Program

Abstract: Deep generative models have made tremendous progress in modeling complex data, often exhibiting generation quality that surpasses a typical human's ability to discern the authenticity of samples. Undeniably, a key driver of this success is enabled by the massive amounts of web-scale data consumed by these models. Due to these models' striking performance and ease of availability, the web will inevitably be increasingly populated with synthetic content. Such a fact directly implies that future iterations of generative models will be trained on both clean and artificially generated data from past models. In addition, in practice, synthetic data is often subject to human feedback and curated by users before being used and uploaded online. For instance, many interfaces of popular text-to-image generative models, such as Stable Diffusion or Midjourney, produce several variations of an image for a given query which can eventually be curated by the users. In this talk we will discuss the impact of training generative models on mixed datasets---from classical training on real data to self-consuming generative models trained on purely synthetic curated data.

Abstract: In this presentation, we will discuss artificial intelligence technology's effect on our society based on Science, Technology, and Society and Human-Computer Interaction. Generative AI is currently becoming a social infrastructure. What kind of effect will the technology have on society, unlike conventional technologies? Especially for ethical aspects, can AI and people interact with each other and create new forms of ethical values? We will discuss these questions while introducing our ongoing research, Human/Stakeholder-in-the-loop AI fairness.

Abstract: Generating images from textual descriptions requires the generative model to make implicit assumptions about the output scene that are not explicitly instructed in the input prompt. These assumptions can reinforce unfair stereotypes related to gender, race, or socioeconomic status. However, measuring and quantifying these social biases in generated images is a big challenge. In this talk, we will explore methods for measuring gender bias in text-to-image models, particularly Stable Diffusion, and discuss how the generated images, when used to train future computer vision models, affect bias in downstream tasks.

Abstract: Recent advances in multimodal generative models have transformed our ability to understand and generate mages, text, and other data forms, leading to groundbreaking applications such as GPT-4o, Gemini, and Sora. Despite their potential, these models are fraught with trustworthiness challenges. For example, they are susceptible to generating harmful content, vulnerable to adversarial attacks, and pose significant privacy risks. This talk will explore the methodologies for identifying these risks and discuss strategies to mitigate them. Additionally, I will introduce MultiTrust, a comprehensive benchmark designed to evaluate the trustworthiness of multimodal generative AI systems.