Pre-training Technique Research Articles

PT-KGNN: A framework for pre-training biomedical knowledge graphs with graph neural networks

Biomedical knowledge graphs (KGs) serve as comprehensive data repositories that contain rich information about nodes and edges, providing modeling capabilities for complex relationships among biological entities. Many approaches either learn node features through traditional machine learning methods, or leverage graph neural networks (GNNs) to directly learn features of target nodes in the biomedical KGs and utilize them for downstream tasks. Motivated by the pre-training technique in natural language processing (NLP), we propose a framework named PT-KGNN (Pre-Training the biomedical KG with GNNs) to learn embeddings of nodes in a broader context by applying GNNs on the biomedical KG.We design several experiments to evaluate the effectivity of our proposed framework and the impact of the scale of KGs. The results of tasks consistently improve as the scale of the biomedical KG used for pre-training increases. Pre-training on large-scale biomedical KGs significantly enhances the drug-drug interaction (DDI) and drug-disease association (DDA) prediction performance on the independent dataset. The embeddings derived from a larger biomedical KG have demonstrated superior performance compared to those obtained from a smaller KG.By applying pre-training techniques on biomedical KGs, rich semantic and structural information can be learned, leading to enhanced performance on downstream tasks. it is evident that pre-training techniques hold tremendous potential and wide-ranging applications in bioinformatics.

A Method for Roof Wireframe Reconstruction Based on Self-Supervised Pretraining

Abstract. In this paper, we present a two-stage method for roof wireframe reconstruction employing a self-supervised pretraining technique. The initial stage utilizes a multi-scale mask autoencoder to generate point-wise features. The subsequent stage involves three steps for edge parameter regression. Firstly, the initial edge directions are generated under the guidance of edge point identification. The next step employs edge parameter regression and matching modules to extract the parameters (namely, direction vector and length) of edge representation from the obtained edge features. Finally, a specifically designed edge non-maximum suppression and an edge similarity loss function are employed to optimize the representation of the final wireframe models and eliminate redundant edges. Experimental results indicate that the pre-trained self-supervised model, enriched by the roof wireframe reconstruction task, demonstrates superior performance on both the publicly available Building3D dataset and its post-processed iterations, specifically the Dense dataset, outperforming even traditional methods.

MixFormer: End-to-End Tracking With Iterative Mixed Attention.

Visual object tracking often employs a multi-stage pipeline of feature extraction, target information integration, and bounding box estimation. To simplify this pipeline and unify the process of feature extraction and target information integration, in this paper, we present a compact tracking framework, termed as MixFormer, built upon transformers. Our core design is to utilize the flexibility of attention operations, and we propose a Mixed Attention Module (MAM) for simultaneous feature extraction and target information integration. This synchronous modeling scheme allows us to extract target-specific discriminative features and perform extensive communication between target and search area. Based on MAM, we build our MixFormer trackers simply by stacking multiple MAMs and placing a localization head on top. Specifically, we instantiate two types of MixFormer trackers, a hierarchical tracker MixCvT, and a non-hierarchical simple tracker MixViT. For these two trackers, we investigate a series of pre-training methods and uncover the different behaviors between supervised pre-training and self-supervised pre-training in our MixFormer trackers. We also extend the masked autoencoder pre-training to our MixFormer trackers and design the new competitive TrackMAE pre-training technique. Finally, to handle multiple target templates during online tracking, we devise an asymmetric attention scheme in MAM to reduce computational cost, and propose an effective score prediction module to select high-quality templates. Our MixFormer trackers set a new state-of-the-art performance on seven tracking benchmarks, including LaSOT, TrackingNet, VOT2020, GOT-10 k, OTB100, TOTB and UAV123. In particular, our MixViT-L achieves AUC scores of 73.3% on LaSOT, 86.1% on TrackingNet and 82.8% on TOTB.

Hybrid Masked Image Modeling for 3D Medical Image Segmentation.

