Adaptive Collaborative Learning with Data Silos

The School of EECS is hosting the following progress review 2 seminar:

Adaptive Collaborative Learning with Data Silos

Speaker: Achmad Ginanjar
Host: Prof Xue Li
Chair: Dr Yadan Luo

Abstract: Open world assumption in model development means that a model may not have enough information to effectively handle data that is completely different or out of distribution (OOD). When a model encounters OOD data, it may suffer a significant decrease in performance. Addressing OOD data requires extensive fine-tuning and experimental trials, which in turn require substantial computational resources. Deep learning has been suggested as a solution and has shown significant improvements. Still, it often requires high-specification hardware, particularly GPUs, which may not always be readily available to common users. Additionally, there is a lack of clear guidance for common users on how to select and evaluate OOD data.

This study delves into detection, evaluation, and prediction tasks within the context of OOD on tabular datasets. It demonstrates how common users can identify OOD data from real datasets and provides guidance on evaluating the OOD selection through simple experiments and visualizations. Furthermore, the study introduces tabular continual contrast learning (TCCL), an enhanced technique specifically designed for tabular prediction tasks. TCCL is more efficient compared to other baseline models, making it useful for general machine learning users with computational limitations in dealing with OOD problems. The study also includes a comprehensive comparison with existing approaches, focusing on both accuracy and computational efficiency.

Bio: Achmad Ginanjar received a Master of Data Science degree from Monash University, Australia in 2019. He is currently working toward a Ph.D. degree with the Electrical Engineering Department at the University of Queensland (UQ), Brisbane, QLD, Australia. Before joining UQ, he is an IT professional from The Indonesia Tax Office and a data analytic lecture at Indonesia State College of Accountancy. His current research interests include machine learning, optimalisation, parallel algorithm.