Selecting optimal deep learning models is often a time-consuming process. To address this challenge, we propose a novel variant of the ant colony optimization (ACO) algorithm. This approach is designed to enhance model selection across various deep learning architectures, with a particular focus on leaf classification tasks. We introduce a new ACO technique specifically tailored for selecting robust models within convolutional neural networks (CNNs). These models are then integrated into an ensemble learning framework known as ensemble CNNs. A distinguishing feature of our proposed evolutionary ACO algorithm is its ability to consistently identify a set of robust CNN models in each iteration. This capability is facilitated by an innovative fitness function and an adaptive learning rate schedule embedded within the ACO algorithm, which optimizes pheromone distribution. Unlike the original ACO algorithm, which consistently selects the same CNN model, our evolutionary approach enables the dynamic discovery of new CNN models. To validate our method, we conducted experiments on two plant leaf datasets: Mulberry and Turkey-plant. Our comparison with existing methods, specifically the ant colony system (ACS) and the max-min ant system (MMAS), demonstrated that the MMAS algorithm outperformed the ACS algorithm. Furthermore, we explored three ensemble learning techniques: unweighted average, weighted average, and cost-sensitive learning. The weighted average method emerged as the most effective ensemble approach, with its parameters determined through a grid search process. The results indicate that the evolutionary ACO algorithm not only facilitates the selection of robust deep learning models but also achieves superior performance compared to the original ACO algorithm when applied to the Mulberry leaf and Turkey-plant datasets.
This manuscript presents a mulberry leaf dataset collected from five provinces within three regions in Thailand. The dataset contains ten categories of mulberry leaves. We proposed this dataset due to the challenges of classifying leaf images taken in natural environments arising from high inter-class similarity and variations in illumination and background conditions (multiple leaves from a mulberry tree and shadows appearing in the leaf images). We highlight that our research team recorded mulberry leaves independently from various perspectives during our data acquisition using multiple camera types. The mulberry leaf dataset can serve as vital input data passed to computer vision algorithms (conventional deep learning and vision transformer algorithms) for creating image recognition systems. The dataset will allow other researchers to propose novel computer vision techniques to approach mulberry recognition challenges. (c) 2024 The Author(s). Published by Elsevier Inc.
Polynomial multiplication is the main operation in lattice-based cryptography. Modular reduction is required to maintain the correctness within finite rings. In this study, the application of lazy reduction is chosen to reduce the frequency of modular operations in the LightSaber cryptosystem. Four polynomial multiplication algorithms, including Schoolbook, Karatsuba (128 and 64 splits), and Toom-3, are evaluated under both conventional and lazy reduction settings within the same polynomial ring domain. The experimental results show that lazy reduction reduces execution time across all evaluated algorithms. The reductions range from 44.52% to 59.15%. When the number of modular reduction operations is considered, they are reduced by more than 99% for Schoolbook and Karatsuba-based methods, and by 57.14% for Toom-3. Under lazy reduction, the number of modular reductions becomes identical across all algorithms. Among the evaluated coefficient-domain methods, Toom-3 achieved the lowest execution time. Karatsuba-based methods provide a more balanced trade-off between performance, memory usage, and structural complexity. This study further examines coefficient growth and evaluates the risk of overflow to ensure numerical reliability. All intermediate values remain within the range of signed 32-bit integers. Furthermore, lazy reduction leads to an increase in peak memory usage across all algorithms, with the most significant increase observed in Schoolbook, while more advanced methods exhibit only modest growth. Since this study focuses on coefficient-domain polynomial multiplication, transform-based approaches such as the Number Theoretic Transform (NTT) are beyond the scope of this work. NTT operates in a different computational domain and involves distinct implementation trade-offs. Accordingly, the reported performance comparisons are restricted to the coefficient-domain methods evaluated in this study. In addition, hardware-specific evaluations are beyond the scope of this work, because the primary objective is to provide an algorithm-level analysis of computational efficiency, modular reduction behavior, and numerical safety under controlled conditions. Under the evaluated Python implementation and the LightSaber parameter configuration, the results show that lazy reduction is an effective technique for reducing computational overhead in coefficient-domain polynomial multiplication. The observed performance remains dependent on the structural characteristics of each multiplication algorithm.
Course transfer credit evaluation is a crucial process in the educational system that helps students transfer their academic credits in accordance with university curriculum requirements. However, the traditional process faces challenges in comparing course content, as responsible instructors must manually review courses, and the transfer credit evaluation process takes considerable time. To address this challenge, we developed a semi-automated course transfer credit system using a Hybrid Method with Dynamic Weighting (HMDW) that combines Cosine Similarity, Levenshtein Distance, and Jaccard Similarity algorithms, and adjusts the weight of each method based on the characteristics of the transfer credit data. The implementation was divided into 4 phases: 1) synthesizing appropriate components and technologies for system development, 2) comparing the performance and accuracy of the proposed algorithms, 3) developing a semi-automated course transfer credit system using the HMDW in a web application format along with evaluating system quality, and 4) studying user acceptance of the system. The results indicate that the HMDW achieved high performance, with an F1-Score of 0.726 and a Recall of 1.000, ensuring that no eligible course for credit transfer was overlooked. In this system, Recall is the primary success criterion because a missed course that a student is eligible to transfer causes direct harm to the student both academically and financially, whereas a False Positive simply adds one item for faculty review. The reported Accuracy of 58.47% reflects standard Accuracy and should be interpreted alongside the Recall and F1-Score.The developed system can recommend courses and has achieved the highest instructor approval rate. Furthermore, experts evaluated the system quality at the highest level, and users also demonstrated the highest level of system acceptance. Although the system requires a high processing time for transfer credit evaluation, this time is acceptable given the increased accuracy. This demonstrates that combining the strengths of multiple algorithms with a dynamic weighting mechanism can improve course-matching performance compared to using a single method alone. The developed system can be implemented in practice to reduce instructors' workload and increase students' opportunities to receive appropriate, rapid transfer credit evaluation.