Search

Treatments for epilepsy include pharmacotherapy or surgery. Recently, focused ultrasound stimulation has been investigated as a promising non-invasive neuromodulation tool for neurological disorders, including epilepsy. To investigate the neuronal dynamics in epilepsy, we acutely stimulated 29 Sprague–Dawley rats to low-intensity focused ultrasound stimulation (LIFUS) three times for 3 minutes. Pentylenetetrazol (PTZ) was injected into the abdominal cavity of the anesthetized rats to induce epilepsy. During the anesthesia, the electroencephalography (EEG) signal was measured and analyzed for 1 h in groups of animals untreated (sham) and treated with LIFUS. The EEG signal was quantitatively processed to show the different characteristics of the frequency change over time and of the band power between sham and treated groups. Histological analyses (Nissl, Iba1, c-Fos, and GAD65) measured the degree of staining of expression factors to confirm the effect of the stimulation on seizure suppression. These results suggest that repetitive LIFUS can effectively reduce epileptic seizure activity by attenuating theta and beta-band oscillation in a PTZ-induced rat model. LIFUS can potentially facilitate hippocampal and cortical cellular recovery by augmenting GABAergic inhibitory neurons via its anti-seizure effect.

Most of the recent studies of reinforcement learning and robotics basically employ computer simulation due to the advantages of time and cost. For this reason, users have to spare time for investigation in order to choose optimal environment for their purposes. This paper presents a survey result that can be a guidance in user’s choice for simulation environments. The investigation result includes features, brief historical backgrounds, license policies and formats for robot and object description of the eight most popular environments in robot RL studies. We also propose a quantitative evaluation method for those simulation environments considering the features and a pragmatic point of view.

In this paper, a new full-bridge (FB) converter with asymmetric duty cycle control for low common mode (CM) noise is proposed in order to achieve high power density. The conventional phase shifted full-bridge converter (PSFB) has a large CM noise, which increases the volume of the CM choke. The proposed converter uses a coupled inductor and an asymmetric duty cycle control to generate the similar voltage variations of the transformer at the primary and secondary side. A CM noise generated from the transformer can be reduced because dv/dt characteristic of the primary side is similar to it of the secondary side. As a result, the proposed converter can achieve high power density by reducing the volume of the CM choke. The performance of the proposed converter is verified by the experimental results with a 56 V/13.4 A 750 W prototype converter.

A promising way to search for microbial life in our solar system's Ocean Worlds is to make use of 3-d microscopes, as these can provide single-image inventories of the complete contents of liquid sample volumes. Two applicable 3d microscopy techniques are digital holographic microscopy and light-field fluorescence microscopy. The former can provide high-resolution imaging information on cellular morphology, structure, index of refraction, and motility, while the latter can identify and locate targeted molecule families, such as lipids and nucleic acids. The combination of this pair of 3-d techniques thus provides a powerful suite of diagnostic tools. We have recently combined both types of microscope into an integrated dual-mode microscope prototype aimed at demonstrating its utility at terrestrial field sites, and the combined instrument has already been taken on an initial foray into the field to assess its performance and shortcomings. Here we describe the design and capabilities of our dual-mode microscope, as well as initial performance measurements obtained during its first field trip to the local seashore.

CMOS X-ray detectors used in industrial and medical equipment should provide a full image depth even for a specific region of interest, and require high resolution, low noise, and wide DR in a wafer-scale detector [1], [4]. To achieve a wide DR, a large integration capacitor is required within the pixel to prevent its saturation at high dose, but this degrades image quality at low dose. To facilitate wide DR (>70dB), a conventional detector uses a column-parallel readout with a programmable gain amplifier (PGA) and an ADC [3]. However, the PGA consumes substantially more power and area than the ADC, and its gain control requires multiple X-ray exposures. The use of switched-capacitor (SC) ΔΣ ADC provides wide DR with an improved noise performance [2], [5]. However, its SC input draws high peak current that must be supplied by pixels and reference drivers, and its complex clock distribution also requires high power consumption.

The surveillance video management system is rapidly expanding its scope of application at the request of citizens and the development of related technologies. In addition, as Cloud Computing and 5G network are applied with AI, scope and function of surveillance systems are being enhanced to intelligent CCTV beyond simple monitoring. However, intelligent CCTV systems with Mobile Edge Computing and 5G, which have the risk of privacy infringement. Accordingly, it is necessary to identify various types of security threats that can be occurred through the cloud based surveillance system and to eliminate the risk of privacy and personal information breaches. So, in this paper, we propose a hierarchical cloud based video surveillance system considering security on the 5G Network.

