Personalization of AI Learning and Usage through Neural Networks: An Evaluation of Feedforward and Convolutional Architectures in the AI LEAP Project

This thesis explores the design and implementation of neural networks for the AI LEAP project, specifically targeted at enhancing educational experiences for children. The study involves the creation of two distinct neural network architectures: a feedforward neural network and a convolutional neural network (CNN). The primary focus was on the feedforward network, where extensive tests were conducted to optimize learning processes using limited data inputs, particularly images, and a small number of tuples. To enhance the learning efficiency, data were pre-processed using one-hot encoding, and data augmentation techniques were employed to artificially expand the training dataset. The runtime adjustments were critical in adapting the networks to work effectively with small datasets, a common challenge in real-world applications. The feedforward network was rigorously tested under various scenarios to assess its performance and adaptability, while the convolutional network was used to compare and benchmark the results. Key metrics were analyzed, demonstrating that strategic data handling and augmentation significantly improved the training outcomes, even with constrained input sizes. The results underscore the potential of tailored neural network models in educational AI, providing a foundation for further development of AI tools that are accessible and engaging for young users. This research highlights the importance of neural network customization to optimize performance in environments with limited data availability, offering insights into the practical applications of AI in educational projects for children.