Masked image modeling (MIM) with transformer backbones has recently been exploited as a powerful self-supervised pre-training technique. The existing MIM methods adopt the strategy to mask random patches of the image and reconstruct the missing pixels, which only considers semantic information at a lower level, and causes a long pre-training time. This paper presents HybridMIM, a novel hybrid self-supervised learning method based on masked image modeling for 3D medical image segmentation. Specifically, we design a two-level masking hierarchy to specify which and how patches in sub-volumes are masked, effectively providing the constraints of higher level semantic information. Then we learn the semantic information of medical images at three levels, including: 1) partial region prediction to reconstruct key contents of the 3D image, which largely reduces the pre-training time burden (pixel-level); 2) patch-masking perception to learn the spatial relationship between the patches in each sub-volume (region-level); and 3) drop-out-based contrastive learning between samples within a mini-batch, which further improves the generalization ability of the framework (sample-level). The proposed framework is versatile to support both CNN and transformer as encoder backbones, and also enables to pre-train decoders for image segmentation. We conduct comprehensive experiments on five widely-used public medical image segmentation datasets, including BraTS2020, BTCV, MSD Liver, MSD Spleen, and BraTS2023. The experimental results show the clear superiority of HybridMIM against competing supervised methods, masked pre-training approaches, and other self-supervised methods, in terms of quantitative metrics, speed performance and qualitative observations. The codes of HybridMIM are available at https://github.com/ge-xing/HybridMIM.

QDETRv: Query-Guided DETR for One-Shot Object Localization in Videos

In this work, we study one-shot video object localization problem that aims to localize instances of unseen objects in the target video using a single query image of the object. Toward addressing this challenging problem, we extend a popular and successful object detection method, namely DETR (Detection Transformer), and introduce a novel approach –query-guided detection transformer for videos (QDETRv). A distinctive feature of QDETRv is its capacity to exploit information from the query image and spatio-temporal context of the target video, which significantly aids in precisely pinpointing the desired object in the video. We incorporate cross-attention mechanisms that capture temporal relationships across adjacent frames to handle the dynamic context in videos effectively. Further, to ensure strong initialization for QDETRv, we also introduce a novel unsupervised pretraining technique tailored to videos. This involves training our model on synthetic object trajectories with an analogous objective as the query-guided localization task. During this pretraining phase, we incorporate recurrent object queries and loss functions that encourage accurate patch feature reconstruction. These additions enable better temporal understanding and robust representation learning. Our experiments show that the proposed model significantly outperforms the competitive baselines on two public benchmarks, VidOR and ImageNet-VidVRD, extended for one-shot open-set localization tasks.

Feasibility of remote screening for high risk valve disease using a new wearable cardiac monitoring patch with synchronized phonocardiogram and electrocardiogram in communities

