Network Models Research Articles

Biological oxygen demand prediction using artificial neural network and random forest models enhanced by the neural architecture search algorithm

Biological oxygen demand prediction using artificial neural network and random forest models enhanced by the neural architecture search algorithm

TCSRNet: a lightweight tobacco leaf curing stage recognition network model

Due to the constraints of the tobacco leaf curing environment and computational resources, current image classification models struggle to balance recognition accuracy and computational efficiency, making practical deployment challenging. To address this issue, this study proposes the development of a lightweight classification network model for recognizing tobacco leaf curing stages (TCSRNet). Firstly, the model utilizes an Inception structure with parallel convolutional branches to capture features at different receptive fields, thereby better adapting to the appearance variations of tobacco leaves at different curing stages. Secondly, the incorporation of Ghost modules significantly reduces the model’s computational complexity and parameter count through parameter sharing, enabling efficient recognition of tobacco leaf curing stages. Lastly, the design of the Multi-scale Adaptive Attention Module (MAAM) enhances the model’s perception of key visual information in images, emphasizing distinctive features such as leaf texture and color, which further improves the model’s accuracy and robustness. On the constructed tobacco leaf curing stage dataset (with color images sized 224×224 pixels), TCSRNet achieves a classification accuracy of 90.35% with 158.136 MFLOPs and 1.749M parameters. Compared to models such as ResNet34, GhostNet, ShuffleNetV2×1.5, EfficientNet-b0, MobileViT-xs, MobileNetV2, MobileNetV3-large, and MobileNetV3-small, TCSRNet demonstrates superior performance in terms of accuracy, FLOPs, and parameter count. Furthermore, when evaluated on the public V2 Plant Seedlings dataset, TCSRNet maintains an impressive accuracy of 97.15% compared to other advanced network models. This research advances the development of lightweight models for recognizing tobacco leaf curing stages, providing theoretical support for smart tobacco curing technologies and injecting new momentum into the digital transformation of the tobacco industry.

A network approach to understanding occupational psychological distress: linking depression, anxiety, and burnout among Chinese healthcare professionals

ObjectivesAs a population at high risk for psychological distress, healthcare workers typically experience varying degrees of anxiety, depression, and burnout. Studies have found that depression and anxiety have a negative impact on the mental health domain of burnout in healthcare workers. However, little is known about the symptom-to-symptom interactions between these psychological outcomes. This study aims to elucidate the characteristics of depression, anxiety, and burnout networks among healthcare workers.MethodsWe recruited 846 healthcare workers from March to April 2023 from three hospitals. A total of 826 healthcare workers completed the General Information Scale, the 16-item Quick Inventory of Depressive Symptomatology–Self-Report (QIDS-SR16), the Self-rating Anxiety Scale (SAS), and the Burnout Clinical Subtype Questionnaire (BCSQ)-36. The network models were constructed using network analysis. The expected influence and the bridge expected influence of nodes were calculated. The stability and accuracy of the network was assessed.ResultsThe results showed that the core symptoms in the symptom network mainly included QIDS8 (Energy/Fatigability), SAS3 (Easily upset or frightened), SAS11 (Dizzy), SAS8 (Tiredness), SAS10 (Tachycardia) and BCSQ3 (Worn-out), and the key nodes connecting these symptoms were QIDS2 (Sad mood), SAS20 (Have nightmares), BCSQ3 (Worn-out), SAS8 (Tiredness), QIDS8 (Energy/Fatigability), QIDS4 (Concentration/decision-making) and SAS4 (Madness).ConclusionUnique pathways of association between burnout, depression, and anxiety were found to exist. Interventions targeting core symptoms can maximize the improvement of depression, anxiety, and burnout, provide a deeper understanding of the relationship between the three conditions, and provide a target and basis for psychological interventions to improve the emotional wellbeing of healthcare workers and enhance their mental health.

