Classification Algorithm Research Articles

The Development of Spinal Endoscopic Ultrasonic Imaging System with an Automated Tissue Recognition Algorithm.

Preclinical experimental study. To develop an intraoperative ultrasound-assisted imaging device, which could be placed at the surgical site through an endoscopic working channel and which could help surgeon recognition of different tissue types during endoscopic spinal surgery (ESS). ESS remains a challenging task for spinal surgeons. Great proficiency and experience are needed to perform procedures such as intervertebral discectomy and neural decompression within a narrow channel. The limited surgical view poses a risk of damaging important structures, such as nerve roots. We constructed a spinal endoscopic ultrasound system, using a 4-mm custom ultrasound probe, which can be easily inserted through the ESS working channel, allowing up to 10 mm depth detection. This system was applied to ovine lumbar spine samples to obtain ultrasound images. Subsequently, we proposed a two-stage classification algorithm, based on a pretrained DenseNet architecture for automated tissue recognition. The recognition algorithm was evaluated using accuracy and consistency. The probe can be easily used in the ESS working channel and produce clear and characteristic ultrasound images. We collected 367 images for training and testing of the recognition algorithm, including images of the spinal cord, nucleus pulposus, adipose tissue, bone, annulus fibrosus and nerve roots. The algorithm achieved over 90% accuracy in recognizing all types of tissues with a Kappa value of 0.875. The recognition times were under 0.1 s using the current configuration. Our system was able to be used in existing ESS working channels and clearly identified at-risk spinal structures in vitro. The pretrained algorithms could identify six intraspinal tissue types accurately and quickly. The concept and innovative application of intraoperative ultrasound in ESS may shorten the learning curve of ESS and improve surgical efficiency and safety.

Retraction: Cost-sensitive classification algorithm combining the Bayesian algorithm and quantum decision tree

Retraction: Cost-sensitive classification algorithm combining the Bayesian algorithm and quantum decision tree

Securing IoT network with hybrid evolutionary lion intrusion detection system: a composite motion optimisation algorithm for feature selection and ensemble classification

ABSTRACT Networks connected to the Internet of Things (IoT) are often vulnerable to attacks. Several existing methods in the intrusion detection system for securing IoT have been presented with ensemble classifier, but it does not accurately classify attack, and also it takes high computation time. With intention of solving the security issues, Intrusion Detection System using Hybrid Evolutionary Lion and Balancing Composite Motion Optimisation Algorithm espoused feature selection with Ensemble Classifier (IDS-IoT-Hybrid ELOA-BCMOA-Ensemble-DT-LSVM-RF-XGBoost) is proposed for Securing IoT Network. At first, data were accumulated from the NSL-KDD data set. Afterward, data is fed to pre-processing, where it restored missing value using mean curvature flow method. At feature selection, optimum features are compiled under Hybrid Evolutionary Lion and Balancing Composite Motion Optimisation Algorithm. Based upon the optimum features, intruders of IoT data are categorised as denial-of-service (DoS), probe, remote to local attack (R2L), user to root attack (U2R), normal (no attack) with the help of Ensemble classifier. Proposed IDS-IoT-Hybrid ELOA-BCMOA-Ensemble-DT-LSVM-RF-XGBoost approach is constructed utilising Python. Then, proposed IDS-IoT-Hybrid ELOA-BCMOA-Ensemble-DT-LSVM-RF-XGBoost approach attains 21.11%, 19.58%, 24.61% and 9.52% higher accuracy; 94.47%, 93.95%, 93.08% and 90.59% lower error rate, and 62.94%, 36.69%, 64.17% and 50.97% less computation time analysed with existing models.

