Faculty of Computing

Permanent URI for this communityhttps://rda.sliit.lk/handle/123456789/4202

Browse

Filters

2020 - 2026607

Advanced Search

Filter by

Settings

Start date: 2020

Search Results

Now showing 1 - 10 of 607
  • Thumbnail Image
    PublicationEmbargo
    Data-centric single teacher guided knowledge distillation for alleviating sub-optimal supervision in image classification
    (Elsevier Ltd, 2026-02-23) Sharma, K; Silva, B. N
    In recent years, larger, deeper, and more complex deep learning models have emerged as a result of advancements in deep learning techniques. Nevertheless, the computational costs have also increased with the growing model size. Thus, Knowledge Distillation has evolved into a cornerstone in contemporary machine learning, facilitating the transfer of knowledge from cumbersome teacher models to more compact student models. However, student learning is persistently challenged by sub-optimal supervision caused by erroneous and ambiguous teacher predictions. Moreover, the learning process is further deteriorated by the complications introduced through frequently encountered noisy labels in real-world datasets. Existing methods often resort to the ensemble of teachers, introducing additional complexity. We propose a novel, simple, and efficient learning method, Corrective Knowledge Distillation (CKD), to alleviate these drawbacks while relying solely on a single-teacher model. The proposed work employs a two-phase learning paradigm. In the initial phase, the teacher selectively teaches extremely confident knowledge to the student, and in the subsequent phase, the student leverages its own past learning experiences, conditioning its knowledge acquisition on the guidance of the teacher. The proposed method consistently exhibits superior performance in addressing sub-optimal supervision, as evidenced by comprehensive experiments on benchmark datasets such as CIFAR-100, CIFAR-100N-Fine, and ImageNet-1K. Notably, CKD surpasses established baselines, achieving substantial accuracy gains of up to 3.53% in real-world scenarios. Furthermore, CKD exhibits exceptional robustness in highly noisy environments, outperforming ensemble techniques by a significant margin of up to 5.18%. Our code is available at https://github.com/Karthick47v2/ckd.
  • Thumbnail Image
    PublicationOpen Access
    Adding Common Sense to Robots Using Ontology
    (International Association of Computer Science and Information Technology, 2025-04-11) Ranathunga, R.A.A.L; Rajapaksha, S
    This work investigates how ontological frameworks might improve robots’ ability to reason using common sense. The goal of the project was to enhance robot decision-making in dynamic real-world situations by developing an ontology-based model retraining technique. The researchers wanted to incorporate organized commonsense knowledge into robotic systems, so they built extensive ontologies that captured knowledge about the physical world and human interactions. The research compared the performance of robots with conventional models (control group) to those with ontology-enhanced models (experimental group) across various measures. The results indicate that this strategy may be used to develop more competent and user-friendly robotic helpers for a variety of sectors, including industry, healthcare, and education. Although the study has limitations related to data quality and experimental design, it does demonstrate the promise of ontology-based techniques to advance autonomous systems and human-robot interactions. Extending ontology databases, multidisciplinary cooperation, and investigating applications in other sectors are some of the future research goals.
  • Thumbnail Image
    PublicationOpen Access
    AI-Driven Adaptive Security for Sensor Networks: Next-Generation Firewalls for Attack Detection
    (John Wiley and Sons, 2025-07-25) Meegammana, N.W; Fernando, H
    Sensor networks are increasingly critical in modern smart environments; however, their limited computational resources expose them to sophisticated cyber threats. Traditional static firewalls and computationally intensive deep learning models are impractical for securing such networks. This study proposes an adaptive next-generation firewall (NGFW) that dynamically switches between shallow and deep AI models based on real-time network load and resource availability. Four neural network models were trained using 20 and 40-feature subsets of the UNSW-NB15 dataset. Two runtime strategies (i) on-demand model loading and (ii) preloaded model switching were developed and evaluated through simulation under real-time conditions. Experimental results indicate that the preloaded approach achieves up to 96% accuracy, 98% precision, and 4-ms inference latency, with a memory footprint of 19 MB, outperforming static AI firewalls in both efficiency and scalability. The proposed NGFW framework offers a resilient and scalable solution for real-time attack detection in resource-constrained environments without requiring frequent model retraining. Future enhancements include hybrid shallow–deep model architectures, continuous federated learning for decentralized adaptability, and the integration of explainable AI to enhance transparency and trustworthiness in edge security deployments.
  • Thumbnail Image
    PublicationOpen Access
    Sri Lankan SMEs’ Performance Through Cloud Computing Adoption: An SEM-ANN Analysis
    (Institute of Electrical and Electronics Engineers Inc., 2025-04-25) Nawaz, S.S; Thelijjagoda, S
