In order to facilitate the deployment of diverse deep learning models while maintaining scalability, modern DNN accelerators frequently employ reconfigurable structures such as Network-on-Chip (NoC) and multi-level on-chip memory hierarchy. To achieve high energy efficiency, it is imperative to store intermediate DNN-layer results within the on-chip memory hierarchy, thereby reducing the need for off-chip data transfers to/from the DRAM memory.Two well-established optimization techniques, node fusion and loop tiling, have proven effective in retaining temporary results within the on-chip buffers, commonly used to minimize off-chip DRAM accesses. In this paper, we introduce MultiFuse, an infrastructure designed to automatically explore multiple DNN layer node fusion techniques, enabling optimal utilization of the on-chip multi-level memory hierarchy.Experimental results demonstrate the effectiveness of our retargetable infrastructure, which outperforms Ansor’s algorithm. Our exploration algorithm achieves a remarkable 70% reduction in Energy-Delay Product (EDP) while gaining a 67x speedup in search time when executing the data-intensive MobileNet model on a single DNN accelerator.

This article proposes a new method for the tracking of three-dimensional (3-D) hand palms from the whole human standing body using fuzzy rule-based fusion of explainable fuzzy features from a monocular video. The characteristics of this method include visually and linguistically explainable fuzzy features and rules and computational efficiency. This article first tracks the 2-D palms using the following four fuzzy features: optical flows; the degree of a pixel in the foreground; skin color information; and the search area around a hand palm candidate from a segmented body. Afterward, a fuzzy system (FS) is proposed to fuse the four fuzzy features to estimate the 2D- palm positions. Localization of the elbows is based on the estimated palm locations, human body skeletons, and body contour. The 2-D palms and elbows are tracked using a modified particle filter. To estimate the depth of each palm, the locations of the palm and elbow are fed as inputs to a neural FS. The 3-D palm tracking result is applied to a robot upper-body imitation system. Experiments with comparisons of different hand palm tracking methods are performed to verify the real-time computational ability and accuracy of the proposed method.

Optical flow algorithm, which estimates the motion detection of consequent video frames, is widely used in surveillance system, Advanced Driver Assistance Systems (ADAS) and object movement estimation in scene analysis. Among different optical flow algorithms, Farneback version provides a better accuracy and brightness-change-resistant displacements by estimating the flow from polynomial domain rather than intensive maps. However, high computation complexity and inconsistent data access patterns make it difficult to be implemented on a hardware platform. In this work, we present a micro-architecture design of Farneback optical flow, which is flexible for optimization with high level Synthesis (HLS) tools. The original software-based implementation was decomposed into functional blocks to balance latency of different stages and flows of data were rearranged to accommodate better memory access patterns. The data flow arrangement is based on a proposed backtrace mechanism, where DRAM accesses of polynomial coefficients in current frame makes consistent traffic patterns, and therefore make it possible to integrate more functional blocks into a deeper pipeline. For several micro-architecture design versions, we demonstrate options of fixed and floating points, optimization techniques such as multiple DMAs and different levels of pipeline integration. We implemented our design on Zedboard Mini-ITX 7045. The results show a 17x end-to-end speedup against a naive HLS version with an image size of 160x120. Considering only the hardware-accelerated part, our FPGA implementation is 40x faster than the naive HLS version with only 50% of the FPGA hardware resources.

The use of the badminton stroke strategy in the evenly matched game is often the key to victory. In this work, a smart racket based on wearable sensors is proposed to collect the data of swing of badminton. A cell phone APP with machine learning techniques is implemented to record stroke types automatically. In each stroke hit event, this prototype system uses Bluetooth earphone to collect the sound for detecting the accuracy time. It uses the data of IMU in each stroke for determining stroke type. Compared to EMU only solution, the system will reduce the false count of stroke hit. Using cloud techniques could record the training and game record in a long period. Overall the accuracy of stroke hit event is almost 100% by using voice print. The data of EMU is classified by Random Forest or SMO. The accuracy for personal model is 95.91%, and it is 7932% for general model. We develop a stroke record system which is combined with Wearable sensor, Mobile platform and Cloud service.

