Bayesian Multi-instance Learning: From Healthcare to Online Review Analysis
Abstract:
This talk explores the advancements in Bayesian multi-instance learning (MIL) and its applications in diverse fields. We will first delve into the utilization of state-of-the-art MIL techniques for cancer detection using T-cell receptor sequences, assessing their ability to address complex biomedical challenges. We will then introduce MICProB, a Bayesian MIL method based on probit regression, emphasizing its advantages in transparency, statistical inference, and explainability in addition to competitive performance. Furthermore, we will present vMMIC, a variational Bayesian multimodal MIC algorithm designed to overcome the limitations of MICProB by handling diverse instance types and improving computational efficiency for large-scale data. Through the lens of these novel approaches, this talk highlights the potential of Bayesian MIL in driving impactful solutions across domains ranging from healthcare to online review analysis.