Investigation of meta-learning to enhance supervised learning
Haque, Md Enamul (2024)
2024
Master's Programme in Computing Sciences
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-05-30
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202405206027
https://urn.fi/URN:NBN:fi:tuni-202405206027
Tiivistelmä
This thesis investigates the application of meta-learning techniques to enhance the supervised machine-learning model’s adaptability, efficiency, and performance. Meta-learning, or "learning to learn," enables models to refine their learning processes based on past experiences, promising significant improvements over supervised machine-learning methods. The primary objectives of this research were to optimize supervised machine learning algorithms and enable these models to manage and adapt iteratively to new data without the need for complete retraining.
Employing the MNIST dataset, the study involved preprocessing, creating paired data for sum prediction, and implementing a train-test-validation split. A base model was developed using a convolutional neural network (CNN) on a subset of the data, which served as a baseline for further enhancements through a modified meta-learning algorithm inspired by Model-Agnostic Meta-Learning (MAML). Subsequent adaptations incorporated online meta-learning techniques to handle data arriving over time, simulating real-world scenarios where data dynamics continuously evolve.
The results demonstrated that the meta-learning models improved learning efficiency and adaptability and enhanced generalization capabilities across new data. These findings substan- tiate the literature on meta-learning’s capacity to significantly upgrade learning processes, high- lighting its practical applications across various fields. The research confirmed that meta-learning frameworks could dynamically adapt and learn from sequential data inputs, showcasing a robust improvement trajectory from the base model to the meta-model and, finally, to the online meta- model.
This study demonstrated the potential of meta-learning within artificial intelligence, particularly in enhancing machine learning models through iterative and adaptive learning. By extending the capabilities of supervised models to adapt to new and changing data swiftly without extensive retraining, meta-learning holds promise for broad applicability in more complex, real-life scenarios. Future research may explore its integration with unsupervised learning models and tackle the challenges of scalability and computational efficiency, further advancing the frontiers of machine learning technology.
Employing the MNIST dataset, the study involved preprocessing, creating paired data for sum prediction, and implementing a train-test-validation split. A base model was developed using a convolutional neural network (CNN) on a subset of the data, which served as a baseline for further enhancements through a modified meta-learning algorithm inspired by Model-Agnostic Meta-Learning (MAML). Subsequent adaptations incorporated online meta-learning techniques to handle data arriving over time, simulating real-world scenarios where data dynamics continuously evolve.
The results demonstrated that the meta-learning models improved learning efficiency and adaptability and enhanced generalization capabilities across new data. These findings substan- tiate the literature on meta-learning’s capacity to significantly upgrade learning processes, high- lighting its practical applications across various fields. The research confirmed that meta-learning frameworks could dynamically adapt and learn from sequential data inputs, showcasing a robust improvement trajectory from the base model to the meta-model and, finally, to the online meta- model.
This study demonstrated the potential of meta-learning within artificial intelligence, particularly in enhancing machine learning models through iterative and adaptive learning. By extending the capabilities of supervised models to adapt to new and changing data swiftly without extensive retraining, meta-learning holds promise for broad applicability in more complex, real-life scenarios. Future research may explore its integration with unsupervised learning models and tackle the challenges of scalability and computational efficiency, further advancing the frontiers of machine learning technology.