learning with biased labels

Deep Neural Networks (DNNs) have gained popularity in a wide range of applications. The remarkable success of DNNs often relies on the availability of high-quality datasets. However, the acquisition of a large amount of well-annotated unambiguous data could be very expensive and sometimes even inaccessibl. Standard training using ambiguous data may produce overly confident models and thus leading to poor generalization.

During my stay at Baidu Research, I have been have been actively engaged in the realm of learning with biased labels, with a specific focus on enhancing model robustness and reliability in the presence of noisy data. This presentation serves as a comprehensive overview of my work. It encompasses the following key components:

• A concise review of various methodologies for learning with noisy labels;
• A specific topic we brought up – addressing the challenge of training data with highly ambiguous labels, such as those that provide an incomplete description of the object. (This part is temporarily deleted from this post.)

I believe that our research may shed some light on how to build a more trustworthy machine learning model, especially in domains where data quality is a critical factor.