A framework for modelling customer invoice payment predictions

By offering clients attractive credit terms on sales, a company may increase its turnover, but granting credit also incurs the cost of money tied up in accounts receivable (AR), increased administration and a heightened probability of incurring bad debt. The management of credit sales, although eminently important to any business, is often performed manually, which may be time-consuming, expensive and inaccurate. Such an administrative workload becomes increasingly cumbersome as the number of credit sales increases. As a result, a new approach towards proactively identifying invoices from AR accounts that are likely to be paid late, or not at all, has recently been proposed in the literature, with the aim of employing intervention strategies more effectively. Several computational techniques from the credit scoring literature and particularly techniques from the realms of survival analysis or machine learning have been embedded in the aforementioned approach. This body of work is, however, lacking due to the limited guidance provided during the data preparation phase of the model development process and because survival analytic and machine learning techniques have not yet been ensembled. In this paper, we propose a generic framework for modelling invoice payment predictions with the aim of facilitating the process of preparing transaction data for analysis, generating relevant features from past customer behaviours, and selecting and ensembling suitable models for predicting the time to payment associated with invoices. We also introduce a new sequential ensembling approach, called the Survival Boost algorithm. The rationale behind this method is that features generated by a survival analytic model can enhance the efficacy of a machine learning classification algorithm.

Enhancing identification performance of cognitive impairment high-risk based on a semi-supervised learning method

BackgroundCognitive assessment plays a pivotal role in the early detection of cognitive impairment, particularly in the prevention and management of cognitive diseases such as Alzheimer’s and Lewy body dementia. Large-scale screening relies heavily on cognitive assessment scales as primary tools, with some low sensitivity and others expensive. Despite significant progress in machine learning for cognitive function assessment, its application in this particular screening domain remains underexplored, often requiring labor-intensive expert annotations. AimsThis paper introduces a semi-supervised learning algorithm based on pseudo-label with putback (SS-PP), aiming to enhance model efficiency in predicting the high risk of cognitive impairment (HR-CI) by utilizing the distribution of unlabeled samples. DataThe study involved 189 labeled samples and 215,078 unlabeled samples from real world. A semi-supervised classification algorithm was designed and evaluated by comparison with supervised methods composed by 14 traditional machine-learning methods and other advanced semi-supervised algorithms. ResultsThe optimal SS-PP model, based on GBDT, achieved an AUC of 0.947. Comparative analysis with supervised learning models and semi-supervised methods demonstrated an average AUC improvement of 8% and state-of-art performance, repectively. ConclusionThis study pioneers the exploration of utilizing limited labeled data for HR-CI predictions and evaluates the benefits of incorporating physical examination data, holding significant implications for the development of cost-effective strategies in relevant healthcare domains.

Anomaly detection in the veterinary antibiotic prescription surveillance system (IS ABV)

Antibiotic resistance is one of the major concerns in veterinary and human medicine and poses a considerable threat to both human and animal health. It has been shown that over- or misuse of antibiotics is one of the primary drivers of antibiotic resistance. To develop the surveillance of antibiotic use, Switzerland introduced the "Informationssystem Antibiotika in der Veterinärmedizin" (IS ABV) in 2019, mandating electronic registration of antibiotic prescriptions by all veterinarians in Switzerland. However, initial data analysis revealed a considerable amount of implausible data entries, potentially compromising data quality and reliability. These anomalies may be caused by input errors, inaccuracies, incorrect or aberrant master data or data transmission and make analysis impossible. To address this issue efficiently, we propose a two-stage anomaly detection framework utilizing machine learning algorithms. In this study, our primary focus was on cattle treatments with either single or group therapy, as they were the species with the highest prescription volume. However, not all outliers are necessarily incorrect; some may be legitimate but unusual antibiotic treatments. Thus, expert review plays a crucial role in distinguishing outliers, that are correct from actual errors. Initially, relevant prescription variables were extracted and pre-processed with a custom-built scaler. A set of unsupervised algorithms calculated the probability of each data point and identified the most likely outliers. In collaboration with experts, we annotated anomalies and established anomaly thresholds for each production type and active substance. These expert-annotated labels were then used to fine-tune the final supervised classification algorithms. With this methodology, we identified 22,816 anomalies from a total of 1,994,170 prescriptions in cattle (1.1 %). Cattle with no further specified production type had the most (2 %) anomalies with 7758 out of 379,995. The anomalies were consistently identified and comprised prescriptions with too high and too low dosages. Random Forest achieved a ROC-AUC score of 0.994, (95 % CI: 0.992, 0.995) and a F1-Score of 0.962 (95 % CI: 0.958, 0.966) for single treatments. The versatility of this framework allows its adaptation to other species within IS ABV and potentially to other prescription-based surveillance systems. If applied regularly to uploaded prescriptions, it should reduce input errors over time, improving the validity of the data in the long term.

