) during the etching of the poly-Si film was investigated in this paper. The results reveal the relationship between the device degradation and the PPID during TFT liquid-crystal-display fabrication. This degradation is caused in part by the damage at the edge of the poly-Si film in plasma exposure. The trapped-state densities Ntrap are measured to clarify the relationship between instability and plasma etching damages. The plasma-process condition substantially affects the PPID of the poly-Si etching process. The main mechanism is the generation of charge trapping states at the poly-Si grain boundary in the damaged edge of the TFT channel active region. The electrical recovery from the plasma damage is also studied with various postetching treatments. Hydrogen-base plasma treatment and laser anneal process are revealed to improve the device characteristics due to reduction of charge trapping states.

We present the light-on results of our 65-inch organic ***light-emitting diode (OLED) panel fabricated by the inkjet printing (IJP) process. The panel uses solution-processable OLED material for the hole injection layer (HIL), the hole transporting layer (HTL) and the emitting layers (EML) for the red, green and blue colors. It is found that fine-tuning the vacuum dry speed after the inkjet printing process can improve the uniformity of the pixels. The results show that we can successfully fabricate a 65-inch OLED panel with a high degree of pixel uniformity.

Organic light‐emitting devices (OLEDs) have shown superior characteristics and are expected to dominate the next‐generation flat‐panel displays. Active‐matrix organic light‐emitting diode (AMOLED) displays, however, have stringent demands on the performance of the backplane. In this paper, the development of thin‐film transistors (TFTs) based on indium gallium zinc oxide (IGZO) on both Gen 1 and 6 glasses, and their decent characteristics, which meet the AMOLED requirements, are shown. Further, several display prototypes (e.g., 2.4” AMOLED, 2.4” transparent AMOLED, and 32” AMLCD) using IGZO TFTs are demonstrated to confirm that they can indeed be strong candidates for the next‐generation TFT technology not only of AMOLED but also of AMLCD (active‐matrix liquid crystal display).

Device-level simulation capabilities have been developed to investigate low-frequency noise behavior in p-type Si0.7Ge0.3/Si heterostructure MOS (SiGe p-HMOS) transistors. The numerical model is based on the impedance field method; it accounts for a trap-induced carrier number fluctuation, a layer-dependent correlated mobility fluctuation, and a Hooge mobility fluctuation in the buried and parasitic surface channels, respectively. Simulations based on such models have been conducted for SiGe p-HMOS transistors, and the results have been carefully correlated with experimental data. Quantitative agreement has been obtained in terms of the noise level dependence on gate biases, drain currents, and body biases, revealing the important role of the dual channels in the low-frequency noise behavior of SiGe p-HMOS devices.

Introduction Population aging is set to increase the elder dependency ratio, causing a rapid increase in healthcare expenditures and financial burden on the government. This study aims to construct an aging score from age-related diseases and to perform longitudinal analyses to examine the association between personal income and aging. Methods A total of 86,838 subjects drawn from a community with 384,617 residents in northern Taiwan were examined. Personal income and aging-related diseases were measured from the National Health Insurance database during 1997–2010. Multiple linear regression and stratified analyses were used. Results Our results indicate that subjects with low personal income (defined as monthly income ≤610 USD) can have high aging scores as calculated by selected aging-related diseases. In stratified analyses, moreover, subjects with low personal income tended to have higher aging scores across different anthropometric groups. The association was particularly evident in subjects >60 years old, and they were noted across multiple strata, including those of catastrophic illness as well as type of primary care setting. Conclusions This study found evidence for association between low income and aging in a population with compulsory health insurance program, where accessibility barriers were low in terms of health care resources and personal socioeconomic status. The economic inequity causes difference in aging process, which deserves for future interventions.

A fast two-point measurement methodology, applicable to nano-scale devices, is introduced to separate electron trapping (Not, e) from interface state contributions (Nit) in hot carrier (HC) induced ΔVt in n-type RMG SOI FinFETs. It is shown that Not, e component is comparable to or dominates over Nit for high-Vg (≥ Vd) hot carrier stress (HCS). The time power-law exponent for HC induced Not, e is larger than 0.2, indicating the observed Not, e is partly due to injection of hot carriers instead of purely parasitic PBTI effect caused by the cold carriers.

