Two papers from BMIC group accepted at the upcoming NeurIPS 2023 conference
We are excited to present two new papers accepted for NeurIPS 2023
"Expert load matters: operating networks at high accuracy and low manual effort” by Sara Sangalli, Ertunc Erdil and Ender Konukoglu:
In human-AI collaboration systems for critical applications, users should set an operating point based on model confidence to delegate decisions to human experts for samples with low confidence, particularly crucial in fields like healthcare. We propose a new loss function that maximizes the area under the 'confidence operating characteristic (COC) curve,' achieving improved network accuracy, fewer human-delegated decisions, and on-par calibration performance on various computer vision and medical image datasets.
"Canonical normalizing flows for manifold learning” by Kyriakos Flouris and Ender Konukoglu:
We propose an advanced method in manifold learning flows, introducing the 'Canonical Manifold Learning Flow' technique, which employs a novel optimization strategy to enforce the adoption of sparse and orthogonal basis functions by the transformation matrix. This innovative approach facilitates a more efficient use of latent space, automatically producing fewer yet non-degenerate dimensions for data representation, and delivering a notably improved approximation of target distributions. This is evidenced by learning orthogonal coordinates of simulated data manifolds and exhibiting lower FID scores in numerous experiments. This method promises to enhance the efficiency and precision of data representation, while introducing a novel optimization method via the metric tensor of a Riemannian manifold.