<p>With high-quality annotation data, edge AI has emerged as a pivotal technology in various domains. Unfortunately, due to sensor errors and discrepancies in data collection, datasets often suffer from noisy labels. Identifying and relabeling all the noisy data becomes imperative, but it’s labor-intensive and time-consuming. To ensure the robustness of resource-constrained edge AI models with noisy labels, in this paper, we propose an efficient selective relabeling method</p><p>leveraging expert knowledge, termed “Co-active”. The method involves three important steps: Noisy Data Identification, Informative Data Selection, and Model Re-training. Initially, we pre-train an encoder model with early stopping to detect noisy instances by analyzing the prediction discrepancies between two classifiers that have been initialized differently. Then we select the most informative noisy data from previous instances by introducing a novel priority scorer that combines entropy and dynamic loss variations. Following this, we utilize the mixup technique to retrain the model. The retraining process involves a dataset that is a composite of clean, relabeled, and data that is potentially clean. This is facilitated by a novel loss function that adeptly balances classification accuracy with regularization terms. Additionally, we introduce strategies for dynamically adjusting the size of the relabel dataset to optimize the labeling budget and enhance model robustness. Our extensive experiments across four datasets demonstrate that Co-active consistently achieves the best performance, with an average improvement of 18.67%.</p>