Feature Compression Research Articles

Empowering legal justice with AI: A reinforcement learning SAC-VAE framework for advanced legal text summarization.

Automated summarization of legal texts poses a significant challenge due to the complex and specialized nature of legal documentation. Despite the recent progress in reinforcement learning for natural language text summarization, its application in the legal domain has been less effective. This paper introduces SAC-VAE, a novel reinforcement learning framework specifically designed for legal text summarization. We leverage a Variational Autoencoder (VAE) to condense the high-dimensional state space into a more manageable lower-dimensional feature space. These compressed features are subsequently utilized by the Soft Actor-Critic (SAC) algorithm for policy learning, facilitating the automated generation of summaries from legal texts. Through comprehensive experimentation, we have empirically demonstrated the effectiveness and superior performance of the SAC-VAE framework in legal text summarization.

Detection of mild sensory hearing loss using a joint reflection-distortion otoacoustic emission profile.

Measuring and analyzing both nonlinear-distortion and linear-reflection otoacoustic emissions (OAEs) combined creates what we have termed a "joint-OAE profile." Here, we test whether these two classes of emissions have different sensitivities to hearing loss and whether our joint-OAE profile can detect mild-moderate hearing loss better than conventional OAE protocols have. 2f1-f2 distortion-product OAEs and stimulus-frequency OAEs were evoked with rapidly sweeping tones in 300 normal and impaired ears. Metrics included OAE amplitude for fixed-level stimuli as well as slope and compression features derived from OAE input/output functions. Results show that mild-moderate hearing loss impacts distortion and reflection emissions differently. Clinical decision theory was applied using OAE metrics to classify all ears as either normal-hearing or hearing-impaired. Our best OAE classifiers achieved 90% or better hit rates (with false positive rates of 5%-10%) for mild hearing loss, across a nearly five-octave range. In summary, results suggest that distortion and reflection emissions have distinct sensitivities to hearing loss, which supports the use of a joint-OAE approach for diagnosis. Results also indicate that analyzing both reflection and distortion OAEs together to detect mild hearing loss produces outstanding accuracy across the frequency range, exceeding that achieved by conventional OAE protocols.

Global–Local Query-Support Cross-Attention for Few-Shot Semantic Segmentation

Few-shot semantic segmentation (FSS) models aim to segment unseen target objects in a query image with scarce annotated support samples. This challenging task requires an effective utilization of support information contained in the limited support set. However, the majority of existing FSS methods either compressed support features into several prototype vectors or constructed pixel-wise support-query correlations to guide the segmentation, which failed in effectively utilizing the support information from the global–local perspective. In this paper, we propose Global–Local Query-Support Cross-Attention (GLQSCA), where both global semantics and local details are exploited. Implemented with multi-head attention in a transformer architecture, GLQSCA treats every query pixel as a token, aggregates the segmentation label from the support mask values (weighted by the similarities with all foreground prototypes (global information)), and supports pixels (local information). Experiments show that our GLQSCA significantly surpasses state-of-the-art methods on the standard FSS benchmarks PASCAL-5i and COCO-20i.

Explainable ensemble learning graphical user interface for predicting rebar bond strength and failure mode in recycled coarse aggregate concrete

Novel study deploys robust machine learning algorithms using newly built comprehensive dataset to predict reinforcing rebar-to-recycled coarse aggregate concrete (RCA) bond strength and failure mode. Prior investigations have solely concentrated on bond strength, resulting in a limited comprehension of the bond failure pattern. Considering the increasing significance of sustainable construction methods, it is crucial to examine both the failure pattern and bond strength to expand the versatility of RCA in various reinforced concrete structures. Accordingly, XGBoost, CatBoost, Random Forest, and LightGBM were trained for this purpose. Model performance was appraised using various statistical metrics, while failure classification performance was assessed using accuracy, recall, and precision indicators. Model performance was ranked using Copeland's algorithm. Feature importance was quantified using SHAP. Coefficient of determination of 0.91 was achieved by XGBoost in predicting bond strength, outperforming other nine analytical models in literature. Failure mode was predicted with accuracy of 94% by CatBoost, XGBoost, and LightGBM. Embedment length and compressive strength features had greatest influence on bond strength and failure mode, respectively. User-friendly graphical interface was developed to harvest ML models in real-world engineering practice. Online free access accurately assigns to any given combination of input features corresponding accurate rebar bond strength and failure mode.