Using Continuous Glucose Monitoring to Passively Classify Naturalistic Binge Eating and Vomiting Among AdultsWith Binge-Spectrum Eating Disorders: APreliminary Investigation.

Binge eating and self-induced vomiting are common, transdiagnostic eating disorder (ED) symptoms. Efforts to understand these behaviors in research and clinical settings have historically relied on self-report measures, which may be biased and have limited ecological validity. It may be possible to passively detect binge eating and vomiting using data collected by continuous glucose monitors (CGMs; minimally invasive sensors that measure blood glucose levels), as these behaviors yield characteristic glucose responses. This study developed machine learning classification algorithms to classify binge eating and vomiting among 22 adults with binge-spectrum EDs using CGM data. Participants wore Dexcom G6 CGMs and reported eating episodes and disordered eating symptoms using ecological momentary assessment for 2 weeks. Group-level random forest models were generated to distinguish binge eating from typical eating episodes and to classify instances of vomiting. The binge eating model had accuracy of 0.88 (95% CI: 0.83, 0.92), sensitivity of 0.56, and specificity of 0.90. The vomiting model demonstrated accuracy of 0.79 (95% CI: 0.62, 0.91), sensitivity of 0.88, and specificity of 0.71. Results suggest that CGM may be a promising avenue for passively classifying binge eating and vomiting, with implications for innovative research and clinical applications.

Empowering the rapid authentication of the botanical origin of monofloral honey by coated blade spray mass spectrometry (CBS-MS)

In the quest for a rapid and cost-effective tool for determining the botanical origin of monofloral honey, the analytical capabilities of coated blade spray mass spectrometry (CBS-MS) were investigated. To this aim, the chemical profiles of 64 honey samples from seven different botanical origins (acacia, dandelion, chestnut, rhododendron, citrus, sunflower, and linden) were captured by the absorbent of the coated blades and then analyzed by mass spectrometry. An exploratory analysis was performed by principal component analysis (PCA) to generate a graphical representation of the CBS-MS data that allows the discovery of patterns, outliers, or relations between types of honey in an unsupervised manner. Additionally, the performances of four different classification algorithms (least absolute shrinkage and selection operator (LASSO), random forest (RF), and neural network (NNET) partial least squares discriminant analysis (PLS-DA)) were built up and compared. The performances of the four classifiers were verified by a 50 times-repeated, 5-fold-cross-validation and permutation test. Although all classifiers performed well, the RF showed significantly higher performances in cross-validation (with area under the curve (AUC) of 0.99, overall accuracy 0.94, Kappa 0.93, sensitivity 0.94, and specificity 0.99). Moreover, the permutation tests showed the models were not overfitted. Finally, to determine the molecular identities of the ions that most contribute to the classification, extracts from the same honey samples were prepared, analyzed by liquid chromatography coupled to high resolution tandem mass spectrometry (LC-HRMS/MS), and the most significant features were annotated. This proof-of-principle work warrants a future large-scale study to validate and challenge this CBS-MS-based method with a greater number of honeys from different years and geographical origins.

Machine Learning Analysis in Diffusion Kurtosis Imaging for Discriminating Pediatric Posterior Fossa Tumors: A Repeatability and Accuracy Pilot Study.