In Machine-to-Machine (M2M) communications, devices can automatically communicate with each other without any intervention by humans. However, LTE-A is primarily designed and optimized for Human-to-Human (H2H) communications. Even though LTE-A has defined Machine-Type Communications (MTC), several parts are still undefined or follow the existing H2H standards. Specifically, the existing paging mechanism in LTE-A will substantially increase the computational load and signaling overhead of the system when there are many M2M devices. Therefore, we propose an effective solution to mitigate this problem. Compared with the traditional paging mechanism defined in LTE-A, the proposed solution can save more energy for M2M devices, reduce signaling overhead and alleviate the network's load significantly.

Collecting data from a tremendous amount of Internet-of-Things (IoT) devices for next generation networks is a big challenge. A large number of devices may lead to severe congestion in Radio Access Network (RAN) and Core Network (CN). 3GPP has specified several mechanisms to handle the congestion caused by massive amounts of devices. However, detailed settings and strategies of them are not defined in the standards and are left for operators. In this paper, we propose two congestion control algorithms which efficiently reduce the congestion. Simulation results demonstrate that the proposed algorithms can achieve 20~40% improvement regarding accept ratio, overload degree and waiting time compared with those in LTE-A.

Tocolysis is a treatment to prevent premature labor and thus can reduce the happening of perinatal and neonatal mortality. In developed and developing countries, the increasing of elderly primigravidae usually accompanies the increase in the preterm birth rate and therefore raises the need for tocolysis. To fulfill the increasing need, more tocolytic centers have been established, such as the one in Mackay Memorial Hospital, Taiwan. However, as emergent situations happen and doctors in charge are not on-site, a tele-cardiotocograph system is needed to help doctors to give proper advice by providing real-time CTG data. Currently, the solution practiced in Mackay Memorial Hospital is to send photos of the cardiotocogram to doctors by email or MMS. In this work, the team from NCTU and ITRI collaborated with Maya, Inc. to develop a remote vital sign access system called FetalCare to improve the solution in Mackay Memorial Hospital by utilizing mobile technology. With the developed system, doctors can remotely access real-time physiological data of patients via our app.

This poster proposes an IoT-enabled mobile e-learning platform to enhance the performance of a learning process (i.e., assembling of fragmented learning time for a trainee) and the efficiency of learning material management. The informatics framework and implementation methodology for the proposed platform are described in the poster. The proposed platform has been used in the emergency department of National Taiwan University Hospital (NTUH) for training junior ultrasound physicians.

Network-wide traffic measurement is of interest to network operators to uncover global network behavior for the management tasks of traffic accounting, debugging or troubleshooting, security, and traffic engineering. Increasingly, sophisticated network measurement tasks such as anomaly detection and security forensic analysis are requiring in-depth fine-grained flow-level measurements. However, performing in-depth per-flow measurements (e.g., detailed payload analysis) is often an expensive process. Given the fast-changing Internet traffic landscape and large traffic volume, a single monitor is not capable of accomplishing the measurement tasks for all applications of interest due to its resource constraint. Moreover, uncovering global network behavior requires network-wide traffic measurements at multiple monitors across the network since traffic measured at any single monitor only provides a partial view and may not be sufficient or accurate. These factors call for coordinated measurements among multiple distributed monitors. In this paper, we present a centralized optimization framework, LEISURE (Load-EqualIzed meaSUREment), for load-balancing network measurement workloads across distributed monitors. Specifically, we consider various load-balancing problems under different objectives and study their extensions to support both fixed and flexible monitor deployment scenarios. We formulate the latter flexible monitor deployment case as an MILP (Mixed Integer Linear Programming) problem and propose several heuristic algorithms to approximate the optimal solution and reduce the computation complexity. We evaluate LEISURE via detailed simulations on Abilene and GEANT network traces to show that LEISURE can achieve much better load-balanced performance (e.g., 4.75× smaller peak workload and 70× smaller variance in workloads) across all coordinated monitors in comparison to a naive solution (uniform assignment) to accomplish network-wide traffic measurement tasks under the fixed monitor deployment scenario. We also show that under the flexible monitor deployment setting, our heuristic solutions can achieve almost the same load-balancing performance as the optimal solution while reducing the computation times by a factor up to 22.5× in Abilene and 800× in GEANT.