High-precision detection of dibutyl hydroxytoluene in edible oil via convolutional autoencoder compressed Fourier-transform near-infrared spectroscopy

The quality of edible oils is closely related to their chemical compositions. Antioxidants have widespread application in edible oil production. In this study, a pioneering detection approach involving the use of a one-dimensional convolutional autoencoder (1D-CAE) was introduced to compress spectral data for assessing antioxidant levels in edible oils. Fourier-transform near-infrared (FT-NIR) characterisation of edible oil samples with varying antioxidant concentrations was also conducted. An 1D-CAE model was developed to compress different pre-processed spectra into a condensed representation. These compressed features were then integrated with a support vector machine and partial least squares regression models to establish correlations for each target. The study examined the influence of pre-processing steps and feature engineering methods on near-infrared spectral analysis through independent or combined model analysis. The findings revealed that features derived from the 1D-CAE model demonstrated remarkable repeatability and can be utilised to construct robust detection models. The experimental results showed that the optimal detection model derived based on the 1D-CAE compression features has an average R2, RPD and RMSE of 0.9953, 15.1664 and 1.2035, respectively, on the prediction set. FT-NIR spectroscopy can be used to accurately detect butylated hydroxytoluene in edible oils. Therefore, autoencoders are an effective tool in spectroscopic analysis, offering promising avenues for future research and application.

Automatic identification and classification of surface defects in small-sized optics

Abstract Precision optics are widely used in lighting systems, imaging systems, and high-precision inspection equipment. The presence of surface defects in optics can seriously affect the design performance of the equipment, so the optics must be rigorously inspected to eliminate defective optics. Currently, engineering products are moving toward smaller sizes, resulting in smaller optics in them also moving towards smaller sizes. However, realizing the surface defect detection of small-sized optics is a great challenge because it requires micrometer-scale high-resolution image acquisition and automatic detection of defects on their surfaces. A machine learning-based automatic surface defect detection method is proposed for surface defects with many types and small sizes. A dark-field micro-scattering imaging system is used to acquire dark-field images. Then the image segmentation and feature analysis are performed on each image to extract the original feature data, and the original feature data are compressed by reliefF algorithm. Based on the compressed feature data, a classification model is built using a support vector machine. The detection method can classify three kinds of defects, namely, dig, scratch, and scuff mark. The accuracy of the method is up to 98%. The experimental results of this study show that the method can automatically and accurately detect a wide range of micro defects occurring in small-sized optics, thus providing valuable insights into the realization of mass production of small-sized optics. In addition, our proposed method provides some ideas for further research in the field of defect detection in small-sized optics.

Analog Front-End for the Readout of LGAD-Based Particle Detectors

This paper presents an analog front-end channel for the readout of Low Gain Avalanche Diodes (LGADs) based particle detectors for the next generation of space-borne experiments. The circuit has been designed in a 65 nm CMOS technology. The charge generated by the detector is integrated by a Charge Sensitive Amplifier (CSA) with dynamic signal compression feature to achieve improved noise performance, along with enhanced time resolution capabilities over the wide range of input charge to be detected (from 45fC to 16pC). An RC-CR shaper with selectable shaping time, a fast discriminator and a Time-to-Amplitude Converter (TAC) complete the readout circuit. An SNR around 300, together with a time resolution of slightly below 60ps, is achieved in post-layout simulations.