Background and purpose: Differentiating pediatric posterior fossa (PF) tumors such as medulloblastoma (MB), ependymoma (EP), and pilocytic astrocytoma (PA) remains relevant, because of important treatment and prognostic implications. Diffusion kurtosis imaging (DKI) has not yet been investigated for discrimination of pediatric PF tumors. Estimating diffusion values from whole-tumor-based (VOI) segmentations may improve diffusion measurement repeatability compared to conventional region-of-interest (ROI) approaches. Our purpose was to compare repeatability between ROI and VOI DKI-derived diffusion measurements and assess DKI accuracy in discriminating among pediatric PF tumors. Materials and methods: We retrospectively analyzed 34 children (M, F, mean age 7.48 years) with PF tumors who underwent preoperative examination on a 3 Tesla magnet, including DKI. For each patient, two neuroradiologists independently segmented the whole solid tumor, the ROI of the area of maximum tumor diameter, and a small 5 mm ROI. The automated analysis pipeline included inter-observer variability, statistical, and machine learning (ML) analyses. We evaluated inter-observer variability with coefficient of variation (COV) and Bland-Altman plots. We estimated DKI metrics accuracy in discriminating among tumor histology with MANOVA analysis. In order to account for class imbalances, we applied SMOTE to balance the dataset. Finally, we performed a Random Forest (RF) machine learning classification analysis based on all DKI metrics from the SMOTE dataset by partitioning 70/30 the training and testing cohort. Results: Tumor histology included medulloblastoma (15), pilocytic astrocytoma (14), and ependymoma (5). VOI-based measurements presented lower variability than ROI-based measurements across all DKI metrics and were used for the analysis. DKI-derived metrics could accurately discriminate between tumor subtypes (Pillai's trace: p < 0.001). SMOTE generated 11 synthetic observations (10 EP and 1 PA), resulting in a balanced dataset with 45 instances (34 original and 11 synthetic). ML analysis yielded an accuracy of 0.928, which correctly predicted all but one lesion in the testing set. Conclusions: VOI-based measurements presented improved repeatability compared to ROI-based measurements across all diffusion metrics. An ML classification algorithm resulted accurate in discriminating PF tumors on a SMOTE-generated dataset. ML techniques based on DKI-derived metrics are useful for the discrimination of pediatric PF tumors.

Thigh-worn accelerometry: a comparative study of two no-code classification methods for identifying physical activity types

BackgroundThe more accurate we can assess human physical behaviour in free-living conditions the better we can understand its relationship with health and wellbeing. Thigh-worn accelerometry can be used to identify basic activity types as well as different postures with high accuracy. User-friendly software without the need for specialized programming may support the adoption of this method. This study aims to evaluate the classification accuracy of two novel no-code classification methods, namely SENS motion and ActiPASS.MethodsA sample of 38 healthy adults (30.8 ± 9.6 years; 53% female) wore the SENS motion accelerometer (12.5 Hz; ±4 g) on their thigh during various physical activities. Participants completed standardized activities with varying intensities in the laboratory. Activities included walking, running, cycling, sitting, standing, and lying down. Subsequently, participants performed unrestricted free-living activities outside of the laboratory while being video-recorded with a chest-mounted camera. Videos were annotated using a predefined labelling scheme and annotations served as a reference for the free-living condition. Classification output from the SENS motion software and ActiPASS software was compared to reference labels.ResultsA total of 63.6 h of activity data were analysed. We observed a high level of agreement between the two classification algorithms and their respective references in both conditions. In the free-living condition, Cohen’s kappa coefficients were 0.86 for SENS and 0.92 for ActiPASS. The mean balanced accuracy ranged from 0.81 (cycling) to 0.99 (running) for SENS and from 0.92 (walking) to 0.99 (sedentary) for ActiPASS across all activity types.ConclusionsThe study shows that two available no-code classification methods can be used to accurately identify basic physical activity types and postures. Our results highlight the accuracy of both methods based on relatively low sampling frequency data. The classification methods showed differences in performance, with lower sensitivity observed in free-living cycling (SENS) and slow treadmill walking (ActiPASS). Both methods use different sets of activity classes with varying definitions, which may explain the observed differences. Our results support the use of the SENS motion system and both no-code classification methods.