Abstract Background The 2010 census data showed that the weighted prevalence rate of valvular disease was 3.8% in China. For many elderly people, COVID-19 pandemic has caused the situation that they have not received medical examination for years. Since many patients with moderate aortic/mitral stenosis/regurgitation do not have significant symptoms, early detection of high-risk groups of valve disease is essential. A new wearable cardiac monitoring patch with synchronized phonocardiogram(PCG) and electrocardiogram(ECG) function is introduced in this study. Unlike handheld auscultation, it minimizes the interference of the environment and the operator, and is not demanding for the operator. Methods We enrolled 3170 patients of over 65 years old. Patients with diagnosed moderate or severe aortic/mitral stenosis/regurgitation, congenital heart disease, hypertrophic cardiomyopathy were excluded. Data acquisition was done by social workers following strict instructions. The patch was attached to the mitral and aortic auscultation area on the patients. ECG and PCG signals can be detected simultaneously and sent to a phone or tablet via Bluetooth. The data would then be auto-uploaded to a cloud-based center for further analysis. The end-to-end algorithm for S1 and S2 heart sounds detection involves a CNN-LSTM model for R-peak detection and a 1-D CNN-based U-Net model for fine-grained segmentation based on the detected heartbeat. The segmentation network takes multi-modal PCG and ECG signals as inputs and applies Contrastive Learning as a pretraining technique for enhancement to a real-world problem: detecting heart murmur from high-risk valve disease. If the results indicate positive, the patient is recommended by family doctors to do ultrasonic cardiogram (UCG) and seek further medical advice (Figure 1). Results 3170 participants were recruited from March 2022 to December 2022 in different communities of Huangpu District in Shanghai(Figure 2). In total, 43.98% were men, and 85.81% of age distribution were among 65-79 years. Among 3170 participants, 119(3.8%) were suspected positive by ECG and PCG. Among 119 patients, 65.55% were among 65-79 years. After telephone follow-up 29(24.3%) patients finished clinical visit and UCG. About 66% were diagnosed as mild aortic/mitral stenosis/regurgitation, 24% were diagnosed as moderate aortic/mitral stenosis/regurgitation, 10% were diagnosed as hypertrophic cardiomyopathy. Conclusions This study demonstrated the feasibility of remote screening for high risk valve disease using a new wearable cardiac monitoring patch in community visit by social workers with no medical background. This technology may improve the efficacy of clinical diagnosis and reduce unnecessary examination, particularly in the absence of doctors' auscultation and UCG. Unfortunately, three quarters of the suspected positive patients lost contact or refused to do UCG due to COVID-19 pandemic. Further investigations are necessary to increase accuracy.Flow chart of community screeningbaseline and disease distribution

MedGCN: An IoT-edge thrombus graph convolutional network for accurate prediction and prescription diagnosis of vascular occlusive diseases from unstructured clinical reports

The timely and accurate prediction and diagnosis of vascular occlusive diseases are pivotal in enhancing patient outcomes. This research addresses this critical healthcare challenge by introducing MedGCN, a novel thrombus graph convolutional network model. MedGCN is specifically designed for the precise prediction of prescriptions and diagnoses based on unstructured clinical diagnostic reports. The model synergistically blends OpenAI’s GPT4 and a uniquely designed Cross Fusion Graph Convolutional Network (CF-GCN) to ensure the meticulous fusion of knowledge. We delve into nine distinct learning tasks, encompassing both prescription and diagnosis and employ a multi-label classification GCN pretraining technique to assess them. Our evaluation underscores MedGCN’s robust predictive prowess across various tasks. By amalgamating cutting-edge AI paradigms with IoT edge computing, this research not only bolsters the efficacy of medical decision-making but also champions patient privacy. The methodologies and findings presented herein hold immense potential for deployment in IoT frameworks, thus proffering swift and precise assistance in medical decision-making and addressing a paramount need in the contemporary healthcare landscape.

Sequential autoencoders for feature engineering and pretraining in major depressive disorder risk prediction.

We evaluated autoencoders as a feature engineering and pretraining technique to improve major depressive disorder (MDD) prognostic risk prediction. Autoencoders can represent temporal feature relationships not identified by aggregate features. The predictive performance of autoencoders of multiple sequential structures was evaluated as feature engineering and pretraining strategies on an array of prediction tasks and compared to a restricted Boltzmann machine (RBM) and random forests as a benchmark. We study MDD patients from Vanderbilt University Medical Center. Autoencoder models with Attention and long-short-term memory (LSTM) layers were trained to create latent representations of the input data. Predictive performance was evaluated temporally by fitting random forest models to predict future outcomes with engineered features as input and using autoencoder weights to initialize neural network layers. We evaluated area under the precision-recall curve (AUPRC) trends and variation over the study population's treatment course. The pretrained LSTM model improved predictive performance over pretrained Attention models and benchmarks in 3 of 4 outcomes including self-harm/suicide attempt (AUPRCs, LSTM pretrained = 0.012, Attention pretrained = 0.010, RBM = 0.009, random forest = 0.005). The use of autoencoders for feature engineering had varied results, with benchmarks outperforming LSTM and Attention encodings on the self-harm/suicide attempt outcome (AUPRCs, LSTM encodings = 0.003, Attention encodings = 0.004, RBM = 0.009, random forest = 0.005). Improvement in prediction resulting from pretraining has the potential for increased clinical impact of MDD risk models. We did not find evidence that the use of temporal feature encodings was additive to predictive performance in the study population. This suggests that predictive information retained by model weights may be lost during encoding. LSTM pretrained model predictive performance is shown to be clinically useful and improves over state-of-the-art predictors in the MDD phenotype. LSTM model performance warrants consideration of use in future related studies. LSTM models with pretrained weights from autoencoders were able to outperform the benchmark and a pretrained Attention model. Future researchers developing risk models in MDD may benefit from the use of LSTM autoencoder pretrained weights.