    This study identifies the determinants of cloud computing adoption and its effect on the performance of Sri Lankan small and medium-sized enterprises (SMEs). The Technology-Organization-Environment (TOE) framework, Technology Acceptance Model (TAM), and individual context were used to derive the study variables. This quantitative cross-sectional study adopted items from previous validated studies. Google Form was employed to collect data, and 418 responses were received from Sri Lankan SMEs. Partial Least Squares Structural Equation Modelling (PLS-SEM) via SmartPLS 4 and Artificial Neural Network (ANN) analysis via IBM SPSS 29 were used for data analysis. Based on the results, all hypotheses are confirmed except for one, and SME performance is significantly affected by cloud computing adoption. This study adds to the existing empirical evidence on cloud computing adoption by introducing an all-inclusive model that integrates the TOE, TAM, and individual factors. This demonstrates the effectiveness of the PLS-SEM/ANN hybrid methodology in analysing the determinants of cloud computing adoption. The significance of top management as a factor is highlighted by providing training and education to employees. Managers can benefit from this result by improving cloud computing adoption among SMEs in Sri Lanka. This is the first study of its kind in Sri Lanka, integrating the TOE, TAM, and individual variables and using a hybrid methodology combining PLS-SEM and ANN.
  • Thumbnail Image
    PublicationOpen Access
    Bridging Language Barriers in Programming Education: Java Programming Assistance Tool for Sinhala Native Speakers
    (International Association of Computer Science and Information Technology, 2025-09-12) Athukorala, K. S.N; De Silva, D.I
    This study presents an innovative programming assistance tool designed to address language barriers faced by Sinhala-speaking novice Java programmers. The tool provides real-time Java code generation and diagram creation based on Sinhala programming queries, enhancing conceptual understanding. Developed using a Design-Based Research methodology, the tool underwent iterative testing with 122 Sinhala-speaking learners, incorporating user feedback to refine usability and performance. Central to the system is Generative Pre-trained Transformer, version 3.5 Turbo, ensuring accurate translations and programming assistance, alongside a transformer-based model that translates Sinhala queries into English for processing. The translation model achieved 91.37% accuracy, with strong Bilingual Evaluation Understudy scores validating its contextual relevance. The tool’s practical applications extend beyond academia, supporting educational institutions, self-learners, and industry professionals in learning and skill development. Statistical evaluation of user performance demonstrated significant improvements in programming comprehension, reinforcing its effectiveness. By promoting inclusivity and expanding access to programming knowledge, this research contributes to the advancement of Sri Lanka’s technology sector and establishes a scalable framework for broader implementation in multilingual programming education. Copyright
  • Thumbnail Image
    PublicationOpen Access
    The Potential of Arboreal Tiger Beetle (Derocrania scitiscabra Walker) as a Natural Enemy for the Control of Fall Armyworm (Spodoptera frugiperda JE Smith)
    (Faculty of Agricultural Sciences, Sabaragamuwa University of Sri Lanka, 2025-01-15) Kasige R.H.; Abeywardhana D.L; Pallewatta N; Perera M.T.M.D.R; Dangalle C.D
    Purpose: The fall armyworm, Spodoptera frugiperda, is a serious pest of maize, sugarcane, and other crops in Sri Lanka. Natural predators and parasitoids are considered the best methods to control this insect pest. The present study investigates the possibility of using an endemic tiger beetle, Derocrania scitiscabra, as a natural predator for the larval stages of S. frugiperda. Research Method: The feeding preferences of D. scitiscabra to live prey versus dead prey, live prey types including different instar stages of S. frugiperda were investigated in the laboratory using choice tests. Fresh minced meat was used as dead prey, while red ants, earthworms and the six larval instar stages of S. frugiperda were used as live prey. Findings: Red ants were the most preferred prey type of D. scitiscabra, and dead prey, earthworms and mature S. frugiperda larvae were not consumed. Early larval instar stages of S. frugiperda were selected as prey, and the beetle showed a high feeding preference for the second larval instar stage. This feeding preference was observed irrespective of whether red ants were present or absent in the same environment. D. scitiscabra may have selected the second larval instars of S. frugiperda due to their small size, high prey density, mobility, and being devoid of injury. Research Limitations: The tests were conducted under laboratory conditions in insectary facilities. However, field investigations are essential to understand the ecological dynamics that affect insect behavior and survival. Originality/value: An endemic beetle is introduced to control S. frugiperda infestations in their early stages of development. The finding may provide an environmentally safe and economically beneficial method to control S. frugiperda.
  • Thumbnail Image
    PublicationEmbargo
    Deep Q-Network-Based Path Planning in a Simulated Warehouse Environment with SLAM Map Integration and Dynamic Obstacles
    (Department of Agribusiness, Universitas Muhammadiyah Yogyakarta, 2025-09-19) Medagangoda, H; Jayawickrama, N; de Silva, R; Samantha K.R.U.U; Abeygunawardhana, P.K.W
    With the rise of e-Commerce and the evolution of robotic technologies, the focus on autonomous navigation within warehouse environments has increased. This study presents a simulation-based framework for path planning using Deep Q-Networks (DQN) in a warehouse environment modeled with moving obstacles. The proposed solution integrates a prebuilt map of the environment generated using Simultaneous Localization and Mapping (SLAM), which provides prior spatial knowledge of static obstacles. The reinforcement learning model is formulated with a state space derived from grayscale images that combine the static map generated by SLAM and dynamic obstacles in real time. The action space consists of four discrete movements for the agent. A reward shaping strategy includes a distance-based reward and penalty for collisions to encourage goal-reaching and discourage collisions. An epsilon-greedy policy with exponential decay is used to balance exploration and exploitation. This system was implemented in the Robot Operating System (ROS) and Gazebo simulation environment. The agent was trained over 1000 episodes and metrics such as the number of actions executed to reach the goal and the cumulative reward per episode were analyzed to evaluate the convergence of the proposed solution. The results across two goal locations show that incorporating the SLAM map enhances learning stability, with the agent reaching a goal approximately 150 times, nearly double the success rate compared to the baseline without map information, which achieved only 80 successful episodes over the same number of episodes. This indicates faster convergence and reduced exploration overhead due to improved spatial awareness.
