Mitigating Biases in Self-consuming Generative Models
Virtual: https://events.vtools.ieee.org/m/486945In this talk, Dr. Ali Siahkoohi highlights the risks of the current industrial AI practices involving training large-scale generative models on vast amounts of data scraped from the internet. This process unwittingly leads to training newer models on increasing amounts of AI-synthesized data that is rapidly proliferating online, a phenomenon Dr. Siahkoohi refers to as ``model autophagy'' (self-consuming models). He shows that without a sufficient influx of fresh, real data at each stage of an autophagous loop, future generative models will inevitably suffer a decline in either quality (precision) or diversity (recall). To mitigate this issue and inspired by fixed-point optimization, a penalty to the loss function of generative models is introduced that minimizes discrepancies between the model's weights when trained on real versus synthetic data. Since computing this penalty would require training a new generative model at each iteration, a permutation-invariant hypernetwork is proposed to make evaluating the penalty tractable by dynamically mapping data batches to model weights. This ensures scalability and seamless integration of the penalty term into existing generative modeling paradigms, mitigating biases associated with model autophagy. Additionally, this penalty improves the representation of minority classes in imbalanced datasets, which is a key step toward enhancing fairness in generative models. Speaker(s): Ali Siahkoohi Agenda: - Invited talk from Dr. Ali Siahkoohi, Assistant Professor in University of Central Florida's Computer Science Department. - Q/A Session Virtual: https://events.vtools.ieee.org/m/486945