This website requires JavaScript.

Data Collection and Quality Challenges in Deep Learning: A Data-Centric AI Perspective

Steven Euijong WhangYuji RohHwanjun SongJae-Gil Lee
Software 2.0 is a fundamental shift in software engineering where machinelearning becomes the new software, powered by big data and computinginfrastructure. As a result, software engineering needs to be re-thought wheredata becomes a first-class citizen on par with code. One striking observationis that 80-90% of the machine learning process is spent on data preparation.Without good data, even the best machine learning algorithms cannot performwell. As a result, data-centric AI practices are now becoming mainstream.Unfortunately, many datasets in the real world are small, dirty, biased, andeven poisoned. In this survey, we study the research landscape for datacollection and data quality primarily for deep learning applications. Datacollection is important because there is lesser need for feature engineeringfor recent deep learning approaches, but instead more need for large amounts ofdata. For data quality, we study data validation and data cleaning techniques.Even if the data cannot be fully cleaned, we can still cope with imperfect dataduring model training where using robust model training techniques. Inaddition, while bias and fairness have been less studied in traditional datamanagement research, these issues become essential topics in modern machinelearning applications. We thus study fairness measures and unfairnessmitigation techniques that can be applied before, during, or after modeltraining. We believe that the data management community is well poised to solveproblems in these directions.