Deep Neural Network with a Characteristic Analysis for Seal Stroke Recognition

Seal characters are derived from ancient Chinese pictographs, naturally inheriting pictographic characteristics and complex structures. As the essential components of seal characters, seal strokes play a vital role in seal character recognition, composition and writing, so accurate recognition of seal strokes can greatly promote the investigation of seal characters. Inspired by curve fitting, we propose a new model called the characteristic analysis neural network (CANN) for seal stroke recognition. Instead of indiscriminate grasping of feature information in regular neural networks, we design an efficient approximation technique based on the piecewise Bezier curves that can effectively facilitate structural compression and lossless feature extraction. The feature extraction capability of Bezier approximation helps the methodology achieve impressive recognition accuracy not only on the seal strokes but also on any curve-based symbols. Furthermore, the hierarchical structure of the deep learning strategy is inherited and improved for better performance with high generalisation. Experiments conducted on different types of strokes verify that CANN obtains superior performance on both seal strokes and other smooth symbols. The robustness and the effectiveness of CANN are also demonstrated with minimal learning cost compared to other state-of-art models.

Radio Signature of the Strong Compression between a Streamer and a Coronal Hole Boundary

We present evidence of the first detection of the radio signature at metric wavelengths of the strong compression between a helmet streamer (HS) and the boundary of a coronal hole (CH) using radio observations from the Callisto MEXICO-LANCE and ALASKA-HAARP systems and white-light observations obtained by the STEREO-A/COR1-COR2 coronagraphs. The event occurred very close to the Sun (∼3.4 solar radii) and produced an intense and unusually broad drifting radio feature at metric wavelengths after a downward-drifting band of emission related to a metric Type II radio burst. The compression is caused by the interaction between an expanding structure (coronal mass ejection/shock) and the HS against the CH boundary. Observations in white light show a sharp compressive feature that propagates radially outward, while STEREO-A/EUVI images show loop oscillations at the same position angle, indicating that the interaction occurs across a range of heights. The loop oscillations cease when the compressive front loses its sharp boundary. This transition indicates a reduction of the density compression at the front and the cessation of the radio emission.

Two-Stream Convolutional Neural Networks for Breathing Pattern Classification: Real-Time Monitoring of Respiratory Disease Patients.

A two-stream convolutional neural network (TCNN) for breathing pattern classification has been devised for the continuous monitoring of patients with infectious respiratory diseases. The TCNN consists of a convolutional neural network (CNN)-based autoencoder and classifier. The encoder of the autoencoder generates deep compressed feature maps, which contain the most important information constituting data. These maps are concatenated with feature maps generated by the classifier to classify breathing patterns. The TCNN, single-stream CNN (SCNN), and state-of-the-art classification models were applied to classify four breathing patterns: normal, slow, rapid, and breath holding. The input data consisted of chest tissue hemodynamic responses measured using a wearable near-infrared spectroscopy device on 14 healthy adult participants. Among the classification models evaluated, random forest had the lowest classification accuracy at 88.49%, while the TCNN achieved the highest classification accuracy at 94.63%. In addition, the proposed TCNN performed 2.6% better in terms of classification accuracy than an SCNN (without an autoencoder). Moreover, the TCNN mitigates the issue of declining learning performance with increasing network depth, as observed in the SCNN model. These results prove the robustness of the TCNN in classifying breathing patterns despite using a significantly smaller number of parameters and computations compared to state-of-the-art classification models.

A SoC-FPGA based readout platform for the free-running AMBER data acquisition system

Owing to a change in the scope of physics research, the AMBER spectrometer at CERN is undergoing an update on its instrumentation and trigger strategy for the data acquisition system. One of the key updates is the adoption of a free-running and trigger-less operation, which differs from the predecessor of AMBER, COMPASS, by working in a continuous mode but in an event-based triggered way.In this article, we present a multichannel data-acquisition platform developed in line with the next generation of trigger-less and free-running data acquisition systems. The platform is based on a pre-existing Mezzanine Sampling ADC board for the analog-to-digital conversion and a Xilinx Zynq Ultrascale+ System on Module for online real-time data processing. We also present the evaluation of the system operating in continuous mode, taking data from a 25-element electromagnetic calorimeter prototype with a muon beam, during the first AMBER pilot run. The acquired data were used for noise and pulse shape studies required for the design of the algorithms for lossless compression and data feature extraction needed for trigger-less operation.