Using Machine Learning for Endoscopic Detection of Low-Grade Subglottic Stenosis: A Proof of Principle.

The current study trains, tests, and evaluates a deep learning algorithm to detect subglottic stenosis (SGS) on endoscopy. A retrospective review of patients undergoing microlaryngoscopy-bronchoscopy was performed. A pretrained image classifier (Resnet50) was retrained and tested on 159 images of airways taken at the glottis, 106 normal-sized airways, and 122 with SGS. Data augmentation was performed given the small sample size to prevent overfitting. Overall model accuracy was 73.3% (SD: 3.8). Precision and recall for stenosis were 77.3% (SD: 4.0) and 72.7 (SD: 4.0). F1 score for the detection of stenosis was 0.75 (SD: 0.04). Precision and recall for normal-sized images were lower at 69% (SD: 4.35) and 74% (SD: 4), with an F1 score of 0.71 (SD: 0.04). This study demonstrates that an image classification algorithm can identify SGS on endoscopic images. Work is needed to improve diagnostic accuracy for eventual deployment of the algorithm into clinical care.

Feature value quantization and reduction process for predicting heart attack possibility and the level of severity by a machine learning model

Heart disease is a prevalent condition nowadays that, if undiagnosed, can be deadly. To predict heart disease, researchers designed many machine learning models. In this study, we propose a model that chooses fewer attribute columns for training, and we use these chosen features to determine the heart problem severity. Correlation Repeated Heat map and Information Gain were used for selecting the features. To train our model we used the UCI Cleveland heart disease dataset. We removed duplicate data to improve the accuracy score, and we also encoded the categorical data collection using the OneHot(OH) encoding method, which can improve prediction accuracy. Support Vector, Logistic Regression, K-Nearest Neighbour, Naive Bayes, Decision Tree, Random Forest, Adaboost, and XGBoost are the eight classifier algorithms that are used in this process overall. Based on repeated heat map correlation, we compare the accuracy score each time. In this proposed method, the Adaboost classification algorithm used by the fbs row heat map achieves the highest accuracy for heart disease detection and it is 92%. By choosing features according to the information gain value, we compare the accuracy score each time in information gain. For both XGBoost and Logistic Regression, we got an accuracy score of 93.44%. However, compared to the XGBoost classification technique, Logistic Regression requires less time. Accuracy, precision, recall, f1-score, sensitivity, specificity, and the AUC of ROC charts were used to evaluate the performance of the model. Overall, the results of our model demonstrate that it is reliable and accurate in identifying cardiac disease and its level of severeness.

The classification algorithms to support the management of the patient with femur fracture

Effectiveness in health care is a specific characteristic of each intervention and outcome evaluated. Especially with regard to surgical interventions, organization, structure and processes play a key role in determining this parameter. In addition, health care services by definition operate in a context of limited resources, so rationalization of service organization becomes the primary goal for health care management. This aspect becomes even more relevant for those surgical services for which there are high volumes. Therefore, in order to support and optimize the management of patients undergoing surgical procedures, the data analysis could play a significant role. To this end, in this study used different classification algorithms for characterizing the process of patients undergoing surgery for a femoral neck fracture. The models showed significant accuracy with values of 81%, and parameters such as Anaemia and Gender proved to be determined risk factors for the patient’s length of stay. The predictive power of the implemented model is assessed and discussed in view of its capability to support the management and optimisation of the hospitalisation process for femoral neck fracture, and is compared with different model in order to identify the most promising algorithms. In the end, the support of artificial intelligence algorithms laying the basis for building more accurate decision-support tools for healthcare practitioners.

