22 April 2025 in Babylonian Journal of Artificial Intelligence
The 2017 paper 'Attention Is All You Need' by Vaswani et al. marked a major paradigm shift in AI. Rather than clinging to tired methods like recurrence or convolutions, its Transformer design boldly flipped the script melding self-attention into a fresh take that not only remade natural language processing but also trickled over into computer vision, robotics, quirky multi-modal setups, and more [1]. At its very core lies a refreshingly simple idea self-attention, which lets models dynamically figure out which bits of the input deserve extra focus instead of being bound by strict, linear routines. This clever tweak kicked the old, rigid rules to the curb and opened up levels of parallel processing that we hadn’t seen before. In many cases, this shift not only ramped up the training speed dramatically but also proved essential in our data-swamped world where speed and nimbleness really do make all the difference.
1 February 2025 in EDRAAK
This research investigates the classification of Iraq's economic performance from 2000 to 2023 using machine learning methodologies, specifically Support Vector Machines (SVM) and Random Forest (RF). The study aims to leverage macroeconomic indicators to develop predictive models for economic state classification. The dataset comprises six key variables gross domestic product, inflation rate, unemployment rate, exchange rate, trade volume, and government spending selected for their economic relevance. The methodology employs a composite Economic Performance Index for binary categorization ("Good" or "Poor"), followed by the application of SVM and RF algorithms. Model performance is evaluated using metrics such as AUC, F1-score, precision, recall, and MCC. Results indicate that SVM demonstrates superior classification accuracy compared to Random Forest, highlighting its effectiveness in capturing complex economic relationships. Gross domestic product, inflation, and government spending are identified as the most influential factors. This research contributes to the growing intersection of machine learning and macroeconomic forecasting, providing policymakers with data-driven insights for economic policy evaluation.
18 April 2025 in Applied Data Science and Analysis
Geometric programming (GP) is a powerful framework for optimizing posynomial and signomial functions, which are widely used in engineering design, economics, and other fields. A critical aspect of solving geometric programs efficiently lies in understanding the convexity properties of the functions involved, as convexity ensures that local minima are also global minima. This work focuses on correcting and verifying the convexity proofs for two specific signomial functions in geometric programming: f 1 (x)=x 1 2 x 2 −1 +x 2 3 −x 1 x 2 and f₂(x) = x₁³x₂⁻² + x₁⁻¹x₂² - x₁x₂⁻¹. Through a detailed analysis, we demonstrate that f1(x) is **not convex** over the domain (x1,x2)> 0, as its Hessian matrix fails to be positive semi-definite for all positive values of x1and x2. For f2(x), the convexity **cannot be conclusively determined** based on the analysis, as the logarithmic transformation does not universally preserve convexity for signomial functions with negative coefficients. Specific cases suggest local convexity, but a general proof for all (x1,x2)> 0 remains elusive. These results highlight the challenges in analyzing signomial functions and emphasize the importance of rigorous convexity verification in geometric programming. The findings have significant implications for optimization feasibility and modeling considerations, particularly in applications where signomial functions with negative coefficients arise. This work provides a foundation for further research into advanced techniques for analyzing and optimizing such functions, ensuring more robust and reliable solutions in geometric programming.