Publications (8)
Automated detection of paddy (Oryza sativa) leaf diseases can support timely field intervention, but public datasets for this task are not always shipped in a directly usable form. This paper documents a reproducible object-detection pipeline built on a 480-image, 6-class Roboflow paddy-disease dataset whose raw export placed all annotated pairs under a single partition and provided only placeholder numeric class labels. We first validate every annotation, then construct a seeded (42), dominant-class-approximate-stratified 70/15/15 train/validation/test split with a post-split integrity check, and fine-tune a COCO-pretrained YOLO11m detector (80 epochs, patience 15, image size 640, AdamW) on the resulting train/validation partitions. Evaluated once on the held-out test split (72 images, 257 instances), the model reaches precision 0.570, recall 0.597, mAP50 0.585, and mAP50-95 0.269, at 46.65 ms/image on a Tesla T4 GPU. Per-class analysis shows AP50 ranging from 0.342 to 0.878 across the six classes, with the numerically dominant class exhibiting the lowest AP50 despite having the most training examples, an effect traced to a high rate of missed (background-classified) detections in the confusion matrix. We report this as a documented, reproducible baseline rather than a state-of-the-art claim, given open questions about class semantics and possible image-level leakage that the available metadata cannot rule out.
This study investigates the associations between weekly generative artificial intelligence (GenAI) tool usage and three student-level academic outcomes — post-semester GPA, skill retention score, and burnout risk level — using a large cross-sectional observational dataset of N = 50,000 undergraduate and graduate students spanning five academic disciplines. Employing covariate-adjusted ordinary least squares regression (ANCOVA), ordered logistic regression, quadratic dose-response modelling, bootstrap mediation analysis, and inverse probability weighting (IPW) with doubly robust correction, we examine whether GenAI tool adoption improves or undermines academic outcomes after controlling for pre-semester GPA, traditional study hours, institutional AI policy, and individual skill characteristics. Bivariate correlation between weekly GenAI hours and post-semester GPA is small and negative (r = −0.019), yet attenuates substantially after adjustment for pre-semester GPA — the dominant predictor (r = 0.927). Perceived AI dependency and examination anxiety partially mediate the GenAI–GPA pathway. Prompt engineering skill level and institutional policy moderate the relationship, with advanced-skill users under permissive policies demonstrating attenuated adverse effects. Non-linear modelling reveals an inverted-U pattern with diminishing or adverse returns beyond moderate weekly use thresholds. High-intensity GenAI users report meaningfully elevated exam anxiety (mean 4.99 vs. 3.82 for low users) and 25% higher perceived AI dependency scores. Debugging and troubleshooting use-cases are associated with the strongest GPA gains (∆ = +0.249), while direct-answer generation shows the weakest benefit (∆ = +0.133). These findings carry direct implications for institutional AI governance, curriculum re-design, and student wellbeing programming.
Artificial intelligence (AI) is restructuring the global labor market through simultaneous displacement of routine-cognitive tasks and the creation of high-skill occupational demand. This study employs a cross-sectional dataset of 3,000 occupational records spanning 10 countries, 10 industries, and 20 distinct job titles to investigate the structural determinants of AI replacement risk and near-term hiring trajectories. Using a combination of descriptive statistics, Spearman rank correlation, one-way ANOVA, Mann-Whitney U, Kruskal-Wallis tests, Ridge regression, and Random Forest modeling, we find that the observed dataset exhibits a notably uniform distribution of AI replacement risk (x̄ = 0.503, σ = 0.261), with no statistically significant differences across automation levels, education strata, upskilling engagement, or AI tool usage in the as-generated synthetic cross-section. Future demand score is near-orthogonal to replacement risk (ρ = 0.008, p = 0.677), suggesting the dataset reflects a balanced design rather than a population drawn from actual occupational risk distributions. Predictive models (Random Forest) are expected to achieve high held-out R² on this dataset given its tabular structure, and multinomial classifiers achieve balanced accuracy for three-class hiring trend prediction. These results are interpreted within a theoretical framework of labor-market polarization, and we draw direct implications for workforce policy, organizational upskilling investment, and occupational re-design strategies. We also document where the synthetic dataset departs from empirical expectations and identify avenues for replication with real-world data.
Pneumonia remains a leading cause of mortality worldwide, particularly affecting children and the elderly. Early and accurate diagnosis is critical for effective treatment and improved patient outcomes. Chest X-ray (CXR) imaging is the most common diagnostic tool, but its interpretation can be subjective and time-consuming, leading to inter-observer variability. This paper presents a deep learning-based approach for automated pneumonia detection from chest X-ray images. We employ a pre-trained ResNet-18 architecture fine-tuned on a publicly available dataset of 5,856 chest X-ray images (NORMAL and PNEUMONIA classes). To address the inherent class imbalance in medical imaging datasets, we implement a weighted loss function and a weighted random sampler during training. Our data augmentation pipeline incorporates random cropping, horizontal flipping, rotation, and color jittering to enhance model generalization. The model achieves a training accuracy of 94.27% and a validation accuracy of 93.30% after a single epoch, with a validation loss of 0.2519. The final model, after applying early stopping and model checkpointing, attains a validation accuracy of 95.11% and a validation loss of 0.2354. Comprehensive eval uation metrics, including precision, recall, F1-score, and ROC AUC, demonstrate the model’s robustness and clinical potential. This research highlights the efficacy of transfer learning and imbalance mitigation techniques in developing reliable computer aided diagnosis systems for pneumonia detection.
