A recommender system presents a personalized experience to each user. One perennial issue affecting current recommendation technologies is the sparsity of data related to user preferences. The overall objective of this proposed research is to address this sparsity problem by a combination of approaches that together enable lifelong learning for recommender systems. This is done by allowing the recommendation model to evolve over time to new users and items and, to transfer over to new product categories. In addition, the proposed recommendation model would have the ability to cross from a source platform that accumulates longer-term preferences to a target platform that seeks to integrate short-term signals and reinforcement learning. This provides a system that is able to learn from longer-term preferences and provide the necessary flexibility for cross platform applications.
This research is an ideation/proof of concept project to develop an interactive Web application, called Slide++, which allows students to self-explore additional content related to their courses, while still being directed by the lesson materials provided by an instructor. Importantly, its primary feature is to provide content augmentation for every slide in the form of learning resources relevant to the slide being viewed. These resources can be of various modalities, including Web pages, videos, or questions and answers (Q&A’s).
This project aims to build an Automatic Programming Coaching system that is based on a combination of AI and software engineering techniques to support students practice coding via formative feedback generation.
Rare events, also known as “black swans”, in financial time series can be seen as sporadic and drastic jumps in financial assets returns. Accurate and timely estimates of future risk associated with rare events are of great importance for finance practitioners, policymakers, and regulators. The research team will leverage the most recent developments in quantum-enhanced Monte-Carlo sampling, stochastic modelling and dimensional reduction to design a set of quantum algorithms for rare event estimation that:
1. Enhance the accuracy in estimating the probability of specific rare events – we anticipate a quadratic scaling improvement, where doubling the iterations for the quantum algorithm will result in an accuracy improvement equivalent to a quadrupling of iterations in its classical counterpart.
2. Reduce systematic error caused by dimensional reduction – when constrained to storing the same amount of past data (e.g., macroeconomic indicators), our quantum model can give more accurate rare event predictions than classical counterparts.
This is a project under the AI Singapore 100 Experiments Programme. It tackles the research challenge of generating good logistics plans and schedules for parcel delivery using AI. uParcel faces the challenges that the number of deliveries daily are in the thousands and the number of drivers delivering a day is in hundreds, which makes it very challenging to match jobs to drivers and encourage job acceptance. Using reinforcement learning coupled with large-scale optimization methods, the research team will develop route optimization, dynamic recommendation, and logistics marketplace matching algorithms for improving operational efficiency. This will also greatly improve city logistics by reducing trips and congestion.
With the advent of e-commerce, users are presented with numerous alternatives to satisfy their everyday needs. Choosing from the available options generally entails the consideration of multiple, often conflicting aspects, the tradeoff among which is assessed differently by different users.
This project proposes PERFLEXO, a new methodology for multi-objective querying centred around three hard requirements, i.e., personalization, flexibility in the preference input, and output-size control. Past approaches have considered these requirements individually, but no existing work satisfies all three of them. On the technical side, the main contributions of the project will centre on PERFLEXO’s ability to process large option-sets (i.e., scalability) and produce shortlists in reasonable time (i.e., responsiveness).
AI models trained offline rely on the accessibility of all classes in training data. When they are updated online to learn new incoming data, they often bias to the patterns of new classes, and thus forget old ones. The problem is known as catastrophic forgetting. This project aims to tackle this issue by task-specific data augmentation. The augmentation for old classes is achieved by distilling from new or open-set data that contain the knowledge of old classes, e.g., shared contexts and sub-parts.
(This is a 6-month extension of the research collaboration with Fujitsu Ltd.) Under the Fujitsu-SMU Urban Computing and Engineering (UNiCEN) Corp Lab, SMU has undertaken the Digital Platform Experimentation (DigiPlex) project with Fujitsu. The project was carried out using the Digital Annealer (DA), a quantum inspired-technology inspired by Fujitsu. Through the DigiPlex project, certain challenges in solving constrained optimization problems using such technology, and promising methods on tuning of the underlying model parameters to improve run time performance, have been identified. This project aims at developing hyper parameter tuning methodology, machine learning techniques, operations research algorithms, and software tools to enhance quantum-inspired techniques for solving large scale real-world combinatorial optimization problems.
The world is experiencing a rapid transition towards a digital society. Although huge number of Internet of Things (IoT) devices are being deployed to provide accurate and real-time sensing and observation of the environment, security and privacy concerns are becoming one of the major barriers for large scale adoption and deployment of IoT. To that end, this project aims to provide IoT devices with privacy-aware authentication and flexible authorisation capabilities to build trust in IoT.
With the widespread adoption of drones in civilian, business, and government applications nowadays, concerns for breaches of safety, security, and privacy by exploiting drone systems are also rising to the highest national level. Malicious entities have used drones to conduct physical and cyber-attacks such as unauthorized surveillance, drug smuggling, armed use, etc. In this project, the research team aims to develop methods and tools for analysing a list of drones to audit drones for detecting anomalies such as malware, data leak, software bugs that could be exploited to conduct criminal/malicious drone activities. The research team will analyse at least five different drone-related criminal/malicious activities from their collaborator and demonstrate how ADrone can assist Drone forensic analysts with the detection of the root causes of activities.