MpbPPI: a multi-task pre-training-based equivariant approach for the prediction of the effect of amino acid mutations on protein-protein interactions.

The accurate prediction of the effect of amino acid mutations for protein-protein interactions (PPI $\Delta \Delta G$) is a crucial task in protein engineering, as it provides insight into the relevant biological processes underpinning protein binding and provides a basis for further drug discovery. In this study, we propose MpbPPI, a novel multi-task pre-training-based geometric equivariance-preserving framework to predict PPI $\Delta \Delta G$. Pre-training on a strictly screened pre-training dataset is employed to address the scarcity of protein-protein complex structures annotated with PPI $\Delta \Delta G$ values. MpbPPI employs a multi-task pre-training technique, forcing the framework to learn comprehensive backbone and side chain geometric regulations of protein-protein complexes at different scales. After pre-training, MpbPPI can generate high-quality representations capturing the effective geometric characteristics of labeled protein-protein complexes for downstream $\Delta \Delta G$ predictions. MpbPPI serves as a scalable framework supporting different sources of mutant-type (MT) protein-protein complexes for flexible application. Experimental results on four benchmark datasets demonstrate that MpbPPI is a state-of-the-art framework for PPI $\Delta \Delta G$ predictions. The data and source code are available at https://github.com/arantir123/MpbPPI.

CT-CBCT deformable registration using weakly-supervised artifact-suppression transfer learning network

Objective. Computed tomography-cone-beam computed tomography (CT-CBCT) deformable registration has great potential in adaptive radiotherapy. It plays an important role in tumor tracking, secondary planning, accurate irradiation, and the protection of at-risk organs. Neural networks have been improving CT-CBCT deformable registration, and almost all registration algorithms based on neural networks rely on the gray values of both CT and CBCT. The gray value is a key factor in the loss function, parameter training, and final efficacy of the registration. Unfortunately, the scattering artifacts in CBCT affect the gray values of different pixels inconsistently. Therefore, the direct registration of the original CT-CBCT introduces artifact superposition loss. Approach. In this study, a histogram analysis method for the gray values was used. Based on an analysis of the gray value distribution characteristics of different regions in CT and CBCT, the degree of superposition of the artifact in the region of disinterest was found to be much higher than that in the region of interest. Moreover, the former was the main reason for artifact superposition loss. Consequently, a new weakly supervised two-stage transfer-learning network based on artifact suppression was proposed. The first stage was a pre-training network designed to suppress artifacts contained in the region of disinterest. The second stage was a convolutional neural network that registered the suppressed CBCT and CT. Main Results. Through a comparative test of the thoracic CT-CBCT deformable registration, whose data were collected from the Elekta XVI system, the rationality and accuracy after artifact suppression were confirmed to be significantly improved compared with the other algorithms without artifact suppression. Significance. This study proposed and verified a new deformable registration method with multi-stage neural networks, which can effectively suppress artifacts and further improve registration by incorporating a pre-training technique and an attention mechanism.

Unsupervised Instance Discriminative Learning for Depression Detection from Speech Signals.