Detection and classification of long terminal repeat sequences in plant LTR-retrotransposons and their analysis using explainable machine learning

BackgroundLong terminal repeats (LTRs) represent important parts of LTR retrotransposons and retroviruses found in high copy numbers in a majority of eukaryotic genomes. LTRs contain regulatory sequences essential for the life cycle of the retrotransposon. Previous experimental and sequence studies have provided only limited information about LTR structure and composition, mostly from model systems. To enhance our understanding of these key sequence modules, we focused on the contrasts between LTRs of various retrotransposon families and other genomic regions. Furthermore, this approach can be utilized for the classification and prediction of LTRs.ResultsWe used machine learning methods suitable for DNA sequence classification and applied them to a large dataset of plant LTR retrotransposon sequences. We trained three machine learning models using (i) traditional model ensembles (Gradient Boosting), (ii) hybrid convolutional/long and short memory network models, and (iii) a DNA pre-trained transformer-based model using k-mer sequence representation. All three approaches were successful in classifying and isolating LTRs in this data, as well as providing valuable insights into LTR sequence composition. The best classification (expressed as F1 score) achieved for LTR detection was 0.85 using the hybrid network model. The most accurate classification task was superfamily classification (F1=0.89) while the least accurate was family classification (F1=0.74). The trained models were subjected to explainability analysis. Positional analysis identified a mixture of interesting features, many of which had a preferred absolute position within the LTR and/or were biologically relevant, such as a centrally positioned TATA-box regulatory sequence, and TG..CA nucleotide patterns around both LTR edges.ConclusionsOur results show that the models used here recognized biologically relevant motifs, such as core promoter elements in the LTR detection task, and a development and stress-related subclass of transcription factor binding sites in the family classification task. Explainability analysis also highlighted the importance of 5’- and 3’- edges in LTR identity and revealed need to analyze more than just dinucleotides at these ends. Our work shows the applicability of machine learning models to regulatory sequence analysis and classification, and demonstrates the important role of the identified motifs in LTR detection.

Predicting Starch Content in Early Potato Varieties Using Neural Networks and Regression Models: A Comparative Study

Starch content serves as a crucial indicator of the quality and palatability of potato tubers. It has become a common practice to evaluate the polysaccharide content directly in tubers freshly harvested from the field. This study aims to develop models that can predict starch content prior to the harvesting of potato tubers. Very early potato varieties were cultivated in the northern and northwestern regions of Poland. The research involved constructing multiple linear regression (MLR) and artificial neural network (ANN-MLP) models, drawing on data from eight years of field trials. The independent variables included factors such as sunshine duration, average daily air temperatures, precipitation, soil nutrient levels, and phytophenological data. The NSM demonstrated a higher accuracy in predicting the dependent variable compared to the RSM, with MAPE errors of 7.258% and 9.825%, respectively. This study confirms that artificial neural networks are an effective tool for predicting starch content in very early potato varieties, making them valuable for monitoring potato quality.

TEESlice: Protecting Sensitive Neural Network Models in Trusted Execution Environments When Attackers have Pre-Trained Models

Trusted Execution Environments (TEE) are used to safeguard on-device models. However, directly employing TEEs to secure the entire DNN model is challenging due to the limited computational speed. Utilizing GPU can accelerate DNN's computation speed but commercial widely-available GPUs usually lack security protection. To this end, scholars introduce TEE-shielded DNN partition (TSDP), a method that protects privacy-sensitive weights within TEEs and offloads insensitive weights to GPUs. Nevertheless, current methods do not consider the presence of a knowledgeable adversary who can access abundant publicly available pre-trained models and datasets. This paper investigates the security of existing methods against such a knowledgeable adversary and reveals their inability to fulfill their security promises. Consequently, we introduce a novel partition before training strategy, which effectively separates privacy-sensitive weights from other components of the model. Our evaluation demonstrates that our approach can offer full model protection with a computational cost reduced by a factor of 10. In addition to traditional CNN models, we also demonstrate the scalability to large language models. Our approach can compress the private functionalities of the large language model to lightweight slices and achieve the same level of protection as the shielding-whole-model baseline.

Runoff prediction using combined machine learning models and signal decomposition

ABSTRACT Accurate forecasting of increasingly unpredictable river runoff is essential for effective water resource management in the face of climate change and human activities. This study uses four machine learning models of long short-term memory neural networks (LSTM), support vector machine (SVM), random forest, and artificial neural network models to improve runoff forecasting accuracy and explore combined forecasting models’ effectiveness. This study develops three advanced combined forecasting models (empirical mode decomposition (EMD)–LSTM, VMD–LSTM, wavelet analysis (WA)–LSTM) by combining preprocessing techniques of EMD, variational mode decomposition (VMD), and WA with the LSTM modeling method. These models use signal decomposition techniques to analyze 41 years of runoff data from the Huanren station (1980–2020). The findings reveal that the LSTM model outperforms the other three individual machine learning models when forecasting days with high runoff. Among the three decomposed combined models, the VMD–LSTM model demonstrates the best overall performance during the validation period, achieving root mean square error, Nash–Sutcliffe efficiency coefficient, and bias values of 52.14 m3/s, 0.96, and −0.002, respectively. The combination of LSTM with signal decomposition techniques shows promising potential for enhancing runoff prediction accuracy, with practical implications for water resource management and flood control strategies.

