Enhancing Machine Learning System Reliability in Healthcare through Uncertainty Estimation and Multi-Modal Learning
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Gnostic Uncertainty Estimation; Machine Learning Reliability; Uncertainty Estimation; Healthcare
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: It is crucial to ensure the dependability of machine learning (ML) systems, especially in areas where safety is a top priority, like healthcare. A tried-and-true method for highlighting the reliability of ML systems during deployment is uncertainty estimation. By successfully using integrated feature sets, sequential and parallel ensemble algorithms have both shown improved ML system performance in multi-modal contexts. We provide Uncertainty-Receptive fusing (URF), a cutting-edge technique that uses uncertainty estimations to improve the fusing of predictions from several base learners. URF, which successively modifies the weighting of the loss function during training in contrast to conventional boosting techniques, is especially successful for multi-modal learning tasks. In order to understand how noise and spatial transformations affect image-based activities, we then offer an image acquisition model that takes these aspects into consideration. We can make predictions with greater accuracy utilizing latent variables thanks to this approach. To quantify uncertainty at the pixel and structure/lesion levels, we use entropy-based uncertainty assessment (EUA). EUA measures the variety within prediction distributions and provides insightful information about the model's confidence. We also present Gnostic Uncertainty Estimation (GUE), which quantifies the model's lack of knowledge regarding the result and helps to comprehend the accuracy of the prediction.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5395
Loading