Uncertainty Estimation Using Variational Mixture of Gaussians Capsule Network for Health Image Classification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Capsule Networks (CapsNets) excel at image recognition but are often deterministic, limiting their trustworthiness.
• Existing CapsNets struggle with uncertainty and can be overconfident on unfamiliar data, hindering adoption in safety-critical fields.

Purpose of the Study:

• To develop a more reliable and interpretable Capsule Network by enabling it to express predictive uncertainty.
• To address the limitations of deterministic models in safety-critical applications.

Main Methods:

• Proposed a novel capsule network architecture utilizing a variational mixture of Gaussians.
• Trained distributions of network weights instead of single weight sets to model uncertainty.

Main Results:

• The proposed model effectively expresses predictive uncertainty on out-of-distribution data.
• Demonstrated faster convergence and reduced computational complexity compared to traditional Bayesian neural networks.
• Achieved performance comparable to state-of-the-art models while avoiding overfitting on smaller datasets.

Conclusions:

• The variational mixture of Gaussians approach enhances CapsNets with transparency, credibility, and reliability.
• This advancement makes CapsNets more suitable for practical adoption in fields like healthcare and autonomous driving.