Seminar Artificial Intelligence - Dr. S. Suh, German Research Center for Artificial Intelligence

Title: Deep Generative Models with Industrial Applications

Abstract:

Deep learning has achieved significant improvements in a variety of tasks in computer vision applications with open image datasets that have a large amount of data. However, in real-world industrial settings, acquiring large datasets remains a challenge, particularly in emerging areas of deep learning. Recently, generative models, ranging from ChatGPT for natural language understanding to applications in vision and audio domains, have become pivotal in shaping the landscape of artificial intelligence. The potential of generative models as a means of data generation and domain adaptation has been studied. This talk will provide an introduction to the core concepts of generative adversarial networks (GANs) and explore different variants of GAN architectures.

Drawing from my experiences, I will introduce how generative models can address the limitations of training data and improve the performance of neural networks. We handle various industrial applications of generative models, including addressing data imbalance issues, virtual try-ons for online shopping, and wearable sensor-based human activity recognition. Through these examples, we illustrate how deep generative models can effectively improve the performance of deep learning applications in industrial contexts.

In addition to leveraging generative models to generate diverse modalities of data, such as images, depth, language, and time-series data, representation learning with generative models will also be introduced. The discussion extends to future directions on how generative models generate embedding features effectively and generalize representations across different modalities.