Search

The main challenge of FSCIL is the trade-off between underfitting to a new session task and preventing forgetting the knowledge for earlier sessions. In this paper, we reveal that the angular space occupied by the features within the embedded area is relatively narrow. Consequently, after the base session training with the converged feature extractor, adding features of new classes easily overlaps with the previously occupied space of previous classes. Furthermore, in contrast to the base session classes, whose features are relatively well aggregated, the features of new session classes are dispersed over large regions. Thus, we propose the Incrementally Learned Angular Representation (ILAR) learning structure to address these issues. During the base session, ILAR attempts to increase the marginal spacing between the distributions of features for each class. In incremental sessions, modifications of features and classifiers corresponding to the previous session classes are limited to preserve the past knowledge while the features of new classes are readjusted. Furthermore, we generate additional features according to the statistics learned from the base session to reduce the variance caused by data shortage. Experiments proceeded on three popular benchmark datasets, including CIFAR100, miniImageNet, and CUB200. We demonstrate that the proposed method achieves state-of-the-art performances by effectively enhancing the new session classification ability while preserving the knowledge of the past sessions.

Time-series data with missing values are a common occurrence in various fields, including healthcare, meteorology, and robotics. The process of imputation aims to fill in the missing values with valid values. Most imputation methods implicitly train models due to the presence of missing values. In this paper, we propose Random Drop Imputation with Self-training (RDIS), a novel training method for time-series data imputation models. In RDIS, we generate extra missing values by applying a random drop to the observed values in incomplete data. We can explicitly train the imputation models by filling in the missing values. Moreover, we utilize self-training with pseudo values to exploit the original missing values. To enhance the quality of pseudo values, we set a threshold and filter them based on entropy calculation. To evaluate the effectiveness of RDIS for imputing time-series data, we test it across several imputation models and obtain competitive results on three real-world datasets.

In the conventional successive cancellation (SC) decoder for polar codes, all the future bits to be estimated later are treated as random variables. However, polar codes inevitably involve frozen bits, and their concatenated coding schemes also include parity bits (or dynamic frozen bits) causally generated from the past bits estimated earlier. We refer to the frozen and parity bits located behind a target decoding bit as its future constraints (FCs). Although the values of FCs are deterministic given the past estimates, they have not been exploited in the conventional SC-based decoders, not leading to optimality. In this paper, with a primary focus on the binary erasure channel (BEC), we propose SC-check (SCC) and belief propagation SCC (BP-SCC) decoding algorithms in order to leverage FCs in decoding. We further devise an improved tree search technique based on stack-based backjumping (SBJ) to solve dynamic constraint satisfaction problems (CSPs) formulated by FCs. Over the BEC, numerical results show that a combination of the BP-SCC algorithm and the SBJ tree search technique achieves the erasure recovery performance close to the dependence testing (DT) bound, a bound of achievable finite-length performance.

The industrial field faces the problem of process optimization by finding the factors affecting the yield of the process and controlling them appropriately. However, due to limited resources such as time and money, optimization is performed using a low evaluation budget. In addition, for process stability, the lower limit of the yield is set so that the yield must be maintained above this limit during optimization. Bayesian Optimization (BO) can be an effective solution in acquiring optimal samples that satisfy a safety constraint given a low evaluation budget. However, many existing BO algorithms have some limitations such as significant performance degradation due to model misspecification, and high computational load. Thus, we propose a practical safe BO algorithm, A-SafeBO, that effectively reduces performance degradation due to model misspecification using only a limited evaluation budget. Additionally, our algorithm performs computations for a large number of observations and high-dimensional input spaces by using Ensemble Gaussian Processes and Safe Particle Swarm Optimization. Here, we also propose a new acquisition function that leads to a wider exploration even under the constraint of safety. This will help deviate from the local optimum and achieve a better recommendation. Our algorithm empirically guarantees convergence and performance through evaluations on several synthetic benchmarks and a real-world optimization problem.

