External Research Grants

In this proposal, we consider a real-world setting that a large model trainer like OpenAI already holds large-scale training data, but it continuously needs more fresh data to update the model or produce specific downstream tasks. Such data sharing mechanism benefits both the model trainer and data provider who should be paid for his contribution. This motivates us to achieve the following four key objectives: (1) Private pre-processing of training data sharing for large model; (2) Secure and private regulation compliance inspection on shared training data with verification and proof; (3) Privacy-preserving dynamic fine-grained training data sharing; and (4) Privacy-preserving inference on large models.

This research project, developed under the CSIRO - A*STAR Research-Industry (2+2) Partnership Program, aims to develop sustainable and responsible AI technologies, with a particular focus on large language models (LLMs). The project's objective is twofold: enhancing environmental sustainability and ensuring compliance with governance standards.

The project conducts research on new conversational AI technologies that understand a user’s mental health conditions and enable a principled strategy to counsel the user. The research will focus on incorporating user personalisation and counselling strategies into the AI models. At the end of project, we hope to create a conversational AI framework that can automate mental health counselling and evaluate its performance.

In the dynamic field of wearable health technology, our proposed research aims to revolutionise how we monitor our health using devices such as smartwatches and earbuds. These devices frequently employ photoplethysmography (PPG), a noninvasive technique that monitors changes in blood volume under the skin, providing valuable insights into cardiovascular health. However, real-world challenges, such as inaccuracies during physical activities and the impact of diverse body postures, impede the realisation of the full potential of PPG technology with respect to these wearable devices.

Our research focuses on a breakthrough concept: incorporating contact pressure (CP) into PPG measurements to address the aforementioned challenges. By analyzing the tightness of wearable devices against the skin, we aim to obtain valuable insights that can help reconstruct high-quality PPG data from the noisy PPG data. Our first contribution is the development of a wearable prototype capable of concurrently measuring CP and PPG. Using this prototype, we will develop intelligent algorithms to mitigate the effects of physical activities and body postures. Finally, we will optimise our energy efficiency and real-time processing methodologies, ensuring prolonged battery life for wearable devices.

We believe that these innovations can substantially improve the accuracy and reliability of health data obtained from wearables, thereby unlocking new capabilities of PPG in health monitoring. The success of this research has the potential to stimulate market growth by establishing a new standard for accuracy and capabilities in wearable devices. Considering the global aging population, our research will considerably impact elderly care, particularly cardiovascular diseases. By improving the reliability of wearable devices, our research can promote an active lifestyle and contribute to overall well-being among the general population.

Artificial intelligence (AI)-driven personal assistants like Apple’s Siri and Amazon’s Alexa have gained significant popularity due to their user-friendly natural-language interfaces that facilitate interaction with the Internet. These agents, embedded in various devices, assist users with tasks such as retrieving information, controlling home automation, and managing calendars. However, current assistants lack awareness of the physical environment; they cannot connect directly to sensors to interpret a user's physical environment, limiting their ability to respond to contextual queries, such as identifying the manufacturer of a toaster pointed out by a user.

Recent advances in AI, particularly with models like GPT-4, have enhanced capabilities for machine-based reasoning that combine visual, verbal, and gestural inputs. Projects like Google’s Project Astra are exploring these multi-modal assistants, which could revolutionize human-computer interactions. Additionally, the integration of sophisticated sensors in mobile and IoT devices, such as LIDAR technology in smartphones, opens the door for new situated agents that can leverage spatial awareness and multimodal reasoning to deliver contextually relevant information.

Developing these situated agents presents significant challenges, notably in energy consumption and computational complexity. High-power sensors like LIDAR are unsuitable for continuous use, and current AI models are too large to run efficiently on mobile devices. To address these issues, the project proposes innovations in energy-efficient sensing techniques and optimization of deep neural networks (DNNs) to minimize latency and energy use. Through innovations in triggering novel spatial sensors on-demand, processing relevant multi-modal sensor data selectively and by exploiting the cached results of recent reasoning actions, these advancements aim to enable responsive and pervasive situated agents that can interact meaningfully with users. The project also seeks to embed these innovations in a prototype conversational agent that can serve as a real-time tutor, responding intelligently to a learner’s queries and enhancing educational experiences.

In recent years, the rapid advancement of Large Language Models (LLMs) has driven significant progress in Generative Artificial Intelligence (GAI). By emulating human language capabilities, GAI has unlocked a multitude of applications, ranging from chatbots and virtual assistants to translation services and content-generation tools. However, GAI has evolved into a double-edged sword, giving rise to ethical concerns surrounding transparency, privacy protection, misinformation, bias and fairness, job displacement, and environmental impact.

In this project, the team aims to quantify and validate firms' ethical awareness of GAI and employ econometrics models and quasi-experiments to comprehend its determinants and document consequences. Through these efforts, the project makes a distinctive contribution to the promising trajectory of developing responsible and ethical GAI systems in business, ultimately fostering a sustainable society.

(This is additional funding to SMU for the existing research project.) Digital wellbeing has arisen in public, governmental and policy discourse as a key measure of a person’s wellbeing through a healthy use of technology. This project aims to identify and measure digital wellbeing for digital readiness, inclusion and safety. 

A trusted digital identity is an essential component for securely and conveniently accessing services and authorizing transactions in cyberspace. With the rapid development of decentralized technologies and applications, such as distributed ledgers, Web3, and decentralized finance, there is an urging demand for decentralized digital identities (DID), also known as self-sovereign identities, which empower end users to create, own and govern their digital identities and assets in an autonomous, reliable, and privacy-preserving manner. The overarching goal of this project is to develop and implement a trusted, versatile, reliable and user centric DID framework covering a complete DID lifecycle. Specifically, the project aims to investigate novel techniques for enabling key components and features that are either missing or inadequately addressed in the existing DID proposals. This research is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative.

This project focuses on “human capital development” that involves, as one of its core pillars, the ability to use AI to improve learning. Through this project, the SMU Research team will work with the MIT team to build up key AI based models for supporting interactive learning by supporting natural multi-modal question answering associated with learning tasks. The SMU team will also develop initial prototypes that embed such interactive learning in online learning platforms.

This project focuses on “immersive workplaces” that involve extensive and interactive co-working by humans and AI-enhanced agents/robots. The research will investigate techniques to combine the latent cognitive state of a human worker, with explicit instructions issued by humans using a natural mix of visual, verbal and gestural cues, to build powerful new capabilities for human-robot co-working in immersive workplaces and industrial settings. The SMU Research team will develop optimized AI models, as well as prototype pervasive systems, that allow a variety of robotic agents to disambiguate, comprehend and respond to commands issued by human workers.