A New Dual-Branch Embedded Multivariate Attention Network for Hyperspectral Remote Sensing Classification

With the continuous maturity of hyperspectral remote sensing imaging technology, it has been widely adopted by scholars to improve the performance of feature classification. However, due to the challenges in acquiring hyperspectral images and producing training samples, the limited training sample is a common problem that researchers often face. Furthermore, efficient algorithms are necessary to excavate the spatial and spectral information from these images, and then, make full use of this information with limited training samples. To solve this problem, a novel two-branch deep learning network model is proposed for extracting hyperspectral remote sensing features in this paper. In this model, one branch focuses on extracting spectral features using multi-scale convolution and a normalization-based attention module, while the other branch captures spatial features through small-scale dilation convolution and Euclidean Similarity Attention. Subsequently, pooling and layering techniques are employed to further extract abstract features after feature fusion. In the experiments conducted on two public datasets, namely, IP and UP, as well as our own labeled dataset, namely, YRE, the proposed DMAN achieves the best classification results, with overall accuracies of 96.74%, 97.4%, and 98.08%, respectively. Compared to the sub-optimal state-of-the-art methods, the overall accuracies are improved by 1.05, 0.42, and 0.51 percentage points, respectively. The advantage of this network structure is particularly evident in unbalanced sample environments. Additionally, we introduce a new strategy based on the RPNet, which utilizes a small number of principal components for feature classification after dimensionality reduction. The results demonstrate its effectiveness in uncovering compressed feature information, with an overall accuracy improvement of 0.68 percentage points. Consequently, our model helps mitigate the impact of data scarcity on model performance, thereby contributing positively to the advancement of hyperspectral remote sensing technology in practical applications.

ICC-BiFormer: A Deep-Learning Model for Near-Earth Asteroid Detection via Image Compression and Local Feature Extraction

Detecting near-Earth asteroids (NEAs) is crucial for research in solar system and planetary science. In recent year, deep-learning methods have almost dominated the task. Since NEAs represent only one-thousandth of the pixels in images, we proposed an ICC-BiFormer model that includes an image compression and contrast enhancement block and a BiFormer model to capture local features in input images, which is different from previous models based on Convolutional Neural Network (CNN). Furthermore, we utilize a larger input size of the model, which corresponds to the side length of the input image matrix, and design a cropping algorithm to prevent NEAs from being truncated and better divide NEAs and satellites. We apply our ICC-BiFormer model into a dataset of approximately 20,000 streak and 40,000 non-streak images to train a binary classification model. The ICC-BiFormer achieves 99.88% accuracy, which is superior to existing models. Focusing on local features has been proven effective in detecting NEAs.

Nondestructive evaluation of harvested cabbage texture quality using 3D scanning technology

Current nondestructive methods for evaluating the texture quality of leafy vegetables have limitations due to their complicated leafy structure. In this study, a promising solution was proposed using 3D scanning technology to nondestructively assess the texture quality of leafy vegetables, and the harvested cabbages were chosen as the experimental samples. The cabbages were scanned to extract the morphological traits, especially for surface features of vein distribution. Results demonstrated that morphological traits exhibited better correlations with texture indices compared to traditional compression features. Texture indices were well predicted based on the XGBoostR algorithm with high R2 values of over 0.89 and low RMSE values. The texture quality of cabbages at different harvesting times analyzed by linear discrimination analysis also showed well-discriminative results with an accuracy exceeding 98.3%. These results successfully indicated that 3D scanning technology was effective in evaluating the texture quality of cabbages, showcasing its potential in the nondestructive texture evaluation of leafy vegetables.

