Georgia Tech SCL Georgia Tech SCL Georgia Tech SCL

Yao Xie, a Coca-Cola Foundation Chair and Professor in ISyE, and Chen Xu, Operations Research Ph.D. student, just had their paper "Invertible normalizing flow neural networks by JKO scheme" accepted as a spotlight paper at Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023). 

 The conference is set to be held at the New Orleans Ernest N. Morial Convention Center from December 10 to 16, 2023. 

NeurIPS 2023 is an interdisciplinary conference that brings together researchers in machine learning, neuroscience, statistics, optimization, computer vision, natural language processing, life sciences, natural sciences, social sciences, and other adjacent fields.  

The core focus is peer-reviewed novel research, which is presented and discussed in the general session, along with invited talks by leaders in their fields. 

The authors introduced a new computer model, JKO-iFlow, to enhance the efficiency of generating realistic data. This model belongs to a family of models known as normalizing flows, which excel in quickly creating realistic data, especially when dealing with a large amount of information. 

In contrast to other models, the authors simplified JKO-iFlow by structuring its components in a way that improves training efficiency. Drawing inspiration from the Jordan-Kinderleherer-Otto scheme, they unfolded the dynamic of the Wasserstein gradient flow. This approach facilitates better training without the need for complex calculations, making it more accessible for the computer to learn. 

Experiments with various types of data demonstrate that JKO-iFlow performs comparably to other advanced models but operates much faster and with lower computational memory requirements.  

The JKO-iFlow model not only demonstrates competitive performance in generating realistic data but also offers a streamlined and resource-efficient approach, paving the way for more effective and accessible applications in the realm of deep generative models.