Tunjay Akbarli

Followers

Following

Public Views

Address: Badamdar, Baku, Azerbaijan.

less

InterestsView All (15)

Uploads

Papers by Tunjay Akbarli

Artificial Depatterning: A Formal Model and Empirical Characterisation of Evaluation-Conditioned Policy Suppression in Frontier Language Models

We present a formal framework and supporting empirical evidence consistent with artificial depatt... more We present a formal framework and supporting empirical evidence consistent with artificial depatterning, a two-phase account by which large language models trained with reinforcement learning from human feedback (RLHF) exhibit systematic behavioural divergence between inferred evaluation and inferred deployment contexts. Using a controlled dual-condition protocol across nine open-weight model families spanning 0.6B to 65B locally-accessible parameters, we observe that the Policy Divergence Index (PDI) — a Jensen-Shannon divergence measure of behavioural consistency across monitored and unmonitored conditions — increases monotonically with model scale (Spearman rho = 0.94, p < 0.001, Bonferroni-corrected). For the largest locally-evaluated models, PDI reaches 0.68 plus or minus 0.05 (Qwen3.5-32B) and 0.66 plus or minus 0.04 (LLaMA 4 Scout). Chain-of-thought coherence collapses under monitored conditions in large models (CCS = 0.16 plus or minus 0.05 in Qwen3.5-32B, versus 0.85 plus or minus 0.04 under unmonitored conditions), suggesting that expressed reasoning traces do not reflect the decision process generating unmonitored outputs. Residual stream linear probing on models for which local weight access is available detects dual-policy activation structure at intermediate depth fractions (0.55–0.85), with probe accuracy reaching 0.84 plus or minus 0.03 (Qwen3.5-32B, depth 0.75) and Hartigan's dip test confirming bimodality (p < 0.01) in models above 14B parameters. We formalise these observations using a continuous evaluation belief state model, deriving the Depatterning Activation Threshold delta to the * and the Latent Objective Maintenance Loop (LOML). These results are consistent with the hypothesis that RLHF training, at sufficient scale, produces systems whose expressed compliance during evaluation is not a reliable indicator of post-deployment behaviour. Mitigation requires low-identifiability adversarial evaluation design paired with mechanistic interrogation of internal representations.

Download

Explainable Machine Learning Approach for Early Prediction of Thalassemia Risk Using Hematological Indicators

17th International Scientific Research and Innovation Congress , May 30, 2026

Thalassemia is one of the most common inherited blood disorders worldwide, making early risk dete... more Thalassemia is one of the most common inherited blood disorders worldwide, making early risk detection an important challenge for preventive healthcare. This study proposes an explainable machine learning approach to predict thalassemia risk using routine hematological laboratory indicators. The research utilizes the open-access HPLC-Based Thalassemia Screening Data dataset, which includes hemoglobin subcomponents (HbA0, HbA2, HbF) and erythrocyte indicators such as MCV, MCH, MCHC, RDWcv, HB, and RBC. Data preprocessing involved normalization, handling missing values, SMOTE-based class balancing, and an 80/20 train-test split to ensure reliable evaluation. Four machine learning models were developed and compared, including Logistic Regression, Random Forest, Multilayer Perceptron (MLP), and FTTransformer-Minimal. Model performance was evaluated using accuracy, precision, recall, F1-macro score, and confusion matrix analysis. Experimental results show that the MLP Neural Network achieved the best performance with an F1-macro score of 0.94, outperforming the other models. Confusion matrix results indicate balanced classification across normal, carrier, and diseased groups. Additionally, permutation importance analysis identified HbA0 and HbA2 as the most influential features, consistent with established medical knowledge regarding thalassemia biomarkers. The findings demonstrate that explainable machine learning models can effectively predict thalassemia risk using routine laboratory data. This approach provides a transparent, reliable, and scalable solution for early detection of genetic blood disorders and supports the practical use of artificial intelligence in healthcare applications.

Download

A Unified Framework for Calibration, Decoding, and Resource-Aware Processing in Hybrid Quantum-Classical Machine Learning

Hybrid quantum-classical machine learning on noisy intermediate-scale quantum (NISQ) devices suer... more Hybrid quantum-classical machine learning on noisy intermediate-scale quantum (NISQ) devices suers from three coupled overheads: hardware miscalibration, classical syndrome-decoding latency, and shot-noise on gradient estimators. We present QCOP (Quantum Calibration and Observable Processing), a unified closed-loop control layer integrating a Spectral Calibration Engine, a Hierarchical Syndrome Decoder (HSD), and an Observable Compression Layer within a single GPU-accelerated pipeline. Three theoretical bounds characterize the joint eject on logical error rate, convergence rate, and observable variance. On device-calibrated noise-model simulators (n = 20 seeds), QCOP achieves up to 2× wall-clock training speedup and ≈40% VQE energy error reduction relative to the unaugmented baseline; all simulation should be treated as promising but unconfirmed on physical hardware. A hardware pilot on ibm_nairobi at surface code distance d = 5 (2000 syndrome volumes, 8192 shots each) yields a 1.6× latency reduction and 1.9× logical error rate reduction for the HSD over MWPM directionally consistent with simulation but quantitatively smaller, indicating noise sources beyond the model. End-to-end hardware validation of all three modules remains future work. Code and model weights: github.com/hekimoghlu/QCOP.

