Stochastic Neural Networks and Their Applications
Faculty Mentor
Dr. Ionut Emil Iacob
Location
Russell Union Ballroom
Type of Research
On-going
Session Format
Oral Presentation
College
College of Science & Mathematics
Department
Department of Mathematical Sciences
Abstract
Traditional deterministic neural networks often struggle with modeling uncertainty and making robust predictions in noisy or sparse data environments. Stochastic Neural Networks (SNNs) address these limitations by incorporating probability distributions directly into the network's architecture, typically by sampling weights, activations, or inputs from defined distributions rather than using fixed scalar values. By leveraging this probabilistic framework, SNNs inherently provide measures of epistemic uncertainty (model uncertainty) and aleatoric uncertainty (data uncertainty). Key methodologies include Bayesian Neural Networks, which utilize variational inference to approximate posterior distributions, and dropout-based techniques like Monte Carlo Dropout. These networks find critical applications in domains requiring robust decision-making under uncertainty, such as: - Autonomous Driving: Assessing the confidence level of object detection to prevent critical failures. - Medical Imaging: Identifying potential diagnostic errors by highlighting areas of high uncertainty. - Financial Modeling: Predicting market trends while quantifying risk exposure. Through probabilistic inference, SNNs offer enhanced generalization capabilities and improved performance in high-stakes applications where knowing how uncertain a prediction is crucial.
Program Description
.
Start Date
4-23-2026 2:30 PM
End Date
4-23-2026 2:45 PM
Recommended Citation
Barone, Arden, "Stochastic Neural Networks and Their Applications" (2026). GS4 Student Scholars Symposium. 242.
https://digitalcommons.georgiasouthern.edu/research_symposium/2026/2026/242
Stochastic Neural Networks and Their Applications
Russell Union Ballroom
Traditional deterministic neural networks often struggle with modeling uncertainty and making robust predictions in noisy or sparse data environments. Stochastic Neural Networks (SNNs) address these limitations by incorporating probability distributions directly into the network's architecture, typically by sampling weights, activations, or inputs from defined distributions rather than using fixed scalar values. By leveraging this probabilistic framework, SNNs inherently provide measures of epistemic uncertainty (model uncertainty) and aleatoric uncertainty (data uncertainty). Key methodologies include Bayesian Neural Networks, which utilize variational inference to approximate posterior distributions, and dropout-based techniques like Monte Carlo Dropout. These networks find critical applications in domains requiring robust decision-making under uncertainty, such as: - Autonomous Driving: Assessing the confidence level of object detection to prevent critical failures. - Medical Imaging: Identifying potential diagnostic errors by highlighting areas of high uncertainty. - Financial Modeling: Predicting market trends while quantifying risk exposure. Through probabilistic inference, SNNs offer enhanced generalization capabilities and improved performance in high-stakes applications where knowing how uncertain a prediction is crucial.
