Search

Model-agnostic meta-learning (MAML) is a well-known optimization-based meta-learning algorithm that works well in various computer vision tasks, e.g., few-shot classification. MAML is to learn an initialization so that a model can adapt to a new task in a few steps. However, since the gradient norm of a classifier (head) is much bigger than those of backbone layers, the model focuses on learning the decision boundary of the classifier with similar representations. Furthermore, gradient norms of high-level layers are small than those of the other layers. So, the backbone of MAML usually learns task-generic features, which results in deteriorated adaptation performance in the inner-loop. To resolve or mitigate this problem, we propose contextual gradient scaling (CxGrad), which scales gradient norms of the backbone to facilitate learning task-specific knowledge in the inner-loop. Since the scaling factors are generated from task-conditioned parameters, gradient norms of the backbone can be scaled in a task-wise fashion. Experimental results show that CxGrad effectively encourages the backbone to learn task-specific knowledge in the inner-loop and improves the performance of MAML up to a significant margin in both same- and cross-domain few-shot classification., Comment: Accepted to WACV2022

Herein, we present a novel triboelectric nanogenerator based on nylon-11/poly-DADMAC and PVDF-TrFE nanofibers and demonstrate as a self-powered human motion sensor (SPHMS). The incorporation of cation poly-DAD MAC into the nylon-11 via the electrospinning method acts as a highly positive triboelectric layer that significantly enhanced dielectric constant by trapping surface charge in the composite materials. The fabricated SPHMS exhibited outstanding performance, including superior electrical performance (power 2.4 mW), excellent mechanical stability (>76k cycles), and sensitivity (1.01 V/KPa). As a proof of concept, the SPHMS was successfully demonstrated for monitoring human motion while walking, jogging, running, and jumping via a smartphone.

In this study, we have firstly developed laser-carbonized Ti 3 C 2 T x MXene-based hierarchical nanofibers (LCHNFs) from electrospun dehydrofluorinated MXene-poly(vinylidene fluoride) (PVDF) nanofibers (DNFs). The incorporation of an optimal amount of MXene promotes the $\beta$ -phase of PVDF by introducing electrostatic interactions. We reveal that the $\beta$ -phase is transformed into sp2-hybridized carbon hexagonal structured by the optimized laser-carbonization process. Finally, a breathable and reusable electrocardiogram (ECG) sensing electronic tattoos (e-tattoos) is fabricated via the one-step laser-scribing and cutting on DNFs. The LCHNF e-tattoos show a record sheet resistance (4 Ω sq−1), skin-contact impedance (23 KΩ cm2 at 10 Hz), and signal-to-noise ratio (SNR) (35.6).

Recent advances in non-invasive epidermal wearable sensors have included various electronic sensors to monitor either physiological signs such as heart rate, skin temperature, and respiration rate, or metabolites. Nevertheless, most of them have focused only on a single electrophysiological measurement at a time, which greatly inhibits the calibration of various biological signals and diagnostic facilities. Herein, we have newly developed an epidermal electrochemical-electrophysiological hybrid patch for precise glucose detecting with pH and temperature correction as well as simultaneously monitoring of electrocardiogram (ECG). The fabricated epidermal hybrid patch has overcome the inconvenience of mandatory attachment of multiple patches on the human body owing to the full integration of various electrochemical and electrophysiological sensors on the same flexible printed circuit board toward advanced glycemic control and continuous ECG monitoring for effective management of chronic diseases.

Building energy optimization has much related with energy saving and building maintenance viewpoint as well. First we face with building modelling problem. Then it is naturally followed the issue by how exact or how efficient it is. Generally, analytical approach meets rather complex and complicate to express with exactly. Hence, we normally design the system model with possible and accessible input/output data with IoT sensor. That is Neural Network (NN), and it has been focused nowadays. In this paper, we will focus on the building modelling with NN based on input and output data such as temperature, working days, humidity, weather characteristics and electrical consumption for IoT applications. Modelling verification will be followed with test data, and performance also evaluated.

Fashion image retrieval (FIR) is a challenging task, which requires searching for exact items accurately from massive collections of fashion products based on a query image. Despite recent advances, FIR still has limitations for application to real-world visual searches. The main reason for this is not only the trade-off between the model complexity and performance, but also the common nature of fashion images captured under uncontrolled circumstances (e.g. varying viewpoints and lighting conditions). In particular, fashion images are vulnerable to shape deformations and suffer from inconsistency between the user's query images and refined product images. Moreover, multiple fashion objects can be present simultaneously within a single image. In this paper, we considered an FIR method that is optimized for the fashion domain. We investigated training strategies and deep models to improve the retrieval performance. The experimental results on three benchmarks from DeepFashion dataset show that considered methods could achieve the significant improvements compared to the previous FIR methods.

In a conventional Cloud Radio Access Network (CRAN) system, due to geographical and cost problems, it is difficult to install additional remote radio head (RRH) with fronthaul links once after the initial CRAN deployment is completed. To solve this problem, we present a new type of CRAN architecture which has simple supplementary radio nodes termed w-RRHs straightly linked to the adjacent existing RRHs, not to a centralized baseband unit (BBU) pool. This paper quantitatively evaluates the potential benefits from the use of extra w-RRHs. Specifically, we provide a quantified guideline on whether or not to permit the usage of w-RRHs from the value of achievable spectral efficiency with regard to the additional delay caused by w-RRHs. Simulation results clearly demonstrate the trade-off between the advantage and penalty of additional w-RRHs, and emphasize the impact of network size on the system performance.

As the interest of the autonomous driving increases, techniques related to the advanced driver assistance system are evolving together. In this paper, we propose a novel driver identification system using convolutional neural network (CNN) with the background removal-based infrared image data augmentation. It helps to identify who a driver is, and provides the customized driving environment. The process for the proposed identification system is as follows. First, we acquire customized individual infrared images in a driving simulation environment. Second, we augment the large amount of data by using the background removal-based method and several image processing techniques. Third, the augmented data is trained by the low-complexity-based CNN method. Finally, we load all trained weights to the forward network for real-time processing. In the experimental results, the proposed system had the memory resource of 4,795 KB, which are up to 49.0822 times smaller than benchmark algorithms, and the average F 1 score of 0.9418 for the driver identification accuracy.