Effective many-body interactions in reduced-dimensionality spaces through neural network models

Accurately describing properties of challenging problems in physical sciences often requires complex mathematical models that are unmanageable to tackle head on. Therefore, developing reduced-dimensionality representations that encapsulate complex correlation effects in many-body systems is crucial to advance the understanding of these complicated problems. However, a numerical evaluation of these predictive models can still be associated with a significant computational overhead. To address this challenge, in this paper we discuss a combined framework that integrates recent advances in the development of active-space representations of coupled cluster (CC) downfolded Hamiltonians with neural network approaches. The primary objective of this effort is to train neural networks to eliminate the computationally expensive steps required for evaluating hundreds or thousands of Hugenholtz diagrams, which correspond to multidimensional tensor contractions necessary for evaluating a many-body form of downfolded effective Hamiltonians. Using small molecular systems (the H2O and HF molecules) as examples, we demonstrate that training neural networks employing effective Hamiltonians for a few nuclear geometries of molecules can accurately interpolate or extrapolate their forms to other geometrical configurations characterized by different intensities of correlation effects. We also discuss differences between effective interactions that define CC downfolded Hamiltonians with those of bare Hamiltonians defined by Coulomb interactions in the active spaces. Published by the American Physical Society 2024

Sim-Boxing Fraud Detection System Using Artificial Neural Network

ABSTRACT Telecommunication fraud is when con artists use pretense to obtain discounted or complimentary services. Globally, prevalent telecommunication fraud results in significant yearly revenue losses for telecom firms. The SIM-box bypass fraud is one tactic scammers use to deceive telecom operators. Installing numerous prepaid SIM cards in a SIM box is known as SIM-box fraud. This makes it possible for fraudsters to place international calls using local phone numbers in the destination countries and receive calls over VoIP (the Internet). By starting the call using the local SIM that is inserted in the SIM box, it appears to be local. This study created a model that employed Artificial Neural Network (ANN) model to more accurately and efficiently identify fraudulent customers; diverted SIM-Box calls using Customer Detail Record (CDR) data. A framework using multi-layer perceptrons, a class of ANN, to identify patterns indicative of SIM-box operations by analysing Call Detail records (CDRs) was developed. The model was trained using 500 iterations on a dataset comprising labelled records of both legitimate and fraudulent calls, using features such as call duration, frequency of calls, timestamp patterns, and originating and terminating number characteristics. The range of prediction accuracy obtained from the developed artificial neural network models was 56.8% to 62.3%. The top model, consisting of five layers, yielded 62.3%. These layers consist of an input layer with neurons, five, nine, and eighteen hidden levels, and one output layer with a single neuron that stands for fraud. It is recommended that the dataset be expanded to include a wider range of SIM-box fraud scenarios from various geographic locations and over different periods. This would improve the model’s robustness and adaptability to changing fraud techniques.

Fast Design Space Exploration for Always-On Neural Networks

An analytical model can quickly predict performance and energy efficiency based on information about the neural network model and neural accelerator architecture, making it ideal for rapid pre-synthesis design space exploration. This paper proposes a new analytical model specifically targeted for convolutional neural networks used in always-on applications. To validate the proposed model, the performance and energy efficiency estimated by the model were compared with actual hardware and post-synthesis gate-level simulations of hardware synthesized with a state-of-the-art electronic design automation (EDA) synthesis tool. Comparisons with hardware created for the Eyeriss neural accelerator showed average execution time and energy consumption error rates of 3.33% and 13.54%, respectively. Comparisons with hardware synthesis results showed an error of 3.18% to 9.44% for two example neural accelerator configurations used to execute MobileNet, EfficientNet, and DarkNet neural network models. Finally, the utility of the proposed model was demonstrated by using it to evaluate the effects of different channel sizes, pruning rates, and batch sizes in several neural network designs for always-on vision, text, and audio processing.

