Latest events

AIML Guest Presentation: Optimisation-centric Generalisations of Bayesian Inference

Dana Rawls — Thu, 27 Jun 2024 13:11:43 +0930

Dr Knoblauch summarises a recent line of research and advocate for an optimization-centric generalisation of Bayesian inference. The main thrust of this argument relies on identifying the tension between the assumptions motivating the Bayesian posterior and the realities of modern Bayesian Machine Learning. Our generalisation is a useful conceptual device, but also has methodological merit: it can address various challenges that arise when the standard Bayesian paradigm is deployed in Machine Learning—including robustness to model misspecification, robustness to poorly chosen priors, and inference in intractable models

AIML Special Presentation: Beyond Sight: Robots Mastering Social and Physical Awareness

Dana Rawls — Thu, 27 Jun 2024 13:05:48 +0930

In the rapidly advancing field of robotics, understanding both social and physical dynamics is crucial for seamlessly integrating robots into dynamic human-centric spaces. Operating effectively in such environments requires a robust visual perception system capable of comprehending physical scenes while anticipating and understanding nuanced human social behaviours.

AIML summer research student presentations

Dana Rawls — Thu, 27 Jun 2024 12:58:09 +0930

Our talented summer research students presented their cutting-edge machine learning projects, from medical breakthroughs to AI advancements. They provided an exciting opportunity to witness the future of technology unfold.

AIML Research Seminar: Computational Algorithms for Human Behaviour Analysis - From Research Endeavor to Industry Relevance

Dana Rawls — Thu, 27 Jun 2024 12:53:22 +0930

Minh provided an overview of his research endeavours and interests, aiming to spark discussions about potential collaborative ventures.

AIML Research Seminar: Anatomically Aware Brain MRI Segmentation of the Cerebral Vasculature

Dana Rawls — Thu, 27 Jun 2024 12:22:28 +0930

As humans, we learn to know what is abnormal by establishing an understanding of what the ‘norm’ actually is. Medical tasks require specific knowledge (e.g. radiologists need to be highly trained), and much is known about healthy, ‘normal’ brain anatomy- documented by anatomical atlases and brain models. This knowledge can then be applied to identify abnormal brain structure from medical images, to identify pathology.

AIML Special Guest Presentation: Exploring the Earth and space with micro-satellites

Dana Rawls — Thu, 27 Jun 2024 12:11:13 +0930

In the past, space utilisation has required long preparation times and high budgets, but the 50 kg optical satellites we have developed and have successfully launched six times can be fabricated in a few years with a budget of 3-7 M USD and can be launched for 1-2 M USD, e.g. using carpool launch opportunities. This means that universities and relatively small companies can own and operate their own satellites, even if they are not national space organisations.

AIML Research Seminar: Improving Efficiency of Foundation Models

Dana Rawls — Thu, 27 Jun 2024 12:06:45 +0930

Large deep learning models or foundation models such as chatGPT or GPT-4 have been the key factor in driving the recent new wave of AI breakthrough, resulting in huge social and economic impacts. However, even GPT-3 (the predecessor of ChatGPT) was trained on half a trillion words and equipped with 175 billion parameters, which required huge computing resource and energy consumption.

AIML Special Presentation: Trustworthy AI

Dana Rawls — Thu, 27 Jun 2024 11:58:54 +0930

Out-of-distribution (OOD) detection aims to let a well-trained classifier tell what it does NOT know, instead of wrongly recognising an unknown object as a known one. For example, for a well-trained flower recognition model, we want it to tell users “I don’t know” when users show a car image to it, instead of telling users that it is a kind of flower. In this talk, Dr Liu presented one advance in OOD detection theory and two recent OOD scores: one based on in-distribution prior and the other based on the pre-trained vision-language model CLIP.

AIML Research Seminar: Crater-based pose estimation for cislunar located spacecraft

Dana Rawls — Thu, 27 Jun 2024 11:48:12 +0930

In this presentation, Sofia demonstrated how we can estimate cislunar located spacecraft through a crater-based pose estimation pipeline. Her research focus lies in the final pose estimation step of the pipeline, where she has conclusively addressed the weaknesses of current pose estimation methods by developing a robust perspective-n-crater algorithm. As part of this research, Sofia interned for three months at Japan’s National Institute of Information and Communication Technology and discussed what it was like working and living in Japan.

AIML Special Presentation: Active Learning with Deep Neural Networks

Dana Rawls — Thu, 27 Jun 2024 11:38:31 +0930

In Active Learning the system gives select data to an expert to annotate, which is costly and should be minimised. Prof Butine proposes the first general Bayesian method to work well in this context, and his experiments show it is the only method consistently better than random.