In this paper, a high throughput context-based adaptive variable length code (CAVLC) encoder design with two-path parallel coefficients procedure is developed for H.264/AVC with block based pipelined and parallel processing schemes. By the two-pipe parallel coefficients processing, the proposed block processing based CAVLC encoder only requires 65 encoding cycles/macroblock (MB) on average. Compared with previous encoder designs, the throughput of the proposed architecture is improved over from 23% to 585%. The proposed encoding core modules have smaller area-time (AT) product performance than other designs. By VLSI implementations, the proposed CAVLC encoder can achieve Digital Cinema 4K, i.e. 4096×2160p@37Hz real-time video encoding.

In this letter, multiple forward and inverse fast algorithm based transforms and their hardware-sharing design for 2 × 2, 4 × 4, and 8 × 8 transforms in H.264/AVC, and the 8 × 8 transform in audio video coding standard, 4 × 4 and 8 × 8 transforms in VC-1, and DCT/IDCT in MPEG-1/2/4 are developed with a low hardware cost for multistandard video coding applications. Compared with the directly combined fast transforms without shares, the proposed low-cost 1-D architecture reduces shifters by 67%, adders by 73%, and gate counts by 53.4%. The hardware-sharing efficiencies of shifters and adders in the proposed 1-D transform design are 32% and 25% more than those in the previous design, respectively. By 0.18-μm CMOS technology, the proposed 2-D transform architecture has less normalized power per mode and larger normalized hardware efficiency than the previous multiple-standard designs. The cost-effective 2-D full pipelined transform achieves multistandard real-time 1080HD at 60-Hz video encoding and decoding applications.

1/2 network capacity is often believed to be the limit of worst-case throughput for mesh networks. However, this paper provides a new worst-case throughput bound, which is higher than 1/2 network capacity, for odd radix two-dimensional mesh networks. In addition, we propose a routing algorithm called U2TURN that can achieve this worst-case throughput bound for odd radix meshes. For even radix meshes, we prove that U2TURN achieves the optimal worst-case throughput, namely, half of network capacity. U2TURN considers all routing paths with at most 2 turns and distributes the traffic loads uniformly in both X and Y dimensions. Theoretical analysis and simulation results show that U2TURN outperforms existing routing algorithms in worst-case throughput. Moreover, U2TURN achieves good average-throughput at the expense of approximately 1.5× minimal average hop count. For asymmetric meshes, we further propose an algorithm called “U2TURN-A” and provide theoretical analysis for different algorithms. Both theoretical analysis and simulation show that U2TURN and U2TURN-A outperform existing algorithms VAL, DOR and O1TURN in both worst-case and average throughput for asymmetric meshes.