The rapid expansion of artificial intelligence as a professional discipline has created an urgent need for data-driven career guidance systems capable of matching practitioners to specialization domains based on measurable skill and experience profiles. This study frames AI domain recommendation as a supervised ten-class classification problem and evaluates five state-of-the-art classifiers — Logistic Regression, Random Forest, XGBoost, LightGBM, and CatBoost — on a dataset of 30,000 AI practitioner profiles across 30 features spanning technical skill scores, experience metrics, productivity indicators, and tooling preferences. Rigorous experimental controls include explicit exclusion of target-leaking domain labels, stratified 60/20/20 train-validation-test splitting, median imputation for three low-missingness columns, and 5-fold stratified cross-validation. LightGBM achieves the highest performance with test accuracy of 83.67%, macro-F1 of 0.7050, weighted F1 of 0.8250, and macro ROC-AUC of 0.9352. SHAP-based explainability identifies largest model parameters million, skill consistency score, and mathematics for ai score as the dominant global predictors. Severe class imbalance between Generative AI (21.4%) and Reinforcement Learning (1.4%) drives differential per-class performance, with Finance AI (F1 = 0.129) and Healthcare AI (F1 = 0.198) remaining severely under-predicted despite class-weighted training. This work establishes the first replicable, leakage-free methodology and performance baseline for AI specialization domain recommendation.
This study investigates the associations between weekly generative artificial intelligence (GenAI) tool usage and three student-level academic outcomes — post-semester GPA,skill retention score, and burnout risk level — using a large cross sectional observational dataset of N = 50,000 undergraduate and graduate students spanning five academic disciplines. Employing covariate-adjusted ordinary least squares regression (ANCOVA), ordered logistic regression, quadratic dose-response modelling, bootstrap mediation analysis, and inverse probability weighting (IPW) with doubly robust correction, we examine whether GenAI tool adoption improves or undermines academic outcomes after controlling for pre-semester GPA, traditional study hours, institutional AI policy, and individual skill characteristics. Bivariate correlation between weekly GenAI hours and post-semester GPA is small and negative (r = −0.019), yet attenuates substantially
after adjustment for pre-semester GPA — the dominant predictor
(r = 0.927). Perceived AI dependency and examination anxiety
partially mediate the GenAI–GPA pathway. Prompt engineering
skill level and institutional policy moderate the relationship,
with advanced-skill users under permissive policies demonstrat
ing attenuated adverse effects. Non-linear modelling reveals an
inverted-U pattern with diminishing or adverse returns beyond
moderate weekly use thresholds. High-intensity GenAI users
report meaningfully elevated exam anxiety (mean 4.99 vs. 3.82
for low users) and 25% higher perceived AI dependency scores.
Debugging and troubleshooting use-cases are associated with
the strongest GPA gains (∆ = +0.249), while direct-answer
generation shows the weakest benefit (∆ = +0.133). These
findings carry direct implications for institutional AI governance,
curriculum re-design, and student wellbeing programming.
This study presents a comprehensive machine learning analysis of 52 humanoid robotic platforms catalogued in the Global Humanoid Robotics Intelligence Dataset (2026), a curated cross-sectional record of the global humanoid robotics industry comprising 68 features spanning physical design, sensory architecture, AI integration, commercial maturity, and geopolitical context. We formulate three complementary analytical tasks: (1) multi-class classification of AI Maturity Score (ordinal 0–5), (2) regression of Commercial Readiness Index (continuous 0–100), and (3) unsupervised clustering to identify latent robot archetypes. Eight classifiers are benchmarked via 5-fold stratified cross-validation and Leave-One-Out (LOO) validation, appropriate for the small-N regime (N = 52). Gradient Boosting achieves the highest 5-fold weighted F1 of 0.930±0.086 and a test set accuracy of 90.9%; the tuned Random Forest achieves LOO accuracy of 86.5% and LOO weighted F1 of 84.8%. For commercial readiness regression, Random Forest Regressor achieves the lowest cross-validated RMSE of 13.20 ± 3.99 with R² = 0.615. KMeans clustering on PCA-reduced features identifies six latent robot archetypes (silhouette-optimal k = 6), capturing the axis from early-stage research prototypes to commercially deployed leaders. Statistical tests confirm that geopolitical tier significantly influences AI maturity (H = 8.371, p = 0.039) and that LLM integration is associated with significantly higher commercial readiness (U = 438.0, p = 0.025). These findings provide actionable intelligence for stakeholders assessing the global humanoid robotics landscape in 2026.
The proliferation of artificial intelligence (AI) tools across professional, educational, and social domains has prompted urgent scholarly inquiry into their long-term effects on human cognitive and psychological development. This study investigates whether measurable dimensions of AI engagement — including usage intensity, dependency score, tool proficiency, and scenario context — combined with individual cognitive and emotional attributes, can predict discrete mental growth trajectories across a global population spanning five decades (2024–2075). Using a dataset of 52,000 records across nine regions, five demographic cohorts, and seven AI developmental eras, we train and evaluate four multiclass classifiers: Logistic Regression, Random Forest, XGBoost, and LightGBM. All models achieve remarkably high performance, with Logistic Regression attaining the highest validation accuracy of 99.71%, macro-averaged F1 of 0.9973, and ROC-AUC of 0.9999 on the held-out test set. SHAP-based explainability analysis identifies Mental Wellbeing Score, Social Interaction Quality, and Emotional Intelligence as the dominant predictors of outcome class membership. These findings suggest that psychological and social dimensions of AI exposure are more strongly associated with mental growth trajectories than raw usage metrics. Implications for AI policy, digital mental health interventions, and human-centered AI design are discussed.