Motion imitation is a fundamental communication skill for a robot; especially, as a nonverbal interaction with a human. Owing to kinematic configuration differences between the human and the robot, it is challenging to determine the appropriate mapping between the two pose domains. Moreover, technical limitations while extracting 3D motion details, such as wrist joint movements from human motion videos, results in significant challenges in motion retargeting. Explicit mapping over different motion domains indicates a considerably inefficient solution. To solve these problems, we propose a three-phase reinforcement learning scheme to enable a NAO robot to learn motions from human pose skeletons extracted from video inputs. Our learning scheme consists of three phases: (i) phase one for learning preparation, (ii) phase two for a simulation-based reinforcement learning, and (iii) phase three for a human-in-the-loop-based reinforcement learning. In phase one, embeddings of the motions of a human skeleton and robot are learned by an autoencoder. In phase two, the NAO robot learns a rough imitation skill using reinforcement learning that translates the learned embeddings. In the last phase, the robot learns motion details that were not considered in the previous phases by interactively setting rewards based on direct teaching instead of the method used in the previous phase. Especially, it is to be noted that a relatively smaller number of interactive inputs are required for motion details in phase three when compared to the large volume of training sets required for overall imitation in phase two. The experimental results demonstrate that the proposed method improves the imitation skills efficiently for hand waving and saluting motions obtained from NTU-DB.

When applying deep reinforcement learning to the motion mimicry problem between teacher and student robots, this paper reports the initial results of how various hyper-parameter configurations affect performance of learning processes and quality of generated motions. The hyperparameters considered in this study include the structure of policies such as convolutional and fully connected networks, the type of activation functions such as ReLU and hyperbolic tangent, and the number of input sequences such as one, four and eight. Under these deep neural network configurations, PPO reinforcement learning algorithm has been applied for learning. In the simulator environment, the teacher NAO robot demonstrates a target action repeatedly, and the learner NAO robot tries to learn that action. The target actions include handshaking and two-arm raising. Our experimental results show that fully connected networks outperform the convolutional counterparts both in training statistics and motion quality. For activation functions, however, we found an interesting mismatch between training and evaluation quality: for example, a configuration with higher rewards does not guarantee less motion discrepancy, which may suggest a new research direction to design better loss and reward functions for robotic motion mimicry.

We present a ParaViewWeb based visual analytics application running on large high-resolution display supporting standard mouse and keyboard interaction. The application relies on SAGE2 for user interaction and multi-display visualization. We also employ a scalable middleware system called "Cloudberry" that allows users to interactively query and analyze large amounts of temporal and spatial data stored on a back end Apache AsterixDB store to enable big data analytics and interactive visualization. Our Visual Analyzing Billion Tweets application shows interactive query and visualization of result from over a billion twitter feeds streamed in real-time to the back end Apache AsterixDB. In our setup, we ran the visual analytics application on a large high-resolution display with a 24-tiled display in a 6 x 4 configuration. We also run a comparative study of the application running on a single 24 inch display and the 24-tiled display with some very interesting findings supporting the benefit of using large high-resolution display for visual analytics.

Keyword-based heat maps are a natural way to explore and analyze the spatial properties of social media data. Dealing with large datasets, there may be many different keywords, making offline pre-computations very hard. Interactive frameworks that exploit database sampling can address this challenge. We present a novel middleware technique called Heatflip, which issues diametrically opposed samples into the temporal and spatial dimensions of the data stored in an external database. Spatial samples provide insights into the temporal distribution and vice versa. The progressive exploration approach benefits from adaptive indexing and combines the retrieval and visualization of the data in a middleware layer. Without any a priori knowledge of the underlying data, the middleware can generate accurate heat maps in 85% shorter processing times than conventional systems. In this paper, we discuss the analytical background of Heatflip, showcase its scalability, and validate its performance when visualizing large amounts of social media data.

As the deep learning (DL) has widely been used for application domains such as image classifications, natural language processing, and speech recognition, various software frameworks have been developed. They provide users with efficient programming interfaces for developing the DL applications. The optimization techniques within these frameworks generally are different from each other, which leads to different processing times for even the same applications. However, it is difficult that end users consider performance differences in processing time due to incompatible programming interface among the DL frameworks. These differences might cause redundant efforts and costs for end users to develop and maintain the applications. In this paper, we introduce an integrated deep learning engine (IDLE), a novel interface working on the top of the existing DL frameworks, which provides a convenient, flexible and scalable programming interface developing the DL applications for end users regardless of DL frameworks. Besides, we also propose a novel adaptive task scheduling scheme for training DL applications in a cluster with different GPUs. We implement our platform on the heterogeneous GPU cluster, and the results show that the proposed scheduling algorithm improves cost efficiency processing various DL applications.

