Data augmentation is an effective method to improve model robustness and generalization. Conventional data pipelines are commonly used as preprocessing modules for neural networks with predefined heuristics restricted differentiability. Some recent works indicated that the differentiable (DDA) could effectively contribute training of policy searching strategies. This survey provides a comprehensive structured overview advances in DDA. Specifically, we focus on fundamental elements including...

In computer-aided design (CAD) community, the point cloud data is pervasively applied in reverse engineering, where analysis plays an important role. While a large number of supervised learning methods have been proposed to handle unordered clouds and achieved remarkable success, their performance applicability are limited costly annotation. this work, we propose novel self-supervised pre-training model for without human annotations, which relies solely on upsampling operation perform...

Point cloud upsampling has been extensively studied, however, the existing approaches suffer from losing of structural information due to neglect spatial dependencies between points. In this work, we propose PU-GAT, a novel 3D point method that leverages graph attention networks learn over baselines. Specifically, first design local–global feature extraction unit by combining and position encoding mine local inter-dependencies across features. Then, construct an up-down-up expansion unit,...

As Android malware is growing and evolving, deep learning has been introduced into detection, resulting in great effectiveness. Recent work considering hybrid models multi-view learning. However, they use only simple features, limiting the accuracy of these approaches practice. In this paper, we propose DeepCatra, a approach for whose model consists bidirectional LSTM (BiLSTM) graph neural network (GNN) as subnets. The two subnets rely on features extracted from statically computed call...

Photovoltaic (PV) forecasting plays a major role in residential and industrial PV installation as well penetration with the grid. An inaccurate power may result increased monetary energy losses. This study proposes metaheuristic-based strategy for accurate using heuristic-based data-driven model. The proposed algorithm integrates dense explorative existing equation knowledge by multilayer perceptron (MLP) network Sigmoid activation functions to predict best coefficients inputs of method is...

In this paper, we propose a runtime configurable approximate computing (RCAC) system for simulated annealing (SA) algorithm. The RCAC integrates hardware to calculate the cost function and lightweight quality estimator, which provides valuable insights into SA algorithm's effectiveness by estimating acceptance rate of each iteration. By utilizing rate, dynamically configures operational mode hardware, making good trade-off between solution computational efficiency. Experimental results...

While a large number of supervised learning methods have been proposed to handle the unordered point clouds and achieved remarkable success, their performance is limited costly data annotation. In this work, we propose novel self-supervised pre-training model for cloud without human annotations, which relies solely on upsampling operation perform feature in an effective manner. The key observation our approach that encourages network capture both high-level semantic information low-level...

Point cloud completion aims to infer a complete shape from its partial observation. Many approaches utilize pure encoderdecoder paradigm in which can be directly predicted by priors learned scans, however, these methods suffer the loss of details inevitably due feature abstraction issues. In this paper, we propose novel framework, termed SPAC-Net, that rethink task under guidance new structural prior, call it interface. Specifically, our method first investigates Marginal Detector (MAD)...

Point cloud completion aims to infer a complete shape from its partial observation. Many approaches utilize pure encoderdecoder paradigm in which can be directly predicted by priors learned scans, however, these methods suffer the loss of details inevitably due feature abstraction issues. In this paper, we propose novel framework,termed SPAC-Net, that rethink task under guidance new structural prior, call it interface. Specifically, our method first investigates Marginal Detector (MAD)...

Approximate computing is a new paradigm. One important area of it designing approximate circuits for FPGA. Modern FPGAs support dual-output LUT, which can significantly reduce the FPGA designs. Several existing works explored use in computing. However, they are limited to small-scale arithmetic circuits. To address problem, this work proposes QUADOL, quality-driven ALS method by exploiting LUTs modern FPGAs. We propose technique approximately merge two single-output (i.e., LUT pair) into...

In this paper, we introduce \textit{DIA}, dissolving is amplifying. DIA a fine-grained anomaly detection framework for medical images. We describe two novel components in the paper. First, \textit{dissolving transformations}. Our main observation that generative diffusion models are feature-aware and applying them to images certain manner can remove or diminish discriminative features such as tumors hemorrhaging. Second, an \textit{amplifying framework} based on contrastive learning learn...

Light Detection and Ranging (LiDAR) has become increasingly popular for on-site wave monitoring due to its high temporal spatial resolution. However, potential application laboratory not been well studied. In this study, a LiDAR was mounted, seven capacitance probes were arranged in flume with slope, study the ability of profiles laboratory. The results demonstrate good agreement between measurements probes, measurement error less than 5 mm. Effective sampling ratio (The number received...

Ternary content addressable memory (TCAM) is a widely used component for high-speed lookup operation. In this work, we advocate novel use of TCAM, i.e., implementing Boolean function. We further leverage approximate match to reduce the resource usage. To achieve this, two extra columns are added TCAM-based architecture. The experimental results show that support implementation any 4-input functions, proposed architecture can 37.5% rows and 12.5% bit cells over conventional

We propose the use of fractals as a means efficient data augmentation. Specifically, we employ plasma for adapting global image augmentation transformations into continuous local transforms. formulate diamond square algorithm cascade simple convolution operations allowing computation on GPU. present TorMentor framework that is totally modular and deterministic across images point-clouds. All can be combined through pipelining random branching to form flow networks arbitrary width depth....