Major Depressive Disorder (MDD) is a severe illness that affects millions of people, and it is critical to diagnose this disorder as early as possible. Detecting depression from voice signals can be of great help to physicians and can be done without any invasive procedure. Since relevant labelled data are scarce, we propose a modified Instance Discriminative Learning (IDL) method, an unsupervised pre-training technique, to extract augment-invariant and instance-spread-out embeddings. In terms of learning augment-invariant embeddings, various data augmentation methods for speech are investigated, and time-masking yields the best performance. To learn instance-spreadout embeddings, we explore methods for sampling instances for a training batch (distinct speaker-based and random sampling). It is found that the distinct speaker-based sampling provides better performance than the random one, and we hypothesize that this result is because relevant speaker information is preserved in the embedding. Additionally, we propose a novel sampling strategy, Pseudo Instance-based Sampling (PIS), based on clustering algorithms, to enhance spread-out characteristics of the embeddings. Experiments are conducted with DepAudioNet on DAIC-WOZ (English) and CONVERGE (Mandarin) datasets, and statistically significant improvements, with p-value 0.0015 and 0.05, respectively, are observed using PIS in the detection of MDD relative to the baseline without pre-training.

Inverse folding based pre-training for the reliable identification of intrinsic transcription terminators.

It is well-established that neural networks can predict or identify structural motifs of non-coding RNAs (ncRNAs). Yet, the neural network based identification of RNA structural motifs is limited by the availability of training data that are often insufficient for learning features of specific ncRNA families or structural motifs. Aiming to reliably identify intrinsic transcription terminators in bacteria, we introduce a novel pre-training approach that uses inverse folding to generate training data for predicting or identifying a specific family or structural motif of ncRNA. We assess the ability of neural networks to identify secondary structure by systematic in silico mutagenesis experiments. In a study to identify intrinsic transcription terminators as functionally well-understood RNA structural motifs, our inverse folding based pre-training approach significantly boosts the performance of neural network topologies, which outperform previous approaches to identify intrinsic transcription terminators. Inverse-folding based pre-training provides a simple, yet highly effective way to integrate the well-established thermodynamic energy model into deep neural networks for identifying ncRNA families or motifs. The pre-training technique is broadly applicable to a range of network topologies as well as different types of ncRNA families and motifs.

An OCR Post-Correction Approach Using Deep Learning for Processing Medical Reports

According to a recent Deloitte study, the COVID-19 pandemic continues to place a huge strain on the global health care sector. Covid-19 has also catalysed digital transformation across the sector for improving operational efficiencies. As a result, the amount of digitally stored patient data such as discharge letters, scan images, test results or free text entries by doctors has grown significantly. In 2020, 2314 exabytes of medical data was generated globally. This medical data does not conform to a generic structure and is mostly in the form of unstructured digitally generated or scanned paper documents stored as part of a patient’s medical reports. This unstructured data is digitised using Optical Character Recognition (OCR) process. A key challenge here is that the accuracy of the OCR process varies due to the inability of current OCR engines to correctly transcribe scanned or handwritten documents in which text may be skewed, obscured or illegible. This is compounded by the fact that processed text is comprised of specific medical terminologies that do not necessarily form part of general language lexicons. The proposed work uses a deep neural network based self-supervised pre-training technique: Robustly Optimized Bidirectional Encoder Representations from Transformers (RoBERTa) that can learn to predict hidden (masked) sections of text to fill in the gaps of non-transcribable parts of the documents being processed. Evaluating the proposed method on domain-specific datasets which include real medical documents, shows a significantly reduced word error rate demonstrating the effectiveness of the approach.

