Towards Efficient Knowledge Reuse from the Perspectives of Data and Models

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

Towards Efficient Knowledge Reuse from the Perspectives of Data and Models

Speaker: Yuxia Fu

Abstract: The rapid growth of deep learning has been accompanied by an increasing scale of models and datasets, resulting in impressive improvements across many tasks. However, reusing knowledge across domains remains inefficient and often requires extensive computational and data resources for retraining or fine-tuning. This research aims to improve the efficiency of knowledge reuse in deep learning by addressing the problem from two complementary perspectives: data and model.

From the data perspective, we investigate how to retain essential task-relevant information while significantly reducing the dataset size. We propose a soft label compression-centric dataset condensation method that generates compact, high-density condensed datasets, enabling effective model training with minimal data. From the model perspective, we first aim to investigate whether a one-for-all vision-language model (VLM) can achieve sufficient generalization capability across diverse tasks and domains. Without loss of generality, we construct a box-referring VQA dataset in the autonomous driving domain to evaluate the transferability of both general-purpose and domain-specific VLMs beyond their original training distributions. Our findings show that these models struggle to transfer knowledge effectively without fine-tuning, indicating challenges in achieving efficient knowledge reuse. In future research, we will further explore knowledge reuse from the model perspective, focusing on merging multiple pretrained models, potentially originating from different tasks or domains, into a unified and efficient representation.

Bio: Yuxia Fu is a PhD student in the School of Electrical Engineering and Computer Science at the University of Queensland. She earned her Master of Data Science degrees from UQ. Her research centres on dataset distillation and model merging, under the supervision of Dr Yadan Luo and Professor Helen Huang.