A comparative study of heterogeneous machine learning algorithms for arrhythmia classification using feature selection technique and multi-dimensional datasets

Arrhythmia, a common cardiovascular disorder, refers to the abnormal electrical activity within the heart, leading to irregular heart rhythms. This condition affects millions of people worldwide, with severe implications on cardiac function and overall health. Arrhythmias can strike anyone at any age which is a significant cause of morbidity and mortality on a global scale. About 80% of deaths related to heart disease are caused by ventricular arrhythmias. This research investigated the application of an optimized multi-objectives supervised Machine Learning (ML) models for early arrhythmia diagnosis. The authors evaluated the model’s performance on the arrhythmia dataset from the UCI ML repository with varying train-test splits (70:30, 80:20, and 90:10). Standard preprocessing techniques such as handling missing values, formatting, balancing, and directory analysis were applied along with Pearson correlation for feature selection, all aimed at enhancing model performance. The proposed optimized RF model achieved impressive performance metrics, including accuracy (95.24%), precision (100%), sensitivity (89.47%), and specificity (100%). Furthermore, the study compared the proposed approach to existing models, demonstrating significant improvements across various performance measures.

An online ensemble classification algorithm for multi-class imbalanced data stream

An online ensemble classification algorithm for multi-class imbalanced data stream

An improved random forest model for music genre classification algorithm based on sparrow search algorithm

In this study, a random forest model optimised based on the sparrow search algorithm is explored for the music genre classification problem. By introducing the sparrow search algorithm, the model is able to search the feature space more comprehensively during the training process, which improves the learning ability of complex data relationships. After 15 iterations of optimisation, the fitness value of the model is improved from 0.2002 to 0.197, showing better convergence. The performance on the training set shows that the model predicts 2,543 music genres with an accuracy of 99.59% and only 10 incorrect predictions, which verifies its excellent performance on the training data. While on the test set, the model's prediction accuracy for 623 music genres is 79.78%, which is slightly lower than the result of the training set, but still reflects good generalisation ability. The results of this paper imply that the improved random forest model based on the sparrow search algorithm has potential application prospects in the music genre classification task. By optimising the parameter settings and feature selection, we not only improve the prediction accuracy and generalisation ability of the model, but also provide new ideas for solving other complex data classification problems. In the future, we can further expand the sample size, try different feature combinations, and conduct in-depth research by combining more domain knowledge in order to further improve the model performance and promote the development of related fields.

Machine and Deep Learning Trends in EEG-Based Detection and Diagnosis of Alzheimer’s Disease: A Systematic Review

This article presents a systematic review using PRISMA methodology to explore trends in the use of machine and deep learning in diagnosing and detecting Alzheimer’s disease using electroencephalography. This review covers studies published between 2013 and 2023, drawing on three leading academic databases: Scopus, Web of Science, and PubMed. The validity of the databases is evaluated considering essential factors such as the arrangement of EEG electrodes, data acquisition methodologies, and the number of participants. Additionally, the specific properties of the databases used in the research are highlighted, including EEG signal classification, filtering, segmentation approaches, and selected features. Finally, the performance metrics of the classification algorithms are evaluated, especially the accuracy achieved, offering a comprehensive view of the current state and future trends in the use of these technologies for the diagnosis of Alzheimer’s disease.

Near-Term Quantum Classification Algorithms Applied to Antimalarial Drug Discovery.