Many significant research achievements in the past few decades have demonstrated that convolutional neural networks (CNNs) could be capable of steering wheel control which is the basic and essential maneuver of autonomous vehicles. We propose an end-to-end steering controller with CNN-based closed-loop feedback for autonomous vehicles that improves driving performance compared to traditional CNN-based approaches. This paper demonstrates that the proposed neural network, DAVE-2SKY, is able to learn to inference steering wheel angles for the lateral control of self-driving vehicles through initial supervised pre-training and subsequent reinforced closed-loop post-training with images from a camera mounted on the vehicle. We used the PreScan simulator and Caffe deep learning framework for training under diversified circumstances in a software in the loop (SIL) simulation environment. We used DRIVE TM PX2 computer to implement a self-driving car for an experimental validation of the proposed end-to-end controller. The performance of the proposed system has been investigated under simulations and on-road tests as well. This work shows that the CNN-based end-to-end controller performs robust steering control even under partially observable road conditions, which indicates the possibility of fully self-driving vehicles controlled by CNN-based end-to-end steering controllers.

Training deep learning model is time consuming, so various researches have been conducted on accelerating the training speed through distributed processing. Data parallelism is one of the widely-used distributed training schemes, and various algorithms for the data parallelism have been studied. However, since most of studies assumed homogeneous computing environment, there is a problem that they do not consider a heterogeneous performance graphics processing unit (GPU) cluster environment. The heterogeneous performance environment leads to differences in computation time between GPU workers in the synchronous data parallelism. Due to the difference of the computation time of one iteration, the straggler problem that fast workers wait for the slowest worker makes training speed slow. Therefore, in this paper, we propose a batch-orchestration algorithm (BOA), reducing the training time by improving hardware efficiency in the heterogeneous performance GPU cluster. The proposed algorithm coordinates local mini-batch sizes for all workers to reduce the training iteration time. We confirmed that the proposed algorithm improves the performance by 23% over the synchronous SGD with one back-up worker when training ResNet-194 using 8 GPUs of three different types.

In recent years, with the growing number of vehicles, the efficient parking management system has become necessary for large buildings. This paper presents a parking occupancy monitoring system which can automatically decide whether a vehicle is parked or empty in each parking space. There are many obstacles in the parking area, such as high diversity in car models, occlusion by other car, moving person, waste trash, and camera lens distortion, making it difficult to detect a vehicle. In order to solve all these problems, this paper proposes to use a part-based and machine learning-based vehicle detection algorithm. We demonstrate that the proposed method performs well on large indoor parking lot dataset which contains the above-mentioned obstacles.

The proliferation of GPS-enabled mobile devices has generated geo-tagged data at an unprecedented rate over the past decade. Data-processing systems that aim to ingest, store, index, and analyze Big Data must deal with such geo-tagged data efficiently. In this paper, among representative, disk-resident spatial indexing methods that have been adopted by major SQL and NoSQL systems, we implement five variants of these methods in the form of Log-Structured Merge-tree-based (LSM) spatial indexes in order to evaluate their pros and cons for dynamic geo-tagged Big Data. We have implemented the alternatives, including LSM-based B-tree, R-tree, and inverted index variants, in Apache AsterixDB, an open source Big Data management system. This implementation enabled comparison in terms of real end-to-end performance, including logging and locking overheads, in a full-function, query-based system setting. Our evaluation includes both static and dynamic workloads, ranging from a "load once, query many" case to a case where continuous concurrent incremental inserts are mixed with concurrent queries. Based on the results, we discuss the pros and cons of the five index variants.

In this paper, an object detection and localization method for bin picking of plastic wrapped objects is described. Since such objects are deformable and have non-Lambertian surfaces, it is difficult to apply conventional feature point approaches or edge based template matching. To solve this problem, we propose a new method which is called “KCS (kernel convolution score)”. It measures the total score of convolution between local image and multiple kernels which are generated according to the characteristics of the target object. The local image having maximum score is considered as the most promising object for picking. In the last part of this paper, the limitation of existing edge and template matching in the given problem is discussed with the experimental results. Performance evaluation shows that the proposed method could localize the object which is appropriate for picking up with 91.67% success rate.

This paper presents the maximum tire-road friction coefficient estimation algorithm which considers about the effect of states. Tire force information is an important factor for active safety system. However, it is difficult to estimate due to the dependency on many states such as vehicle speed, tire pressure, and tire wear. In this paper, several experimental researches about the effect of states on the maximum friction coefficient and previous maximum tire-road friction coefficient estimation algorithms are reviewed and summarized. The influential states and the estimation method which doesn't require extra sensors were determined and combined. The proposed algorithm consists of two parts: an interacting multiple models (IMM) based maximum tire-road friction coefficient estimation and an updating sequence based on the effect of vehicle speed. To validate the algorithm, the closed-loop simulation with the advanced emergency braking system (AEBS) has been conducted. It has been shown that the proposed estimation algorithm could enhance the performance of AEBS algorithm.