Ground segmentation of outdoor LiDAR point clouds remains challenging due to its sparse and irregular nature. This paper presents SectorGSnet: a ground segmentation framework for outdoor LiDAR point clouds, aiming to accomplish this task efficiently and effectively. The framework consists of a sector encoder module and a segmentation module. The former module introduces a novel bird’s-eye-view (BEV) sector partition strategy that discretizes the point cloud into sectors of varying sizes to enhance the 2D representation ability of the point cloud. Then, the points within each sector are fed into a multimodal PointNet encoder to obtain the corresponding sector feature map. In the latter module, a lightweight segmentation network next learns binary classification from the sector feature map, and point labels are restored from the sector segmentation results. Our proposed framework is trained and evaluated on SemanticKITTI against state-of-the-art 2D projection-based approaches and achieves an excellent balance between performance and computational complexity. We conducted inference at 170.6 Hz on a desktop PC with a GTX2080Ti GPU, and also experimented on a resource-limited platform with only 10 watts of power.

Real-time hand gesture recognition plays a vital role in human-computer interaction (HCI). Recent radar-based hand gesture recognition methods have focused on achieving high classification accuracy using deep neural network (DNN)-based classifiers. However, the hand gesture recognition system should not only classify the gestures accurately but also detect out-of-distribution (OOD) samples to be used in real-world HCI scenarios with high reliability. Recognition systems without OOD detection capability misclassify unintended gestures in silence, especially in real-time scenarios. To tackle this problem, we propose a real-time hand gesture recognition system that can simultaneously classify hand gestures and detect OOD samples by using a Frequency Modulated Continuous Wave (FMCW) radar sensor. First, we design radar data processing technique and Transformer encoder-based classifier to achieve high classification accuracy. Second, the relative Mahalanobis distance (RMD)-based OOD detection method is adopted to increase the reliability of the proposed system. Finally, one in-distribution dataset and two OOD datasets are collected to verify the proposed system. The proposed system achieves a classification accuracy of 93.95% on the in-distribution dataset. We conduct the OOD detection experiments with two OOD datasets for which the proposed system reports AUROC values of 92.96% and 92.84%, respectively. Furthermore, the feasibility of the proposed system is certified through a real-time experimental demonstration.

Nowadays, the welding process is essential in various manufacturing industrial fields, such as aerospace, vehicle production, and shipbuilding. The welding defects caused in the process need to be monitored as they can cause serious accidents and losses. Traditional computer vision methods in an industrial application are inefficient when the detection targets have variations in shape, scale, and color because the detection performance depends on the hand-crafted features. To overcome this limitation, deep learning models, such as the convolutional neural network (CNN), are applied to industrial defect detection. These CNN-based models trained on static images, however, have a low performance that cannot meet the industrial requirements. To deal with the challenge, bidirectional Convolutional Recurrent Reconstructive Network (bi-CRRN) is proposed for welding defect detection and localization based on welding video. Spatio-temporal data, specifically the forward and backward sequences, are considered in our bi-CRRN to get high detection performance. Moreover, an automatic defect detection equipment is developed to weld a material and monitor the welding bead simultaneously. We demonstrate that the proposed bi-CRRN outperforms the other segmentation network models in welding defect detection.

Since the Convolutional Neural Network (CNN) has surfaced and fascinated the world, many researchers have exploited CNN for image classification, object detection, semantic segmentation, etc. However, the conventional CNNs have a pyramidal structure and were designed to process images which have the same size. Although some CNNs can accept images of various sizes, performance is degraded for images smaller than the size of images used for training. In this paper, we propose MarsNet, a CNN based end-to-end network for multi-label classification with an ability to accept various size inputs. In order to allow the network to accept such images, dilated residual network (DRN) is modified to get higher resolution feature maps, and horizontal vertical pooling (HVP) is newly designed to efficiently aggregate positional information from the feature maps. Furthermore, multi-label scoring module and threshold estimation module are employed to serve the purpose of multi-label classification. We verify the effectiveness of the proposed network through two distinctive experiments. We first verify our model by inspecting and classifying multiple types of defects occurred in PCB screen printer using solder paste inspection (SPI) datasets. Secondly, we verify our network using VOC 2007 dataset. Our network is pioneering in that no research has attempted to accomplish multi-label classification for defects in addition to being able to take input images of various sizes in SPI field.