A Novel Tool for Soybean Pods and Branches Recognition by Deep Learning Based Approach

ABSTRACTPhenotypic trait identification is crucial in cultivating new soybean varieties with high yield and quality. The traditional soybean phenotypic trait identification relies on manual pod counting and plant height measuring with a ruler. The heavy workload causes the data collected by human resources to be extremely prone to error. Therefore, developing an efficient and high‐quality method to obtain phenotypic data of soybean pods and branches is urgently needed. Three network models including ResNet‐101, Swin‐S and ConvNeXt‐S are compared in this study, and the ConvNeXt‐S model is identified as optimal, with a mAP@0.5 of 0.95, which could reach 74.2%. A deep learning–based approach soybean plant phenotype detection and data storage system is developed, including information on plants, pods and branches. The R2 between the number of pods detected by the system and the true value reached 0.995. These results indicate that the system is more accurate and stable than the manual phenotype identification. Our study paves the way for reducing economic and time costs as well as improving phenotype identification efficiency and accuracy in detecting soybean phenotypes.

Stochastic step-wise feature selection for Exponential Random Graph Models (ERGMs).

This study introduces a novel methodology for endogenous variable selection in Exponential Random Graph Models (ERGMs) to enhance the analysis of social networks across various scientific disciplines. Addressing critical challenges such as ERGM degeneracy and computational complexity, our method integrates a systematic step-wise feature selection process. This approach effectively manages the intractable normalizing constants characteristic of ERGMs, ensuring the generation of accurate and non-degenerate network models. An empirical application to nine real-life binary networks demonstrates the method's effectiveness in accommodating network dependencies and providing meaningful insights into complex network interactions. Particularly notable is the adaptability of this methodology to both directed and undirected networks, overcoming the limitations of traditional ERGMs in capturing realistic network structures. The findings contribute to network analysis, offering a robust framework for modeling and interpreting social networks and laying a foundation for future advancements in statistical network analysis techniques.

Optimizing quality of service forecasting in mobile networks through modified walrus optimization and multivariate approaches

This paper presents Ensemble-based Service Quality Prediction (EAQP), an automated method for predicting service quality under changing mobile network conditions. EAQP incorporates data preparation methods such as transformation, purification, & imputation, and then performs feature extraction utilizing statistical, geographical, as well as temporal approaches. An improved feature selection method, using a unique weighting approach and optimized by a modified Walrus Optimization Algorithm, improves the accuracy of predictions. EAQP utilizes a variety of prediction models such as support vector regression, recurrent neural network models, bi-directional short-term long-term memory networks, extreme learning machines, along with multi-layer perceptron neural networks to enhance predictive accuracy. EAQP uses complex optimization algorithms and ensemble learning approaches to provide precise and dependable predictions about service quality in real-time. This helps in proactive network management as well as improvement. This comprehensive approach shows potential for boosting network efficiency, optimizing the distribution of resources, and enhancing the end-user experience when using mobile communications systems.

Exploration of simultaneous transients between cerebral hemodynamics and the autonomic nervous system using windowed time-lagged cross-correlation matrices: a CENTER-TBI study

BackgroundTraumatic brain injury (TBI) can significantly disrupt autonomic nervous system (ANS) regulation, increasing the risk for secondary complications, hemodynamic instability, and adverse outcome. This retrospective study evaluated windowed time-lagged cross-correlation (WTLCC) matrices for describing cerebral hemodynamics–ANS interactions to predict outcome, enabling identifying high-risk patients who may benefit from enhanced monitoring to prevent complications.MethodsThe first experiment aimed to predict short-term outcome using WTLCC-based convolution neural network models on the Wroclaw University Hospital (WUH) database (Ptraining = 31 with 1,079 matrices, Pval = 16 with 573 matrices). The second experiment predicted long-term outcome, training on the CENTER-TBI database (Ptraining = 100 with 17,062 matrices) and validating on WUH (Pval = 47 with 6,220 matrices). Cerebral hemodynamics was characterized using intracranial pressure (ICP), cerebral perfusion pressure (CPP), pressure reactivity index (PRx), while ANS metrics included low-to-high-frequency heart rate variability (LF/HF) and baroreflex sensitivity (BRS) over 72 h. Short-term outcome at WUH was assessed using the Glasgow Outcome Scale (GOS) at discharge. Long-term outcome was evaluated at 3 months at WUH and 6 months at CENTER-TBI using GOS and GOS-Extended, respectively. The XGBoost model was used to compare performance of WTLCC-based model and averaged neuromonitoring parameters, adjusted for age, Glasgow Coma Scale, major extracranial injury, and pupil reactivity in outcome prediction.ResultsFor short-term outcome prediction, the best-performing WTLCC-based model used ICP-LF/HF matrices. It had an area under the curve (AUC) of 0.80, vs. 0.71 for averages of ANS and cerebral hemodynamics metrics, adjusted for clinical metadata. For long-term outcome prediction, the best-score WTLCC-based model used ICP-LF/HF matrices. It had an AUC of 0.63, vs. 0.66 for adjusted neuromonitoring parameters.ConclusionsAmong all neuromonitoring parameters, ICP and LF/HF signals were the most effective in generating the WTLCC matrices. WTLCC-based model outperformed adjusted neuromonitoring parameters in short-term but had moderate utility in long-term outcome prediction.