Microwave Heartprint: A novel non-contact human identification technology based on cardiac micro-motion detection using ultra wideband bio-radar

The existing one-time identity authentication technology cannot continuously guarantee the legitimacy of user identity during the whole human-computer interaction session, and often requires active cooperation of users, which seriously limits the availability. This study proposes a new non-contact identity recognition technology based on cardiac micro-motion detection using ultra wideband (UWB) bio-radar. After the multi-point micro-motion echoes in the range dimension of the human heart surface area were continuously detected by ultra wideband bio-radar, the two-dimensional principal component analysis (2D-PCA) was exploited to extract the compressed features of the two-dimensional image matrix, namely the distance channel-heart beat sampling point (DC-HBP) matrix, in each accurate segmented heart beat cycle for identity recognition. In the practical measurement experiment, based on the proposed multi-range-bin & 2D-PCA feature scheme along with two conventional reference feature schemes, three typical classifiers were selected as representatives to conduct the heart beat identification under two states of normal breathing and breath holding. The results showed that the multi-range-bin & 2D-PCA feature scheme proposed in this paper showed the best recognition effect. Compared with the optimal range-bin & overall heart beat feature scheme, our proposed scheme held an overall average recognition accuracy of 6.16% higher (normal respiration: 6.84%; breath holding: 5.48%). Compared with the multi-distance unit & whole heart beat feature scheme, the overall average accuracy increase was 27.42% (normal respiration: 28.63%; breath holding: 26.21%) for our proposed scheme. This study is expected to provide a new method of undisturbed, all-weather, non-contact and continuous identification for authentication.

Boosted unsupervised feature selection for tumor gene expression profiles

AbstractIn an unsupervised scenario, it is challenging but essential to eliminate noise and redundant features for tumour gene expression profiles. However, the current unsupervised feature selection methods treat all samples equally, which tend to learn discriminative features from simple samples. Besides, it is difficult for these methods to guide the feature selection process by multiple feature subspaces comprehensively. In order to address these problems, a boosted unsupervised feature selection method (BoostUFS) is proposed for tumour gene expression profiles. Specifically, the authors design a boosting scheme to sequentially learn multiple compressed feature subspaces by focusing on ambiguous samples. The uncertainty of samples and the confidence of feature subspaces can be evaluated adaptively by minimising the overall loss of feature subspaces learning. Furthermore, we provide a consensus objective function with L2,1‐norm regularisation to combine these weighted feature subspaces and select discriminative features. Extensive experiments on several real‐world datasets of tumour gene expression profiles are carried out to demonstrate the effectiveness of the proposed method.

Poorly differentiated thyroid cancer: Clinical, pathological, mutational, and outcome analysis.

Poorly differentiated thyroid cancer (PDTC) remains a challenge not only for pathologists and surgeons because of the difficulties associated with the diagnostic process and the compelling need for difficult thyroidectomy, but it is also of high clinical relevance because it is responsible for mortality in non-anaplastic follicular cell-derived thyroid cancer. Cases of PDTC within a 30-year period were reviewed by two independent pathologists. Histological features like atypical mitosis, necrosis, capsular, and vascular invasion were studied. Mutation analysis was done for BRAF, RET/PTC, RAS, and PI3KCA, and P53 was performed using immunohistochemistry. There were 39 patients with a median age of 53 years; 14 patients were more than 55 years of age. At presentation, 38.4% had compressive features and the median tumor size was 9 cm. At presentation, 67.7% had an extrathyroidal extension (ETE). R0 resection was achieved in 41%, with 12 cases resulting in a difficult thyroidectomy. Necrosis was seen in 65.7% and mitosis in 73.3% with well-differentiated components in 41%. The commonest mutation was RAS (23.1%). Survival was higher in the operable group (54.26, 95% confidence interval [CI]: 30.83-77.70 vs. 20.25, 95% CI: 0-54.07) months, respectively; however, 10-year survival was only 5% and only the tumor size and presence of mitosis were independent risk factors. PDTC presents with worrisome features like large size, ETE, and rapid growth. Aggressive surgical resection with extended/radical thyroidectomy may result in better loco-regional control and improved survival. RAS was the frequent mutation detected. It is worthwhile to identify prognostic factors that can predict the course of PDTC.