In recent years, in the field of education technology, artificial intelligence tutors have come to be expected to provide individualized educational services to help learners achieve high levels of academic success. To this end, AI tutors need to be able to understand the current status and preferences of a learner and then suggest appropriate learning contents accordingly. However, it is challenging to monitor learner status and preferences continually and to recommend appropriate educational services. In this paper, we propose an individualized AI tutor as an integrated system of three developmental learning networks (DLNs) by extending a deep adaptive resonance theory (Deep ART) network, a neural network capable of incremental learning. Specifically, the learner status DLN is able to easily add new input channels about learner status without disrupting existing classifiers. The learner preference DLN is to categorize learner preferences based on frequency as well as sequence of events. The learner experience DLN is updated to immediately reflect alteration of the educational effectiveness in the current classification. Our AI tutor is currently embedded in a commercialized mobile application for teaching the Korean language to children. Experimental results show that the AI tutor application efficiently helps children learn the Korean language.

In the field of surface mount technology (SMT), early detection of defects in production machines is crucial to prevent yield reduction. In order to detect defects in the production machine without attaching additional costly sensors, attempts have been made to classify defects in solder paste printers using defective solder paste pattern (DSPP) images automatically obtained through solder paste inspection (SPI). However, since the DSPP images are sparse, have various sizes, and are hardly collected, existing CNN-based classifiers tend to fail to generalize and over-fitted to the train set. Besides, existing studies employing only multi-label classifiers are less helpful since when two or more defects are observed in the DSPP image, the location of each defect can not be specified. To solve these problems, we propose a dual-level defect detection PointNet (D3PointNet), which extracts point cloud features from DSPP images and then performs the defect detection in two semantic levels: a micro-level and a macro-level. In the micro-level, a type of printer defect per point is identified through segmentation. In the macro-level, all types of printer defects appearing in a DSPP image are identified by multi-label classification. Experimental results show that the proposed D3PointNet is robust to the sparsity and size changes of the DSPP image, and its exact match score was 10.2% higher than that of the existing CNN-based state-of-the-art multi-label classification model in the DSPP image dataset.

Due to the development of short-range radar with high-resolution, the radar sensor has a high potential to be used in real human-computer interaction (HCI) applications. The radar sensor has advantages over optical cameras in that it is unaffected by illumination and it is able to detect the objects in an occluded environment. This paper proposes a hand gesture recognition system for a real-time application of HCI using 60 GHz frequency-modulated continuous wave (FMCW) radar, Soli, developed by Google. The overall system includes signal processing part that generates range-Doppler map (RDM) sequences without clutter and machine learning part including a long short-term memory (LSTM) encoder to learn the temporal characteristics of the RDM sequences. A set of data is collected from 10 participants for the experiment. The proposed hand gesture recognition system successfully distinguishes 10 gestures with a high classification accuracy of 99.10%. It also recognizes the gestures of a new participant with an accuracy of 98.48%.

In this paper, we address a number of weaknesses of the windowed decoding of spatially coupled low-density parity-check (SC LDPC) codes and propose three modifications that simultaneously improve its performance, complexity, and latency. An effective termination method of the windowed decoding and the reuse of edge messages of previous target symbols provide a good performance-latency tradeoff when compared with the conventional windowed decoder. Also, we propose a scheme that lowers the error floor, in which the amplified edge messages of the previous window are used in the present window. The proposed windowed decoding, consisting of the three schemes, provides a significant performance gain with smaller latency. The validity of the new windowed decoding is verified by the evaluation with codes from different SC LDPC ensembles.