  • Thumbnail Image
    ItemOpen Access
    Optimizing Inotropic Infusion With Cluster Specific AI Decision Models and Digital TwinsNair, V
    (Institute of Electrical and Electronics Engineers, 2025-06-20) Nair, V. S; Niranga, G. D. H; Aryalakshmi C.S; Sathyapalan, D. T; Madathil, T; Pathinarupothi, R.K
    Inotropes are critical care medications essential for maintaining normal blood pressure (BP) in hospitalized patients. Titrating infusion rates of inotropes such as noradrenaline, vasopressin, and adrenaline based on fluctuating BP presents significant challenges in critical care settings. Typically, clinicians set a constant infusion rate for one hour, which may not accommodate the dynamic variability of BP inherent in critically ill patients, potentially leading to inadvertent hypotension or hypertension. Conventional feedback controllers, including fuzzy logic controllers (FLC), struggle to adapt to complex BP variations due to fixed algorithms and intracohort variability in drug responses. We propose an AI-enhanced closed-loop noradrenaline infusion control mechanism utilizing long short-term memory (LSTM) networks. This approach captures variability in drug responses through clustering of patients using LSTM autoencoders and K-means algorithms, subsequently developing LSTM-based decision models for infusion rates tailored to clusters. Additionally, a digital twin cardiac model serves as a simulation tool for validating the impact of inotropic infusion as indicated by the decision model. Comparative performance analyses demonstrate that our AI-enhanced closed-loop feedback method outperforms conventional systems like FLC regulators and pharmacokinetic-pharmacodynamic (PK-PD) models while ensuring patient safety as well as reducing the workload of clinicians.
  • Thumbnail Image
    ItemEmbargo
    Wet-Neuromorphic Computing: A New Paradigm for Biological Artificial Intelligence
    (Institute of Electrical and Electronics Engineers, 2025-03-31) Perera, J; Balasubramaniam, S; Somathilaka, S; Wen, Q; Li, X; Kasthurirathna, D; Roohi, A; Nelson, M. T
    As we delve into a life governed by artificial intelligence (AI), ongoing research continues to discover new forms of intelligence that are efficient and closely mimic an organism’s brain in terms of performance. This article presents a new concept termed wet-neuromorphic computing, in which biological cells or organisms are leveraged to perform computational tasks using their natural molecular functions. We map key neuromorphic properties to natural biological computing observed in bacteria, 3-D organoids, and Caenorhabditis elegans. To expand beyond the inspiration of the brain to create conventional neuromorphic computing, the study presents a case study that demonstrates bacterial AI computing using the gene regulatory neural network derived from Escherichia coli’s gene regulatory network for pattern recognition, validated through wet lab experiments. Finally, challenges and future directions are discussed.
  • Thumbnail Image
    PublicationOpen Access
    Enhancing Organizational Threat Profiling by Employing Deep Learning with Physical Security Systems and Human Behavior Analysis
    (Science and Information Organization, 2025) Senevirathna D.H; Gunasekara W.M.M; Gunawardhana K.P.A.T; Ashra M.F.F; Fernando, H; Abeywardena, K. Y
    Organizations need a comprehensive threat profiling system that uses cybersecurity methods together with physical security methods because advanced cyber-threats have become more complex. The objective of this study is to implement deep learning models to boost organizational threat identification via human behavior assessment and continuous surveillance activities. Our method for human behavior analysis detects insider threats through assessments of user activities that include logon patterns along with device interactions and measurement of psychometric traits. CNN, together with Random Forest classifiers, has been utilized to identify behavioral patterns that indicate security threats from inside the organization. Our model uses labeled datasets of abnormal user behavior to properly differentiate between normal and dangerous user activities with high accuracy. The physical security component improves surveillance abilities through the use of MobileNetV2 for real-time anomaly detection in CCTV video data. The system receives training to detect security breaches and violent and unauthorized entry attempts, and specific security-related incidents. The combination of transfer learning and fine-tuning methodologies enables MobileNetV2 to deliver outstanding security anomaly detection alongside low power requirements, thus it fits into Security Operations Centers operations. Experiments using our framework operate on existing benchmark collection sets that assess cybersecurity, together with physical security threats. Experimental testing establishes high precision levels for detecting insider threats along with physical security violations by surpassing conventional rule-based methods. Security Operation Centers gain an effective modern threat profiling solution through the application of deep learning models. The investigation generates better organization defenses against cyber-physical threats using behavioral analytics together with intelligent surveillance systems.