From Single-Modal to Multi-Modal: Towards Efficient AI in Dataset Distillation

The School of EECS is hosting the following HDR Progress Review 1 Confirmation Seminar:

From Single-Modal to Multi-Modal: Towards Efficient AI in Dataset Distillation

Speaker: Bowen Yuan

Abstract: As AI systems continue to scale, data efficiency has emerged as a critical bottleneck limiting practical deployment and accessibility. Dataset distillation serves as a promising approach to synthesizing compact, information-dense datasets that preserve the essential learning characteristics of much larger datasets while dramatically reducing computational and storage requirements.

We propose a novel color-oriented redundancy reduction framework that addresses overlooked inefficiencies in the color space of distilled datasets. By introducing an efficient palette network that transforms 8-bit images into fewer bits while preserving essential discriminative features, our framework effectively quantizes distilled images with less storage budget. Apart from single-modal dataset distillation, a cross-disciplinary study serves as a strategic benchmarking for understanding the challenges of extending dataset distillation to multi-modal settings, including preserving cross-modal correspondences and domain-specific reasoning capabilities.

Bio: Bowen Yuan is currently a PhD student in the School of Electrical Engineering and Computer Science (EECS) at the University of Queensland, where he completed his undergraduate and Honours degrees in Computer Science. His research interests center on efficient AI, dataset distillation and vision-language models, supervised by Professor Helen Huang and Dr. Zijian Wang.