Workshop 1: Foundations for Robust AI: Self-Supervised Learning

Time: Nov 28th, 9:00 - 12:30
Organizers:
Dr. Saimunur Rahman, CSIRO, Brisbane, Australia
Dr. David Hall, CSIRO, Brisbane, Australia
Dr. Stephen Hausler, CSIRO, Brisbane, Australia
Dr. Peyman Moghadam, CSIRO, Brisbane, Australia

This workshop will bring together the Australasian community in self-supervised learning (SSL), one of the most successful and rapidly expanding fields of artificial intelligence research in recent times. SSL does not learn to overfit a solution for a specific task but instead seeks to learn highly generalisable and robust representations of the world through unlabelled data, which can then be applied across a multitude of tasks. As a critical driving force behind popular applications such as ChatGPT, the potential of SSL cannot be underestimated, and the need for local expertise will be critical for the region to avoid dependence on models trained overseas that may not consider local needs. This workshop will focus on how SSL can serve as the foundational cornerstone for practical AI applications that directly benefit Australasia and provide a forum for discussion on how the community should move forward in its research.

Workshop 2: Federated Learning in Australasia: When FL meets Foundation Models

Time: Nov 28th, 9:00 - 12:30
Organizers:
A/Prof. Guodong Long, University of Technology Sydney, Australia
Asst Prof. Han Yu, Nanyang Technological University, Singapore
Dr. Tao Shen, University of Technology Sydney, Australia

The aim of this workshop is to bring together academic researchers and industry professionals within the Australasian region to discuss, explore, and address the potentials and challenges of incorporating federated learning (FL) techniques with foundation models. This workshop serves as a platform to share insights and innovative solutions in developing robust and secure AI systems that handle the unique issues brought about by foundation models, such as the high number of learnable parameters leading to edge computing and communication challenges, privacy and security concerns due to learning from a vast amount of data, and limited personalization possibilities. Through fostering collaboration and exchanging ideas, the workshop seeks to push the boundaries of current understanding and applications of federated learning in the context of foundation models, thereby propelling the advancement of AI technology in Australasia.

Workshop 3: Artificial Intelligence Enabled Trustworthy Recommendations

Time: Nov 28th, 9:00 - 12:30
Organizers:
Dr. Shoujin Wang, University of Technology Sydney, Australia
Dr. Rocky Tong Chen, The University of Queensland, Australia
A/Prof. Hongzhi Yin, The University of Queensland, Australia
Prof. Lina Yao, CSIRO’s Data61, The University of New South Wales, Australia
Prof. Fang Chen, University of Technology Sydney, Australia
Prof. Guiqiong Xu, Shanghai University, China

The aim of this workshop is to engage with active researchers from both academia and industry communities and deliver the state-of-the-art research insights into trustworthy recommender systems (RSs) and the social impact of recommendation techniques. With the advancement of data science and AI, more and more powerful RS models have been developed to provide very accurate recommendation results to well match users’ preferences. However, during the past decades, most of the existing work only focus on the improvement of recommendation accuracy while ignoring some other very significant aspects, such as the trustworthiness and social impact of an RS. Here the trustworthiness can be specific to multiple aspects, including the robustness, explainability and security/privacy-preserving of the RS models, the explainability and fairness of the recommendation results as well responsible recommendation results. The social impact refers to the various influence of an RS on both individuals and groups, both internally and externally, in our society and the influence may be positive or negative. However, on one hand, we are in the era of information explosion and the cyberspace where the recommendation happens is becoming more and more complex, uncertain and dynamic, and is filled with more and more noise, attacks, and bias. The accuracyoriented RSs may not be compatible with such complicated cyberspace, which triggers the emerging need of the consideration of trustworthiness of RSs. On the other hand, RSs generally aims to benefit the end users and the society by providing recommendation service on nearly every corner of our daily live. However, sometime, RSs may generate negative impact to the users and the society. For example, an RS may limit a user’s interaction scope on items by tending to only recommending those limited popular items to the user, and thus leads to information echo chamber. Therefore, it is the right time to draw the attention of both researchers and engineers in data science to the trustworthiness and the social impact of RSs to lead RSs towards social good. In addition to accuracy, we should focus more on the social influence of an RS, so that we can provide responsible recommendation services that contribute to a sustainable and harmonious digital and psychical world.

Workshop 4: Machine Learning for Data-driven Optimization

Time: Nov 28th, 9:00 - 12:30
Organizers:
Xilu Wang, Bielefeld University, Germany
Xiangyu Wang, Bielefeld University, Germany
Shiqing Liu, Bielefeld University, Germany
Prof Yaochu Jin, Bielefeld University, Germany

Optimization is pervasive in scientific and industrial fields, such as artificial intelligence, data mining, bioinformatics, software engineering, scheduling, manufacturing, and economics. The objective functions in practical problems usually exhibit complex functional characteristics, such as multi-modality, large scale, sparsity, constraints, dynamic environment, expensive evaluation, and black box, posing challenges for traditional exact solvers. For example, the evaluation of the objective or constraint functions may involve costly physical experiments or intensive computer simulations, thereby limiting the amount of available data. Moreover, in some cases, the function evaluations can only be calculated on the basis of a large amount of data. Consequently, data-driven optimization has emerged as a powerful solution for these problems to reduce the computational cost and reshape the way we tackle complicated optimization problems, where typically machine learning techniques and optimizers are combined.

Recently, data-driven optimization has attracted increasing attention in learning-based combinatorial optimization. Combinatorial optimization tasks encompass a broad range of domains, including logistics, scheduling, transportation, and manufacturing. Traditional methods based on exact solvers and heuristic-based approaches suffer from huge computational complexity from scratch when tackling large-scale instances. With the development of graph representation learning, neural combinatorial optimization has emerged as a promising paradigm for solving classical combinatorial optimization problems. It empowers neural networks to not only handle the inherent complexity of combinatorial optimization challenges but also to learn powerful information embedded in graph representations. However, existing work on data-driven neural combinatorial optimization still has some limitations such as the dependency on high-quality data, the complexity of training and the restricted transferability among different tasks.