Machine Learning Application for High-Speed FTIR Absorption Spectra Analysis

"With this work our main objective is to find the best general (baseline) model for analyzing unknown spectra using Fourier infrared transformed spectroscopy (FT-IR) coupled with machine learning (ML) algorithms. This goal allows us to identify the best methodology applied for inline analysis of different experimental spectra for qualitative structural information obtained with types of structures that generate absorption or emission peaks. This methodology opens new perspectives for automated data processing using flexible algorithms and machine learning to encode experimental data for future applications. The results provide a good perspective on Machine Learning algorithms for applied sciences research. For our case study (FTIR experimental data) our model allows for peak feature extraction with a relative low, close to machine standard deviation, error budget. The best model identified is a specialized model but the standard, fully connected network models are evaluated."

Burning Mouth Syndrome Underlying Factors: A Roadmap From a Network Perspective.

To investigate the relationship between biological, psychological, and social factors underlying Burning Mouth Syndrome (BMS). A case (n = 40) and control (n = 42) study containing 80 variables was examined using two network models based on regularized partial correlations (n = 82). The structure of the associative pathways with the BMS was revealed. Direct associations involved Gastrointestinal Alterations (0.23), Vitamin D Deficiency (0.29), Musculoskeletal Alterations (0.29), Symptom Severity Score 2 (SSS2) (0.22), Cortisol Variation (0.10), Interpersonal Sensitivity (0.04), Hostility (0.03). Global Severity Index, Symptom Severity Score 1, Psychoticism, Obsession-Compulsion, Depression, Anxiety, and Somatization were indirectly related. The SSS2 was the most influential on BMS accuracy. Gastrointestinal alterations andvitamin D deficiency show a significant influence on BMS while cortisol mediates in multiple associative pathways between musculoskeletal alterations, gastrointestinal alterations, vitamin D deficiency, non-restorative sleep, fatigue, and cognitive problems. In addition to anxiety and depression, psychoticism, interpersonal sensitivity, and hostility stand out as psychological factors that seem to be related to a lack of vitamin D. None of the factors studied seem to have a relevant predictive potential for BMS, except for nonspecific symptoms of central sensitization.

The Effect of Pre- and Post-Processing Techniques on Tree Detection in Young Forest Stands from Images of Snow Cover Using YOLO Neural Networks

A neural network model for individual tree detection was developed based on the YOLOv4 architecture, which underwent additional preprocessing and postprocessing steps. The preprocessing step involved expanding the dataset by randomly cutting fragments from images, calculating anchor box sizes using the K-means clustering algorithm, and discarding anchor boxes that were too small a priori. The existing post-processing block of the YOLO architecture was modified by giving more weight to false positives in the error function and using the non-maximum suppression algorithm. Baseline neural networks from the YOLOv4 and YOLOv5 architectures, each in two versions (pre-trained and not pre-trained on the MS COCO dataset), were used for comparison without any additional modifications. In the overgrown experimental field, multi-season aerial copter surveys and ground counts were conducted on several sample plots to gather data. Comparison of multi-season aerial photographs with ground-count data showed that the best images in terms of the percentage of visually identifiable trees were those taken during the snowy season and when there was no foliage. Using these images and some additional images, we manually created a dataset on which we trained and tested neural network models. The model we developed showed significantly better results (2 to 10 times better) on the mAP 0.5 metric compared to the alternatives we considered.