Computational approaches are widely applied in drug discovery to explore properties related to bioactivity, physiochemistry, and toxicology. Over at least the last 20 years, the exploitation of machine learning on molecular data sets has been used to understand the structure-activity relationships that exist between biomolecules and druggable targets. More recently, these methods have also seen application for phenotypic screening data for neglected diseases such as tuberculosis and malaria. Herein, we apply machine learning to build quantum Quantitative Structure Activity Relationship models from antimalarial data sets. There is a continual need for new antimalarials to address drug resistance, and the readily available in vitro data sets could be utilized with newer machine learning approaches as these develop. Furthermore, quantum machine learning is a relatively new method that uses a quantum computer to perform the calculations. First, we present a classical-quantum hybrid computational approach by building a Latent Bernoulli Autoencoder machine learning model for compressing bit-vector descriptors to a size that can be adapted to quantum computers for classification tasks with limited loss of embedded information. Second, we apply our method for feature map compression to quantum classification algorithms, including a completely novel machine learning algorithm with no analogy in classical computers: the Quantum Fourier Transform Classifier. We apply both these approaches to build quantum machine learning models for small-molecule antimalarials with quantum simulation software and then benchmark these quantum models against classical machine learning approaches. While there are many challenges currently facing the development of reliable quantum computers, our results demonstrate that there is potential for the use of this technology in the field of drug discovery.

Spectral sensor‐based device for real‐time detection and severity estimation of groundnut bud necrosis virus in tomato

AbstractA machine learning‐based approach was utilized to develop a device for groundnut bud necrosis virus (GBNV) disease severity detection and estimation in tomato plants (Solanum lycopersicum L.). The study involved inoculating tomato plants with GBNV, monitoring changes in morphological and spectral characteristics, evaluating machine learning algorithms (decision tree [DT] classifier) for analysis and classification of disease severity, and developing and validating a device for disease detection and severity estimation. Spectral data analysis revealed distinct patterns in reflectance, with notable peaks observed in the 680 and 760 nm bands, while reflectance remained low and constant beyond 900 nm. Machine learning techniques, specifically a DT model, were employed to classify disease severity based on spectral data with high accuracy (95.01% training accuracy and 93.65% testing accuracy). The model identified the near‐infrared band as highly correlated (correlation coefficient of 0.82) with disease severity. Furthermore, a compact handheld device integrating a spectral sensor, organic light‐emitting diode display, and Raspberry Pi 3B was developed for real‐time disease severity estimation. The device demonstrated robust performance, accurately predicting disease severity at different growth stages, even in the absence of visible symptoms. Additionally, disease severity percentages obtained via reverse transcription polymerase chain reaction were used to validate the accuracy of the device's estimations. Its responsive nature, with estimated response times ranging from milliseconds to seconds, facilitates timely interventions in agricultural settings. Overall, this interdisciplinary approach, combining spectral analysis, machine learning, and device development, presents a promising solution for efficient disease monitoring and management in agriculture, contributing to enhanced crop health and food security.

Optimizing Convolutional Neural Networks for Image Classification on Resource-Constrained Microcontroller Units

Running machine learning algorithms for image classification locally on small, cheap, and low-power microcontroller units (MCUs) has advantages in terms of bandwidth, inference time, energy, reliability, and privacy for different applications. Therefore, TinyML focuses on deploying neural networks on MCUs with random access memory sizes between 2 KB and 512 KB and read-only memory storage capacities between 32 KB and 2 MB. Models designed for high-end devices are usually ported to MCUs using model scaling factors provided by the model architecture’s designers. However, our analysis shows that this naive approach of substantially scaling down convolutional neural networks (CNNs) for image classification using such default scaling factors results in suboptimal performance. Consequently, in this paper we present a systematic strategy for efficiently scaling down CNN model architectures to run on MCUs. Moreover, we present our CNN Analyzer, a dashboard-based tool for determining optimal CNN model architecture scaling factors for the downscaling strategy by gaining layer-wise insights into the model architecture scaling factors that drive model size, peak memory, and inference time. Using our strategy, we were able to introduce additional new model architecture scaling factors for MobileNet v1, MobileNet v2, MobileNet v3, and ShuffleNet v2 and to optimize these model architectures. Our best model variation outperforms the MobileNet v1 version provided in the MLPerf Tiny Benchmark on the Visual Wake Words image classification task, reducing the model size by 20.5% while increasing the accuracy by 4.0%.