http://210.212.227.212:8080/xmlui/handle/123456789/554 2026-05-17T00:01:38Z 2026-05-17T00:01:38Z William, Tom Jacob Sumod, Sundar http://210.212.227.212:8080/xmlui/handle/123456789/571 2024-07-08T05:43:42Z 2024-06-30T00:00:00Z

STUDENT-CENTERED AUTOMATED FEEDBACK USING TRANSFORMER BASED APPROACH William, Tom Jacob; Sumod, Sundar Effective communication between teachers and students is essential in today’s educational environment. This project presents a transformer model that has been designed for the ed- ucational environment. With the use of this model, students can have personalized learning experiences and ask academic questions and while getting immediate answers generated by the model. The personalized feedback is given by using a deep learning model. Due to lack of dataset existing in education domain, a dataset on education domain is developed using web scraping techniques, which scrapes or extracts the relevant data from internet. The data which is obtained through web scrapping is given to the model for training. The deep learn- ing model used is Generative Pre-trained Transformer-2 (GPT-2) a transformer model which then analyses the pattern and gives the customized feedback back to student. The GPT-2 was used for training on the data set. After performing the validation between reference and generated text, the evaluation metrics were obtained. The evaluation metrics showed very low BLEU score and ROUGE score. To increase the value of BLEU score and ROUGE score, data augmentation was performed. Data augmentation technique such as synonym replacement was performed to increase the data size. Parameter efficient fine-tuning tech- niques such as Low rank Adaption is applied to reduce the size of the model which in turn can produce the results of the model at much less computation time and improved memory efficiency. The integration of this model is expected to transform the educational system by creating an atmosphere in which students are inspired to learn and teachers can skilfully meet each student’s unique learning needs.

2024-06-30T00:00:00Z Shanimol, Shajan Christy, D Ponnan http://210.212.227.212:8080/xmlui/handle/123456789/570 2024-07-08T05:39:38Z 2024-06-30T00:00:00Z

DEEP LEARNING BASED ARRHYTHMIA DISEASE DETECTION USING ELECTROCARDIOGRAM SIGNALS Shanimol, Shajan; Christy, D Ponnan The primary objective of this work is to use the patient’s electrocardiogram (ECG) to develop a method for early diagnosis of Arrhythmia disease. Traditional methods of diagnosing Ar- rhythmia often require extensive tests and a laborious workup in a hospital setting. Improved diagnostic technologies are needed to enable early and precise detection of cardiovascular diseases, particularly Arrhythmias. In this work, deep learning is integrated with signal processing techniques to detect Arrhythmia disease, using a dataset from the well-known MIT-BIH Arrhythmia Database. The first step is to convert the raw ECG information into visual representations using Wavelet transforms. These transforms capture precise temporal and frequency properties, unveiling detailed information about the underlying complexities of the heart rhythm. A Convolutional Neural Network (CNN) model specifically designed for Arrhythmia detection is then trained using this transformed data. Deep learning uti- lizes pre-existing knowledge from established architectures, enhancing the model’s ability to detect patterns indicative of Arrhythmia diseases. The proposed methodology is carefully evaluated on a testing set using performance metrics such as accuracy, sensitivity, and speci- ficity. These performance metrics are crucial indicators of the model’s ability to accurately distinguish between normal, web, and even cardiac rhythms. Overall, this research aims to develop a novel approach combining deep learning and signal processing techniques to enhance the detection of Arrhythmia diseases using ECG data.

2024-06-30T00:00:00Z Sachin, John Thomas Imthias, Ahamed T P http://210.212.227.212:8080/xmlui/handle/123456789/569 2024-07-08T05:35:01Z 2024-06-30T00:00:00Z

PATH PLANNING IN ROBOTICS USING HYBRID Q-LEARNING APPROACH Sachin, John Thomas; Imthias, Ahamed T P Reinforcement learning is a technique that enables agents to learn optimal behaviors through interactions with their environment, using rewards and penalties to shape their actions. In this project, we address the challenge of enabling a mobile robot to navigate through environments such as a factory layout or a hospital setting while avoiding collisions with obstacles. The main objective of the agent is to navigate through these environments without colliding with any of the static or dynamic obstacles in its way. The robot, or the agent, is equipped with three proximity sensors that determine the proximity of the obsta- cles in their respective directions. The learning is achieved through the combination of a reinforcement learning approach known as the Q-learning algorithm, which is a value-based reinforcement learning technique, and the A* algorithm, a heuristic-based search algorithm. Q-learning is notable for its ability to handle problems with various state and action spaces, as well as its simplicity and versatility in various applications, including robotics, game play- ing, and autonomous systems. However, Q-learning faces the difficulty of large state spaces. To address this, we consider a heuristic approach to handle both large spaces and testing in unknown and dynamic environments. For testing and visualization, Python’s Pygame library is involved. The agent undergoes training within a grid-based environment. The hybrid approach of combining Q-learning with the A* algorithm ensures faster learning and lesser computational time. This combination leverages the strengths of both methods, with Q-learning providing robust policy learning and A* offering efficient pathfinding through heuristic search. This ensures that the agent learns to efficiently navigate complex environ- ments while minimizing computational overhead, ultimately enhancing its ability to operate autonomously and safely in real-world scenarios.

2024-06-30T00:00:00Z Reniya, Shajahan Fousia M, Shamsudeen http://210.212.227.212:8080/xmlui/handle/123456789/568 2024-07-08T05:31:32Z 2024-06-30T00:00:00Z

Financial and Investment Management System using Transformers Reniya, Shajahan; Fousia M, Shamsudeen In the contemporary landscape of stock market analysis, the utilization of advanced nat- ural language processing techniques has become increasingly prevalent. This work presents the application of the Phi 2 Transformer Model, a compact yet potent language model, for the classification of stock news articles as either indicative of a buy or sell recommendation. Leveraging the Phi 2 model’s demonstrated accuracy of 89% in stock news classification, this study contributes to the development of sophisticated decision support systems for investors.The methodology involves preprocessing a comprehensive dataset of stock news ar- ticles, incorporating relevant contextual features such as industry-specific terminology, and market trends. Notably, the closing price of the respective company serves as a crucial deter- minant in analyzing the news’s trend and discerning the overall market trend. Through the integration of this contextual information with the Phi 2 Transformer Model, the system can effectively classify incoming news articles into actionable buy or sell recommendations.The proposed model offers valuable insights for investors by automating the time-consuming pro- cess of manually analyzing news articles and market trends. By harnessing the predictive capabilities of the Phi 2 model, coupled with real-time market data, investors can make more informed decisions, potentially enhancing their investment strategies and overall port- folio performance.

2024-06-30T00:00:00Z