Mo Ahsan Ahmad, a Master of Science candidate in the Department of Mathematics and Statistics, will present their Masters project titled Advancing Generative Modeling and Applications with Boltzmann Machines, Restricted Boltzmann Machines, and Sum-Product Networks on Friday, December 20, 2024 at 10:30 AM online on Microsoft Teams.
The examination committee includes supervisors Dr. Ejaz Ahmed and Dr. Yifeng Li and Supervisory Committee Member Dr. Pouria Ramazi.
Students (both graduate and undergraduate) as well as other members of the Brock Community are invited to attend. If you are interested in the presentation, please contact Neil Marshall at nmarshall@brocku.ca for a link to the team. Please join with your microphones and camera turned off.
Keywords: Probabilistic Models, Boltzmann Machines, Restricted Boltzmann Machines, SPNs, FMNIST Dataset, Model Performance
Abstract:
We live in the era of advanced machine learning methodologies with promising applications in generative probabilistic modeling. This study explores advanced machine learning methodologies with promising applications in probabilistic modeling and real-world problem-solving. The investigation focuses on Boltzmann Machines (BMs), Restricted Boltzmann Machines (RBMs), and Sum-Product Networks (SPNs), emphasizing their ability to analyze complicated data distributions and reconstruct meaningful outcomes. BMs and RBMs, as energy-based probabilistic models, provide a strong foundation for capturing patterns in a variety of datasets. SPNs, with their hierarchical structure, allow for scalable probabilistic inference and efficient data representation. Using the Fashion MNIST dataset as a benchmark, this work demonstrates the practical performance of these models, highlighting reconstructed images, and precise predictions, alongside quantitative performance metrics. These findings are relevant for a variety of applications, such as image synthesis, object detection, and pattern recognition in domains like healthcare diagnostics and scientific research. The results highlight the distinct strengths of each strategy: the scalability and inference effectiveness of SPNs, as well as the capacity of BMs and RBMs to efficiently reconstruct and model data distributions. By providing a comparative analysis of these approaches, the study provides practical insights for both researchers and developers, demonstrating generative models’ revolutionary capability in developing machine learning and deep learning applications.