Improving spatial variation of ground-level PM2.5 prediction with contrastive learning from satellite imagery

Convolutional Neural Networks (CNNs) are a promising technique to predict highly localized fine particulate matter (i.e., PM2.5 levels) based on high-resolution satellite imagery. Unfortunately, CNNs typically require large amounts of supervised data to perform well, whereas this application generally has lots of unsupervised data (all satellite imagery) and relatively sparse supervised data (measurements from ground sensors). Previous work used transfer learning from another visual task to initialize the CNN weights; however, we hypothesize that standard transfer learning strategies would bias the CNN to focus on irrelevant details of the image for our applications. Instead, we develop a novel framework called Spatiotemporal Contrastive Learning (SCL) to pre-train the CNN. We test both regular contrastive learning and SCL on predicting PM2.5 levels from satellite images in two different cities, Delhi and Beijing, and compare to CNNs with parameters initialized randomly and by transfer learning. Our results show that regular contrastive learning and our SCL frameworks both manage to better capture spatial variation of ground-level PM2.5 concentrations compared to traditional initialization schemes, and that this performance gap increases as the number of ground sensors decreases, implying that the approach will be even more valuable in cities with fewer ground sensors. Our work demonstrates that contrastive learning is a powerful pre-training technique to build better spatial maps of PM2.5, and can be broadly applied in related situations.

An intelligent early warning system of analyzing Twitter data using machine learning on COVID-19 surveillance in the US

The World Health Organization (WHO) declared on 11th March 2020 the spread of the coronavirus disease 2019 (COVID-19) a pandemic. The traditional infectious disease surveillance had failed to alert public health authorities to intervene in time and mitigate and control the COVID-19 before it became a pandemic. Compared with traditional public health surveillance, harnessing the rich data from social media, including Twitter, has been considered a useful tool and can overcome the limitations of the traditional surveillance system. This paper proposes an intelligent COVID-19 early warning system using Twitter data with novel machine learning methods. We use the natural language processing (NLP) pre-training technique, i.e., fine-tuning BERT as a Twitter classification method. Moreover, we implement a COVID-19 forecasting model through a Twitter-based linear regression model to detect early signs of the COVID-19 outbreak. Furthermore, we develop an expert system, an early warning web application based on the proposed methods. The experimental results suggest that it is feasible to use Twitter data to provide COVID-19 surveillance and prediction in the US to support health departments’ decision-making.

Deep learning model construction for a semi-supervised classification with feature learning

Several deep models were proposed in image processing, data interpretation, speech recognition, and video analysis. Most of these architectures need a massive proportion of training samples and use arbitrary configuration. This paper constructs a deep learning architecture with feature learning. Graph convolution networks (GCNs), semi-supervised learning and graph data representation, have become increasingly popular as cost-effective and efficient methods. Most existing merging node descriptions for node distribution on the graph use stabilised neighbourhood knowledge, typically requiring a significant amount of variables and a high degree of computational complexity. To address these concerns, this research presents DLM-SSC, a unique method semi-supervised node classification tasks that can combine knowledge from multiple neighbourhoods at the same time by integrating high-order convolution and feature learning. This paper employs two function learning techniques for reducing the number of parameters and hidden layers: modified marginal fisher analysis (MMFA) and kernel principal component analysis (KPCA). The MMFA and KPCA weight matrices are modified layer by layer when implementing the DLM, a supervised pretraining technique that doesn't require a lot of information. Free measuring on citation datasets (Citeseer, Pubmed, and Cora) and other data sets demonstrate that the suggested approaches outperform similar algorithms.

RSTC: A New Residual Swin Transformer for Offline Word-Level Writer Identification

Writer identification has steadily progressed in recent decades owing to its widespread application. Scenarios with extensive handwriting data such as page-level or sentence-level have achieved satisfactory accuracy, however, word-level offline writer identification is still challenging owing to the difficulty of learning good feature representations with scant handwriting data. This paper proposes a new Residual Swin Transformer Classifier (RSTC), comprehensively aggregating local and global handwriting styles and yielding robust feature representations with single-word images. Local information is modeled by the Transformer Block through interacting strokes. Global information is featurized by holistic encoding using the Identity Branch and Global Block. Moreover, the pre-training technique is exploited to transfer reusable knowledge learned from a task similar to writer identification, strengthening the model’s representation of handwriting features. The proposed method is tested on the IAM and CVL benchmark datasets and achieves state-of-the-art performance, which demonstrates RSTC’s superior modeling capability on word-level writer identification.