With rapid development of the Internet, e-learning system has become more and more popular. Currently, to solve the issue of sharing and reusing of teaching materials in different e-learning system, Sharable Content Object Reference Model (SCORM) is the most popular standard among existing international standards. In SCORM standard, the Sequencing and Navigation (SN) defines the course sequencing behavior, which controls the sequencing, selecting and delivering of a course, and organizes the content into a hierarchical structure, namely Activity Tree (AT). However, the structures with complicated sequencing rules of Activity Tree (AT) in SCORM make the design and creation of course sequences hard. Therefore, how to provide a user-friendly authoring tool to efficiently construct SCORM compliant course becomes an important issue. However, before developing the authoring tool, how to provide a systematic approach to analyze the sequencing rules and to transform the created course into SCORM compliant are our concerns. Therefore, in this paper, based upon the concept of Object Oriented Methodology (OOM), we propose a systematic approach, called Object Oriented Course Modeling (OOCM), to construct the SCORM compliant course. High-Level Petri Nets (HLPN), which is a powerful language for system modeling and validation, are applied to model the basic sequencing components, called Object-Oriented Activity Tree (OOAT), for constructing the SCORM course with complex sequencing behaviors. Every OOAT as a middleware represents a specific sequencing behavior in learning activity and corresponding structure with associated sequencing rules of AT in SCORM. Thus, these OOATs can be efficiently used to model and construct the course with complex sequencing behaviors for different learning guidance. Moreover, two algorithms, called PN2AT and AT2CP, are also proposed to transform HLPN modeled by OOATs into a tree-like structure with related sequencing rules in Activity Tree (AT) and package the AT and related physical learning resources into a SCORM compliant course file described by XML language, respectively. Finally, based upon the OOCM scheme, a prototypical authoring tool with graphical user interface (GUI) is developed. For evaluating the efficiency of the OOCM approach compared with existing authoring tools, an experiment has been done. The experimental results show that the OOCM approach is workable and beneficial for teachers/instructional designers.

Background Cdk1 (cyclin-dependent kinase 1) is critical regulator of the G2-M checkpoint. Cyclin-dependent kinase pathways are considered possible targets for cancer treatment; however, the prognostic role of Cdk1 in colorectal cancer is still controversial. Therefore, we attempted to determine the impact of Cdk1 on the clinical outcome of colorectal cancer patients to further identify its role in colorectal cancer. Methods Cdk1 immunoreactivity was analyzed by immunohistochemistry (IHC) in 164 cancer specimens from primary colorectal cancer patients. The medium follow-up time after surgery was 3.7 years (range: 0.01 to 13.10 years). The prognostic value of Cdk1 on overall survival was determined by Kaplan-Meier analysis and Cox proportional hazard models. Results All samples displayed detectable Cdk1 expression with predominant location in the cytoplasm and nucleus. A high Cdk1 nuclear/cytoplasmic (N/C) expression ratio was correlated with poor overall survival (5-year survival rate: 26.3% vs 46.9%, N/C ratio ≥1.5 vs N/C ratio

AMOLED displays based on metal oxide film as active channel layer are fabricated. To achieve cost-competitive process in OLED manufacturing, solution-processed HIL, HTL and pixilated EML (RGB) layers followed by evaporation process for other layers are demonstrated in this work. Meanwhile, the AMOLED backplanes possessing low power consumption and high frequency were fabricated on metal oxide TFTs with high mobility (23.13 cm2/Vs) technology. A 14” qHD AMOLED in 79 ppi-resolution was realized with IGZO based active matrix panel.

We demonstrated a defect free 32-inch large scaled amorphous indium-gallium-zinc-oxide based active-matrix organic light-emitting diode television panel. The panel achieves good image quality, such as high contrast ratio, a wide viewing angle, and fast response time, and the module thickness is only 3 mm. The thin film transistor shows excellent characteristic and the compensation pixel circuit is used. The side by side organic light-emitting diode device is realized by fine metal mask.