During recent years cloud service providers have successfully provided reliable and flexible resources to cloud users. For example Amazon Elastic Block Store (Amazon EBS) and Simple Storage Service (Amazon S3) provides users storage in the cloud. Despite the tremendous efforts cloud service providers have devoted to the availability of their services, the interruption is still inevitable. Therefore just as an Internet service provider will not count on a single network provider, a cloud user should not depend on a single cloud service provider either. However, cloud service providers provide different levels of services. A more costly service is usually more reliable. As a result it is an important and challenging problem to choose among a set of service providers to fit one's need, which could be budget, failure probability, or the amount of data that can survive failure. The goal of this paper is to select cloud service providers in order to maximize the benefits with a given budget. The contributions of this paper include a mathematical formulation of the cloud service provider selection problem in which both the object functions and cost measurements are clearly defined, algorithms that selects among cloud storage providers to maximize the data survival probability or the amount of surviving data, subject to a fixed budget, and a series of experiments that demonstrateDuring recent years cloud service providers have successfully provided reliable and flexible resources to cloud users. For example Amazon Elastic Block Store (Amazon EBS) and Simple Storage Service (Amazon S3) provides users storage in the cloud. Despite the tremendous efforts cloud service providers have devoted to the availability of their services, the interruption is still inevitable. Therefore just as an Internet service provider will not count on a single network provider, a cloud user should not depend on a single cloud service provider either. However, cloud service providers provide different levels of services. A more costly service is usually more reliable. As a result it is an important and challenging problem to choose among a set of service providers to fit one's need, which could be budget, failure probability, or the amount of data that can survive failure. The goal of this paper is to select cloud service providers in order to maximize the benefits with a given budget. The contributions of this paper include a mathematical formulation of the cloud service provider selection problem in which both the object functions and cost measurements are clearly defined, algorithms that selects among cloud storage providers to maximize the data survival probability or the amount of surviving data, subject to a fixed budget, and a series of experiments that demonstrate that the proposed algorithms are efficient enough to find optimal solutions in reasonable amount of time, using price and fail probability taken from real cloud providers. that the proposed algorithms are efficient enough to find optimal solutions in reasonable amount of time, using price and fail probability taken from real cloud providers.

Network-wide traffic measurement is important for various network management tasks, ranging from traffic accounting, traffic engineering, and network troubleshooting to security. Existing techniques for traffic measurement tend to be sub-optimal due to poor choice of monitor deployment location or due to constantly evolving monitoring objectives and traffic characteristics. It is not feasible to dynamically reconfigure/redeploy monitoring infrastructure to satisfy such evolving measurement requirements. In this paper, we present a distributed measurement-aware traffic engineering protocol based on a game-theoretic re-routing policy that attempts to optimally utilize existing monitor locations for maximizing the traffic measurement gain while ensuring that the traffic load distribution across the network satisfies some traffic engineering constraint. We introduce a novel cost function on each link that reflects both the measurement gain and the traffic engineering (TE) constraint. Individual routers compete with each other (in a game) to minimize their own costs for the downstream paths, i.e., each router dynamically gathers its cost information for upstream routers and use it to locally decide how to adjust traffic split ratios for each destination to the next-hop routers among these multiple equal-cost paths. Our routing policy guarantees not only a provable Nash equilibrium, but also a quick convergence without significant oscillations to an equilibrium state in which the measurement gain of the network is close to the best case performance bounds We evaluate the protocol via simulations using real traces/topologies (Abilene, AS6461 and GEANT). The simulation results show fast convergence (as expected from the theoretical results), improved measurement gains (e.g., 12 % higher) and much lower TE-violations (e.g., up to 100X smaller) compared to static, centralized measurement-aware routing framework in dynamic traffic scenario.

Achieving high coverage is an important goal in design verification. Fixing coverability problems found at the verification stage, however, can require tremendous effort. To address this problem, we propose a flow for analyzing code and variable-toggle coverability at the early-RTL block-level stage. In addition, we devise a novel technique to analyze the coverability problems so that engineers can resolve the issues more efficiently. By identifying coverability problems at early RTL design stages, design verifiability can be improved, thus reducing the effort required at the verification phase.

Network-wide traffic measurement is important for various network management tasks, ranging from traffic accounting, traffic engineering, network troubleshooting to security. Previous research in this area has focused on either deriving better monitor placement strategies for fixed routing, or strategically routing traffic sub-populations over existing deployed monitors to maximize the measurement gain. However, neither of them alone suffices in real scenarios, since not only the number of deployed monitors is limited, but also the traffic characteristics and measurement objectives are constantly changing. This paper presents an MMPR (Measurement-aware Monitor Placement and Routing) framework that jointly optimizes monitor placement and dynamic routing strategy to achieve maximum measurement utility. The main challenge in solving MMPR is to decouple the relevant decision variables and adhere to the intra-domain traffic engineering constraints. We formulate it as an MILP (Mixed Integer Linear Programming) problem and propose several heuristic algorithms to approximate the optimal solution and reduce the computation complexity. Through experiments using real traces and topologies (Abilene , AS6461 , and GEANT ), we show that our heuristic solutions can achieve measurement gains that are quite close to the optimal solutions, while reducing the computation times by a factor of 23X in Abilene (small), 246X in AS6461 (medium), and 233X in GEANT (large), respectively.