A cross-sectional study of pre- and posttraining evaluation of inhaler use technique among outpatients with bronchial asthma or chronic obstructive pulmonary disease at a tertiary care hospital in India.

Management of asthma and chronic obstructive pulmonary disease (COPD) includes use of inhalers as an integral component of drug delivery. Awareness about inhaler devices along with demonstration of the usage of inhaler technique aids in the optimization of therapeutic outcome. This study aimed to assess the correct use of inhaler technique following pre- and posttraining sessions for the management of COPD and asthma among pulmonary outpatients at a tertiary care hospital. This prospective cross-sectional study was carried out among patients diagnosed with asthma or COPD and prescribed with inhaler medication in the form of pressurized metered-dose inhaler (pMDI), MDI with spacer, or dry powder inhaler (DPI). The inhaler device use by the patients was assessed initially, followed by a demonstration on inhaler device technique, and reassessed post training using a checklist. Comparison of the median total score of pretraining and posttraining inhaler technique was analyzed by Wilcoxon signed rank test. P < 0.05 was considered statistically significant. Out of 144 patients, 55.6%, 27.7%, and 16.7% of them were prescribed pMDI, MDI with spacer, and DPI, respectively. Post inhaler technique training, about 79.2% of the patients were able to demonstrate the inhaler technique correctly compared to 52.1% prior to training. A statistically significant difference in the median score of inhaler technique has been observed before and after training. This study reports a significant improvement in the correct use of inhaler technique post training. In addition, the most frequent error among inhaler users was revealed to be in the breath actuation.

Self-supervised pretraining for transferable quantitative phase image cell segmentation.

In this paper, a novel U-Net-based method for robust adherent cell segmentation for quantitative phase microscopy image is designed and optimised. We designed and evaluated four specific post-processing pipelines. To increase the transferability to different cell types, non-deep learning transfer with adjustable parameters is used in the post-processing step. Additionally, we proposed a self-supervised pretraining technique using nonlabelled data, which is trained to reconstruct multiple image distortions and improved the segmentation performance from 0.67 to 0.70 of object-wise intersection over union. Moreover, we publish a new dataset of manually labelled images suitable for this task together with the unlabelled data for self-supervised pretraining.

Deep Learning for SWIPT: Optimization of Transmit-Harvest-Respond in Wireless-Powered Interference Channel

In this paper, we consider a wireless-powered two-way communication, called <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">transmit-harvest-respond</i> , with co-channel interference. The two-way communication considered here comprises three steps: i) transmitters send data signals, ii) receivers decode information and harvest energy simultaneously from the received signals using a policy of time switching (TS) or power splitting (PS), and iii) receivers transmit responses back to transmitters using this harvested energy. We aim to find the transmit power and energy harvesting ratios that maximize the sum rate of the forward links while ensuring a minimum rate requirement for each backward link. Due to the non-convexity and NP hardness of the optimization problem considered here, we first derive suboptimal solutions using an iterative algorithm (IA) on the basis of asymptotic strong duality. In view of the high computation time of the IA, we then design an efficient deep neural network (DNN) framework and novel training strategy as a means of combining supervised and unsupervised training. Specifically, DNNs are pre-trained using the suboptimal solutions obtained by the IA in a supervised manner, as a means of initialization; further training is then applied to DNNs using a well-designed loss function in an unsupervised manner to enhance performance. Simulation results reveal that the pre-training technique using IA solutions is beneficial for improving the performance of the DNN. The proposed hybrid scheme thus achieves near-optimal performances with a lower computation time, compared with the use of IA or DNN alone.