محمدجواد ميرشكاري حقيقي

The success of deep neural networks largely relies on large, labeled datasets, which are costly and time-consuming to acquire. To address this, methods like crowdsourcing and extracting labels from web texts are used to reduce costs and time. Crowdsourcing involves non-experts for data labeling, leading to potential errors and inconsistencies. Similarly, while extracting labels from web texts is efficient, it often suffers from errors due to noise in web data.Unfortunately, these methods have resulted in noise entering real-world datasets, and the presence of incorrect labels in the learning process leads to a decrease in model accuracy and efficiency, highlighting the importance of developing robust methods for identifying and managing incorrect labels. Addressing this challenge requires approaches for identifying and correcting incorrect labels to maintain the accuracy and robustness of deep learning models in practical applications. In this research, to provide a solution for correcting incorrect labels in datasets, we first eva‎luate metrics for separating noisy data, followed by the introduction of a stability metric for distinguishing noisy data. Then, in order to correct incorrect labels and learn in a noisy space, we focus on training a robust model in a noisy environment using this metric alongside a framework for correcting incorrect labels and selecting reliable samples, combined with semi-supervised learning.Subsequently, we eva‎luate the proposed method from two aspects: its ability to correct incorrect labels and learn in a noisy space. For this eva‎luation, we utilize datasets with both artificial and natural noise, such as CIFAR10N and ANIMAL10N.In the case of artificial noise, the model's accuracy on test data has improved on average by 0.88% for noise levels below 50%, while in natural noise, it has improved by an average of 0.14%. Additionally, the proposed method has successfully corrected on average up to 73% of incorrect labels across various artificial noise levels, achieving an F1 score of 98.29 in detecting incorrect labels.

برچسب‌هاي نادرست , مجموعه داده‌هاي نويزي , شبكه‌هاي عصبي عميق , اصلاح برچسب , مدل مقاوم , پايداري برچسب

incorrect labels , noisy datasets , deep neural network , label correction , semi-supervised learning , robust model , label stability

mohammad javad mirshekari haghighi

Dr.mohammad reza mohammadi