Existing action recognition models are designed by following the typical single-dataset training-testing paradigm which cannot generalise well and inevitably suffer from severe accuracy drop issue, where these recognition models are deployed to another dataset with different distributions.
Honours and Minor Thesis projects
Cryptographic applications require a careful implementation to avoid side-channel attacks that reveal secret information to an attacker (e.g. via run-time measurements). In particular, for floating point arithmetic it is known that the timing of some basic arithmetic operations and functions on some CPUs depends on the input values , and thus the timing may leak secret information when the input contains secret values. Constant-time implementation tries to mitigate such run-time timing leakage on typical devices.
This is one of our CSIRO Next Generation AI Graduates projects:
We have an existing product that focus on mid-market service or product delivery companies. The product is offering a community solution that offers support around a specific product or service. We want to increase its capabilities to offer smart content moderation and smart responses.
Traditional recommendation generation approaches mostly need to either collect and use expert rules or train learning models to generate recommendations. Generative AI tools show high potential to facilitate this process as their model has been trained on a large volume of scientific studies. They can generate content in a specific sentiment or mirror one’s lexicon by following instructions from the user. This feature contributes to improving the user experience, particularly when it is used for seeking health-related recommendations.…
Politeness plays an important part in facilitating people's daily communication. There have been quite some studies developed to demonstrate that polite communication is beneficial for instructors to build solidarity and rapport with students, which, in turn, facilitate students to learn and gain better performance. However, it remains largely unexplored to what extent instructors and teachers provide polite feedback to students in educational discussion forums.
Following the success of the Human Genome Project, the entire scientific community witnessed a large data explosion in genomics, which was also aided by advances in molecular biology technologies such as next-generation sequencing. These high-throughput technologies enable comprehensive molecular profiling of cancer cell lines, including gene expression. Regardless of the use of gene-based assays, they provide abundant genomic information for identifying participating genes (biomarkers) that contribute to the chemoresistance process in cancer cells.…
The mobility of people and actions they perform provides valuable insights into the management and optimization of the inbuilt environment. Camera-based solutions are simple, fast and cost-effective but might cause privacy concerns. In this project we aim to develop camera-based solutions using human skeletons and artificial intelligence models or algorithms to preserve privacy while being able to gain insights into the human activities within the inbuilt environment.
The need: Early detection and diagnosis of eye conditions is critically important as many diseases, including diabetic retinopathy, glaucoma, and age-related macular degeneration, often show minimal or even no symptoms. Glaucoma is called the "silent thief of sight" since it progressively damages the eyes without any noticeable signs.
This project relates to the visualisation of the source of data used in scientific experiments, and their results. The visualisation focus is graphs.
Trust in the results of scientific experiments and scientific modelling relies on knowing how they have been derived – that is, the ‘scientific workflow’ that led to their production. Being able to reproduce the scientific workflow that led to such results is critical in ensuring trust, confidence and transparency .
In our current day and age, there is an exponential growth in multimodal data, especially the transition of social media from text-based communications to video formats which can be observed with the rise of TikTok, Youtube, and Instagram Reels. This shift requires a shift in how we analyze multimodal data as we will have to move away from traditional text sentiment analysis such as TextCNN.