Deep learning model for human-intuitive shoeprint reconstruction

Shoeprint reconstruction is essential in forensic science, but it is also challenging due to various inconsistencies in the patterns, textures, sizes, abrasions, etc. of shoeprints. The computational reconstruction of sharper and more complete shoeprints is conventionally conducted using handcrafted features, and it often requires human intervention. Prior studies using end-to-end machine learning approaches are limited in number and have not achieved a high level of performance. In this paper, we propose a model named ShoeRec, which employs variational autoencoder (VAE) as a component in a U-Net-like architecture to reconstruct missing regions and borders in shoeprint images. ShoeRec incorporates skip connections to preserve key patterns and employs VAE in the bottleneck to facilitate the reconstruction of desired shoeprints with the restoration of detail as perceived by humans. As a U-Net, the model skips the contextual information from the encoder to the decoder, and the compressed features in the latent space via VAE optimize the probabilistic distribution for reconstructing complete shoeprints. The reconstruction operation is automatically tuned according to the objective function, so as to reduce the structural correlation between the original and projected shoeprint and restore the absent information in the desired shoeprints. To the best of our knowledge, ShoeRec is the first deep learning infusion model that specializes in shoeprint reconstruction. The shoeprints reconstructed by ShoeRec have a close match with the originals in terms of both structures and patterns, and ShoeRec outperforms state-of-the-art generative models in human evaluation.

Guided Wave-Based Early-Stage Debonding Detection and Assessment in Stiffened Panel Using Machine Learning With Deep Auto-Encoded Features

Abstract Debonding between stiffener and base plate is a very common type of damage in stiffened panels. Numerous efforts have been made for debonding assessment in the stiffened panel structure using guided wave-based techniques. However, these studies are limited to the detection of through-the-flange-width debonding (i.e., full debonding). This paper attempts to develop a methodology for the detection and assessment of early-stage debonding (i.e., partial debonding) in the stiffened panel using machine learning (ML) algorithms. An experimentally validated finite element (FE) simulation model is used to create an initial guided wave dataset containing several debonding scenarios. This dataset is processed through a data augmentation process, followed by feature extraction involving higher harmonics of guided waves. Thereafter, the extracted feature is compressed using a deep autoencoder model. The compressed feature is used for hyperparameter tuning, training, and testing of several supervised ML algorithms, and their performance in the identification of debonding zone and prediction of its size is analyzed. Finally, the trained ML algorithms are tested with experimental data showing that the ML algorithms closely predict the zones of debonding and their sizes. The proposed methodology is an advancement in debonding assessment, specifically addressing early-stage debonding in stiffened panels.

Gringotts: An Encrypted Version Control System With Less Trust on Servers

Version Control System (VCS) plays an essential role in software supply chain, as it manages code projects and enables efficient collaboration. For a private repository, where source code is a high-profile asset and needs to be protected, VCS' security is extremely important. Traditional (unencrypted or encrypted) VCS solutions rely on a trusted service provider to host the code and enforce access control, which is not realistic enough for real-world threats. If the service provider peep in or the hackers break into the repository, the read & write privilege to the sensitive code is totally lost. Therefore, we consider whether one can relax the assumption on the server by introducing a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">covert adversary</i> , namely, it may act maliciously, but will not misbehave if it can be caught doing so. However, protecting sensitive code and enforcing access control on a covert adversarial server is a challenging task. Existing encryption-based VCS solutions failed to address this challenge, as they offered limited access control functionalities, introduced heavy key management overhead or storage overhead. Moreover, the crucial feature of compression of the source files were missing in an encrypted and versioned storage. To address these problems, we introduce <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Gringotts</small> , an end-to-end encrypted VCS, tailored for read & write access control, version control and source file compression. We present a formal model and propose a scheme with detailed analysis. We also implement and evaluate <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Gringotts</small> on top-10 most starred code projects on GitHub. The results demonstrate that <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Gringotts</small> introduces low latency (less than 0.3 s) for commit encryption and decryption, supports fine-grained access control and rich version control functionalities with practical performance.