Network-wide traffic measurement is of interest to network operators to uncover global network behavior for the management tasks of traffic accounting, debugging or troubleshooting, security, and traffic engineering. Increasingly, sophisticated network measurement tasks such as anomaly detection and security forensic analysis are requiring in-depth fine-grained flow-level measurements. However, performing in-depth per-flow measurements (e.g., detailed payload analysis) is often an expensive process. Given the fast-changing Internet traffic landscape and large traffic volume, a single monitor is not capable of accomplishing the measurement tasks for all applications of interest due to its resource constraint. Moreover, uncovering global network behavior requires network-wide traffic measurements at multiple monitors across the network since traffic measured at any single monitor only provides a partial view and may not be sufficient or accurate. These factors call for coordinated measurements among multiple distributed monitors. In this paper, we present a centralized optimization framework, LEISURE (Load-EqualIzed measurement), for load-balancing network measurement workloads across distributed monitors. Specifically, we consider various load-balancing problems under different objectives and study their extensions to support different deployment scenarios. We evaluate LEISURE via detailed simulations on Abilene and GEANT network traces to show that LEISURE can achieve much better load-balanced performance (e.g., 4.75X smaller peak workload and 70X smaller variance in workloads) across all coordinated monitors in comparison to naive solution (uniform assignment) to accomplish network-wide traffic measurement tasks.

Reset is one of the most important signals in many designs. Since reset is typically not timing critical, it is handled at late physical design stages. However, the large fanout of reset and the lack of routing resources at these stages can create variant delays on different targets of the reset signal, creating reset recovery problems. Traditional approaches address this problem using physical design methods such as buffer insertion or rerouting. However, these methods may invalidate previous optimization efforts, making timing closure difficult. In this work we propose a formal method to calculate reset recovery slacks for registers at the register transfer level. Designers and physical design tools can then utilize this information throughout the design flow to reduce reset problems at later design stages.

Most synthesis tools perform optimizations based on the design itself and do not utilize the information present in the verification environment. Not using such information greatly limits the optimization capabilities of synthesis tools, which is especially serious for circuit customization because most environment constraints are encoded in the testbench. To exploit verification intention, we propose a methodology that utilizes functional assertions for design optimization. To support circuit customization, we also propose a property mining technique that can extract properties from the design under the constraints in the testbench. Our experimental results show that these methods can reduce design size after synthesis, and the optimization is orthogonal to other existing circuit customization methods.

A Shrew attack, which uses a low-rate burst carefully designed to exploit TCP's retransmission timeout mechanism, can throttle the bandwidth of a TCP flow in a stealthy manner. While such an attack can significantly degrade the performance of all TCP-based protocols and services including Internet routing (e.g., BGP), no existing scheme clearly solves the problem in real network scenarios. In this paper, we propose a simple protection mechanism, called SAP (Shrew Attack Protection), for defending against a Shrew attack. Rather than attempting to track and isolate Shrew attackers, SAP identifies TCP victims by monitoring their drop rates and preferentially admits those packets from the victims with high drop rates to the output queue. This is to ensure that well-behaved TCP sessions can retain their bandwidth shares. Our simulation results indicate that under a Shrew attack, SAP can prevent TCP sessions from closing, and effectively enable TCP flows to maintain high throughput. SAP is a destination-port-based mechanism and requires only a small number of counters to find potential victims, which makes SAP readily implementable on top of existing router mechanisms.