Automatic Assessment of AK Stage Based on Dermatoscopic and HFUS Imaging—A Preliminary Study

Background: Actinic keratoses (AK) usually occur on sun-exposed areas in elderly patients with Fitzpatrick I–II skin types. Dermatoscopy and ultrasonography are two non-invasive tools helpful in examining clinically suspicious lesions. This study presents the usefulness of image-processing algorithms in AK staging based on dermatoscopic and ultrasonographic images. Methods: In 54 patients treated at the Department of Dermatology of Poznan University of Medical Sciences, clinical, dermatoscopic, and ultrasound examinations were performed. The clinico-dermoscopic AK classification was based on three-point Zalaudek scale. The ultrasound images were recorded with DermaScan C, Cortex Technology device, 20 MHz. The dataset consisted of 162 image pairs. The developed algorithm includes automated segmentation of ultrasound data utilizing a CFPNet-M model followed by handcrafted feature extraction. The dermatoscopic image analysis includes both handcrafted and convolutional neural network features, which, combined with ultrasound descriptors, are used in support vector machine-based classification. The network models were trained on public datasets. The influence of each modality on the final classification was evaluated. Results: The most promising results were obtained for the dermatoscopic analysis with the use of neural network model (accuracy 81%) and its combination with ultrasound scans (accuracy 79%). Conclusions: The application of machine learning-based algorithms in dermatoscopic and ultrasound image analysis machine learning in the staging of AKs may be beneficial in clinical practice in terms of predicting the risk of progression. Further experiments are warranted, as incorporating more images is likely to improve classification accuracy of the system.

Insights into food security: affordability perspectives from Bayesian belief network models

PurposeThis study investigates the dependencies between the Global Food Security Index (GFSI) and its affordability-related indicators using Bayesian belief network (BBN) models. The research also aims to prioritise these indicators within a probabilistic network setting.Design/methodology/approachThe research utilises BBN models to analyse data from 113 countries in 2022. Nine indicators related to food affordability, including income inequality, safety net programmes and trade freedom, are examined to understand their impact on food security. The methodology involves statistical modelling and analysis to identify critical factors influencing food security and to provide a comprehensive understanding of the global food affordability landscape.FindingsThe study reveals that income inequality, the presence and efficacy of safety net programmes and the degree of trade freedom are significant determinants of food affordability and overall food security outcomes. The analysis reveals marked disparities in performance across different countries, highlighting the need for context-specific interventions. The findings suggest that improving safety net programmes, implementing trade policy reforms and addressing income inequality are crucial for enhancing food affordability and security.Originality/valueThis research contributes to the literature by using BBN models to comprehensively analyse the relationship between the GFSI and affordability-related indicators. The study provides novel insights into how different socioeconomic factors influence food security across a diverse range of countries. The study offers actionable recommendations for policymakers to address food security challenges effectively, thereby supporting the development of more equitable and resilient food systems globally.

Efficient on-chip training of large-scale optical neural network through block adjoint training algorithm

MZI-based block optical neural networks (BONNs), which utilize block matrix multiplication to achieve large-scale network models, have garnered significant attention but still lack efficient training algorithms. In this article, by calculating the original field and adjoint field for the block matrices in BONNs and directly updating the phase values of all phase shifters within the optical mesh, we propose an on-chip block adjoint training (BAT) algorithm for large-scale BONNs. To demonstrate the effectiveness of our proposed algorithm, the trained BONNs are applied in image classification tasks for MNIST and SVHN datasets. The calculated results demonstrate that the performance of the BAT algorithm (95.915% for the MNIST dataset and 82.64% for the SVHN dataset) is competitive with the traditional gradient algorithm based on artificial neural networks (96.238% and 84.182%), but the BONNs can infer 1.5 times and 1.3 times faster than artificial neural networks, respectively. By studying the influence of the block size and the inputted position of the padded zero signals, we demonstrate that the BAT algorithm based on the BONNs with 12 block sizes can achieve higher performance by adding the padded zero signals to the same side beside the normal inputted signals. Additionally, we demonstrate that substituting the complete weight matrices with unitary matrices to construct BONNs is an efficient way to reduce both the system area and the required trainable parameters. Finally, we demonstrate the relatively good robustness of the BAT algorithm and the imprecision alleviation method by using on-chip retraining. Notably, our proposed BAT algorithm shows excellent potential for more complex tasks and network models.