Objects with non-Lambertian reflecting surfaces suffer severe appearance changes according to the change of observer's angle of view. Thus it is difficult to obtain a consistent object model. This paper addresses the problem of bin picking for such objects without using explicit object models. Due to the above-mentioned reasons, traditional pattern matching approaches are difficult to apply. Furthermore, 3D sensors often fail to capture these objects correctly, which also disables the approaches based on 3D object models. Considering these limitations, this paper proposes a novel algorithm that tries to find a reasonable pickup point by investigating the status of the pile in the bin instead of trying to perfectly isolate single object from the pile and measure full pose information. For the purpose, feature point matching between stereo cameras is adopted to retrieve approximate 3D information of the pile. Additionally, by associating these corresponding point pairs with the binary patches around the point, corresponding binary patch pairs are detected, which is evaluated by a proposed objective function to select the most appropriate candidate for the pickup operation.

In this paper we present a master-slave system to program dual arm industrial robots. Since traditional robot motion programming apparatus, teach pendants, usually deal with single arm motions, it is difficult to teach dual arm motions requiring synchronization of the arms. Furthermore, a teach pendant is 1-dimensional position input device which is not sufficient to teach 3-dimensional motions so that it requires quite much time for teaching complex motions. To deal with this problem, we developed a passive exoskeleton as a master device for a master-slave teleoperation system. The proposed teleoperation-based robot programming method was verified through experiments on via point teaching. An operational space framework is adopted to control the slave arm.

Instant search is an emerging information-retrieval paradigm in which a system finds answers to a query instantly while a user types in keywords character-by-character. Fuzzy search further improves user search experiences by finding relevant answers with keywords similar to query keywords. A main computational challenge in this paradigm is the high-speed requirement, i.e., each query needs to be answered within milliseconds to achieve an instant response and a high query throughput. At the same time, we also need good ranking functions that consider the proximity of keywords to compute relevance scores. In this paper, we study how to integrate proximity information into ranking in instant-fuzzy search while achieving efficient time and space complexities. We adapt existing solutions on proximity ranking to instant-fuzzy search. A naive solution is computing all answers then ranking them, but it cannot meet this high-speed requirement on large data sets when there are too many answers, so there are studies of early-termination techniques to efficiently compute relevant answers. To overcome the space and time limitations of these solutions, we propose an approach that focuses on common phrases in the data and queries, assuming records with these phrases are ranked higher. We study how to index these phrases and develop an incremental-computation algorithm for efficiently segmenting a query into phrases and computing relevant answers. We conducted a thorough experimental study on real data sets to show the tradeoffs between time, space, and quality of these solutions.

Apple's iPhone has opened the market of various smart devices and changed the way of computing in daily life. In traditional PCs, mouse and keyboard are representative input devices, but, in smart devices, touch screen and motion recognition become major input methods. Especially, in game applications, motion recognition is more important to dynamically interact with users. In this paper, we propose a library of motion recognition for general interactive applications on smart devices. The proposed library provides the functions of registering motions in the development phase and recognizing a real-time motion in the running phase. By using the proposed library, any application can create and register its customized motions and recognize user's real-time motion from the registered motions in the common interface.

In this paper, we present the experiences on the OPRoS component based programming by applying OPRoS to a shooting behavior for a small humanoid robot. We also compare OPRoS with the conventional robot programming method. From the experience and comparison, we expect that user will get some help in choosing the programing methods and the way of programming in OPRoS.

In this paper, we have designed shoes to provide active performance which enables the dancer to interact with sounds and images. This novel physical interface which we named 'the Shadow Dancer' consists of two systems. The first system is wearable physical interface that uses procedural input to enable dancer to create rhythmic pattern. The second system is sound and image output using speaker and display to interact against dancers. The combined system provides communication of interactive dancer. We have used rhythm of step dance as outline application for the field of interactive dance.

The dependence of oxide charge-to-breakdown (Q/sub BD/) and device degradation on the combined post annealing of RTA and FA in the tungsten polycide gate technology have been experimentally investigated. The experimental results suggest that Q/sub BD/ and degradation are improved by lower temperature and shorter time of RTA. Whereas the FA/RTA annealing sequence is more advantageous for improving Q/sub BD/, the RTA/FA annealing sequence is good for improving device degradation.

This paper proposes a new simple architecture of digital dividing system in dual-modulus prescaler phase-locked loop for wireless communications. In this new architecture a swallow counter is not employed while the same total division ratio as in a conventional system can be obtained. This simple architecture shows advantages in reducing power consumption and gate-counts and is suitable for small die area and low power applications. The circuit is designed in a standard 0.35 um CMOS process.