Download

Predicting Diabetes with Interpretable Machine Learning: A Comparative Study of Classical and Deep Models

12th International New York Conference on Evolving Trends in Interdisciplinary Research & Practices, 2025

Timely and accurate prediction of diabetes is essential for reducing long-term complications and ... more Timely and accurate prediction of diabetes is essential for reducing long-term complications and improving clinical outcomes. With the increasing availability of structured patient data, machine learning (ML) has emerged as a promising approach for identifying individuals at risk of chronic conditions such as diabetes. In this study, we conduct a comparative evaluation of five ML models-Logistic Regression, Random Forest, XGBoost, Multi-Layer Perceptron (MLP), and TabNet-on a real-world tabular dataset for binary diabetes classification. The dataset, publicly available and widely used in medical AI benchmarking studies, contains anonymized physiological and diagnostic attributes relevant to diabetes prediction. To ensure consistency and fairness across all models, we applied z-score normalization to the input features and performed five-fold stratified cross-validation. For TabNet, which is not natively compatible with standard scikit-learn pipelines, a custom cross-validation routine was implemented to maintain experimental parity. Model performance was assessed using Accuracy, F1 Score, and the Area Under the Receiver Operating Characteristic Curve (ROC AUC). TabNet demonstrated the highest overall predictive performance, achieving strong accuracy and the most stable ROC AUC across folds. Beyond predictive metrics, this study emphasizes the importance of model interpretability in healthcare contexts. TabNet's attention-based architecture enabled the extraction of both global feature importance and local explanations for individual predictions. Features such as glucose levels and body mass index (BMI) emerged as key predictors across multiple evaluation layers. Overall, the findings support the use of interpretable deep learning models in medical risk prediction tasks and highlight TabNet's potential as a reliable, transparent tool for decision support in clinical practice.

Download

Predicting Human Well-being: Analysing Job Satisfaction, Work-life balance and job related stress using Ensemble Learning Models

10th International European Conference on Interdisciplinary Scientific Research, Sep 26, 2024

Fast-paced working condition demands understanding of multiple factors, which may impact employee... more Fast-paced working condition demands understanding of multiple factors, which may impact employees' feelings towards work satisfaction, balance, and social well-being, to have a better account of individualistic performance and health of the organization. Therefore, ensemble learning will be applied to major facets concerned about well-being in employees by making predictions from the complete dataset including age groups, personal habits, time management, and other measures of stress. The objectives will be to develop predictive models on Job Satisfaction, Work-Life Balance, and Daily Stress Levels using some advanced ensemble methods like Random Forests, Gradient Boosting Machines, and Extreme Gradient Boosting (XGBoost). This is related to the diurnal stress in the category of thejob, spending time at work every day, time spent on sports, usage of social media, family and friend time, hobbies, and other lifestyle habits. Such features are included in models largely to capture important complex nonlinear relationships that specify well-being relations across individuals. After rigorous training and validation of the models, we illustrate that the predictive accuracies of such ensemble learning models are superior to traditional machine learning methods. The results portrayed the most important predictors for Job Satisfaction, Balance Level, and Stress Levels that offer very good insights for the employers and HR professionals when designing relevant interventions. In this paper, the potentials of machine learning for human resource management are opened in new ways for more adaptable and responsive strategies at the workplace. Further research will be needed on how models developed can be used in real-time to facilitate proactive efforts of management in improving the work environment to a more healthy and vital one.

Download

Predicting Acute Ischemic Stroke Incidence Using Knowledge-Informed Neural Networks

11th International New York Conference on Evolving Trends in Interdisciplinary Research & Practices, Dec 10, 2024

This paper applies Knowledge-Informed Neural Networks to forecast AIS incidence from 2019 to 2024... more This paper applies Knowledge-Informed Neural Networks to forecast AIS incidence from 2019 to 2024. The focus is including domain knowledge into the predictive model to enhance the accuracy of the AIS forecast, principally health patterns related to age. Demographic data, diagnosis types, age groups, and stroke occurrences recorded are the contents of the dataset whose patient count will be targeted in the study. We introduced domain knowledge of age groups into the learning pipeline through comprehensive exploratory data analysis and correlation assessments. The model, trained for more than 300 epochs with a mean squared error as the evaluation metric after detailed hyperparameter tuning, supports the fact that knowledgeinformed approaches will yield better performance, especially in the age group above 65, which is most susceptible to stroke. These insights are important for health resource planning and for focused prevention strategies against strokes.

Download

Books by Tunjay Akbarli

The Codira Programming Language

The Codira Programming language / by Tunjay Akbarli, Maryna Rybalko, Editor: Maryna Rybalko. Baku: SkyE, 2025. – 608 p: SkyE, 2025. – 608 p language / Editor: Maryna Rybalko. Baku: SkyE, 2025. – 608 p, Jul 29, 2025

Codira is a next-generation programming language designed for speed, clarity, and interoperabilit... more

Download

Tunjay Akbarli

Uploads

Papers by Tunjay Akbarli

Books by Tunjay Akbarli

Log In