This project shall (i) enhance the generalizability of hierarchical multi-agent learning and control framework for heterogeneous agents in a range of scenarios and (ii) develop algorithms to analyse and explain the learned behaviour models at the various levels.
In this project, the problem being addressed is to provide an accurate, low-latency, minimal maintenance indoor localisation solution to locate organisational resources. Our goal is to achieve this without any form of Wi-Fi fingerprinting, without any knowledge of the location of the Wi-Fi Access Points (AP), and without the availability of any maps of the indoor spaces being used. We plan to achieve this by leveraging the new 802.11mc Wi-Fi Fine-Time Measurement standard in pure 1-sided mode that allows time-of-flight measurements to be made between a client device and any AP. These measurements will then be used with inertial data to jointly optimise both the location of the device and the location of the APs.
This project aims to leverage the pre-existing but currently underutilized educational technological infrastructure in secondary schools in Bangladesh in partnership with the Government, and to evaluate the effectiveness of the proposed interventions on learning outcomes and other socio-economic outcomes such as child labour and early marriage incidents. The proposed interventions include conducting training sessions for headmasters and teachers on ICT integration and blended-learning strategies for digital remedial education, and the provision of offline educational videos focusing on English and Mathematics to students.
SMU has been awarded with additional funding the extension of the Singapore International Dispute Resolution Academy (SIDRA). A platform for thought leadership in international dispute resolution theory, practice and policy, SIDRA leads the way through projects, publications and events that promote dynamic and inclusive conversations on how to constructively engage with and resolve differences and disputes at global, regional and national levels. In particular, SIDRA differentiates itself through its focus on applied research that has practical impact on industry.
The proposed project studies how to automate fact-checking (FC) based on neural language models. FC is the investigative process of verifying and reporting the accuracy of claims to help people make decisions based on facts rather than misinformation. Our proposed project, named FrankLM, targets Fact-check and report automation via neural knowledge Language Modeling to enable the applicability of LLMs for FC. We aim to improve task accuracy by 20% for claim verification and explanation generation, improve task accuracy by 15% for reasoning, and achieve over 90% of human performance in report generation. FrankLM will benefit FC and improve the accuracy, explainability, and trustworthiness of AI systems, and it will open new opportunities to apply them in various sectors including media, healthcare, finance, and education.
This research/project is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG3-RP-2024-035).
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.
