HAICO-CN: Human-AI Collaboration By Cluster-wise Noisy-Label Augmentation
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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: Human-Ai Collaboration, Noisy-label learning, Multi-rater learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: The intricate dynamics of human-AI collaboration presents an ongoing chal- lenge. While recent research incorporates human behaviors into machine learn- ing model design, most utilise single global confusion matrix or human behavior model, disregarding potential personalization to user. To address this gap, we propose HAICO-CN, a human-AI collaborative method that enhances human-AI joint decision-making by training personalized models using a novel cluster-wise noisy-label augmentation technique. During training, HAICO-CN first identifies and clusters noise label patterns within the multi-rater data sets, followed by a cluster-wise noisy-label augmentation method that generates enough data to train a collaborative human-AI model for each cluster. During inference, the user fol- lows an onboarding process, allowing HAICO-CN to select a cluster-wise human- AI model based on the user’s noisy label patterns, thereby enhancing human-AI joint decision-making performance. HAICO-CN is simple to implement, model- agnostic, and effective. We propose new evaluation criteria for assessing human- AI collaborative methods and empirically evaluate HAICO-CN across diverse datasets, including CIFAR-10N, CIFAR-10H, Fashion-MNIST-H, and Chaoyang histopathology, demonstrating HAICO-CN’s superior performance compared to state-of-the-art human-AI collaboration approaches.
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: 5191
Loading