New Data-Centric AI Class from MIT

In traditional machine learning classes, students learn methods to create effective models for specific datasets. However, data is often disorganized in real-world situations, and improving models may not be the only solution to improve performance. Instead of assuming the data is fixed, it can be beneficial to enhance the dataset itself. Data-Centric AI (DCAI) is an emerging field that investigates techniques to enhance datasets, which is often the best approach to enhancing performance in practical machine learning applications. Although experienced data scientists have traditionally accomplished this through trial and error and intuition, DCAI offers a systematic engineering discipline for improving data.

This course on DCAI is the first of its kind, teaching students algorithms for identifying and resolving typical problems in machine learning data and constructing better datasets, with a particular focus on data used in supervised learning tasks, such as classification. The course material emphasizes practical skills essential in real-world machine learning applications rather than the mathematical details of how specific models operate. Students who take this course can learn valuable techniques not typically included in most machine-learning classes. These techniques can assist in overcoming the "garbage in, garbage out" issue prevalent in many practical machine-learning applications.