A Shrew attack, which uses a low-rate burst carefully designed to exploit TCP's retransmission timeout mechanism, can throttle the bandwidth of a TCP flow in a stealthy manner. While such an attack can significantly degrade the performance of all TCP-based protocols and services including Internet routing (e.g., BGP), no existing scheme clearly solves the problem in real network scenarios. In this paper, we propose a simple protection mechanism, called SAP (Shrew Attack Protection), for defending against a Shrew attack. Rather than attempting to track and isolate Shrew attackers, SAP identifies TCP victims by monitoring their drop rates and preferentially admits those packets from victims with high drop rates to the output queue. This is to ensure that well-behaved TCP sessions can retain their bandwidth shares. Our simulations indicate that under a Shrew attack, SAP can prevent TCP sessions from closing, and effectively enable TCP flows to maintain high throughput. SAP is a destination-port-based mechanism and requires only a small number of counters to find potential victims, which makes SAP readily implementable on top of existing router mechanisms.

Remote medical monitoring with medical sensors and wireless communications has recently gained attention for the potential savings and improved quality in patient health-care. Currently, many mobile wireless medical infrastructures connect to the Internet in a point-to-point fashion, such as the cellular network. However, our research investigates new models for remote patient care that exploit P2P networking among patients and healthcare providers (nurses, doctors, emergency personnel). In this paper, we identify several medical applications based on P2P Health Networking focusing on two specific scenarios where nurses and patients are both equipped with Bluetooth devices. Nurses opportunistically collect, share and upload data in a P2P fashion. During an emergency, the nearest nurse is alerted via Bluetooth Enhanced Inquiry Response mechanism. The simulation and test-bed experimentation demonstrate that Bluetooth P2P networking is both feasible and cost-effective in remote medical monitoring.

A low power CMOS Voltage Reference Circuit was designed and implemented by TSMC 0.18-μm CMOS process. The voltage reference circuit uses the VGSdifference between two MOSFETs operating in the weak-inversion region to generate the voltage with positive temperature coefficient. The reference voltage can be obtained by combining the weighted VGSdifference with weak-inversion VGSvoltage, which has a negative temperature coefficient. This circuit provides a nominal reference voltage of 621 mV, a temperature coefficient of 11.5 ppm/°C in [-20°C~120°C] from a 1.5V supply voltage. The line regulation of the reference voltage is 6 mV/V when the supply voltage is increased from 1.5V to 3V. The chip area is 0.132 mm2and dissipates 17.25 μW at room temperature. By connecting a 0.22 μF loading capacitor, the measured noise density at 100 Hz and 100 kHz is 0.14 μV/√Hz and 22.2 μV/√Hz, respectively.

In this paper, we present a systematic recursive formula for bit-wise decomposition of M-ary symbol metric. The decomposed bit metrics can be applied to improve the performance of a system where the information sequence is binary-coded and interleaved before M-ary modulated. A traditional receiver designed for certain system is to de-map the received M-ary symbol into its binary isomorphism so as to facilitate the subsequent bit-based manipulation, such as hard-decision decoding. With a bit-wise decomposition of M-ary symbol metric, a soft-decision decoder can be used to achieve a better system performance. The idea behind the systematic formula is to decompose the symbol-based maximum-likelihood (ML) metric by equating a number of specific equations that are drawn from squared-error criterion. It interestingly yields a systematic recursive formula that can be applied to some previous work derived from different standpoint. Simulation results based on IEEE 802.11a/g standard show that at bit-error-rate of 10-5, the proposed bit-wise decomposed metric can provide 3.0 dB, 3.9 dB and 5.1 dB improvement over the concatenation of binary-demapper, deinterleaver and hard-decision decoder respectively for 16QAM, 64QAM and 256QAM symbols, in which the in-phase and quadrature components in a complex M2-QAM symbol are independently treated as two real M-PAM symbols. Further empirical study on system imperfection implies that the proposed bit-wise decomposed metric also improves the system robustness against gain mismatch and phase imperfection. In the end, a realization structure that avails the recursive nature of the proposed bit-decomposed metric formula is addressed