Machine Intellection

Advancing and applying AI for economic and societal impact, A*STAR I²R develops algorithms to extract insight from complex patterns in data in a wide variety of domains, including advanced manufacturing, semiconductor, healthcare, education, urban, finance, and transportation.

Our capabilities enable us to handle diverse data modalities, as well as data and computational constraints that are present in real-world applications.

We delivered AI solutions with great flexibility, precision, and automation to several public agencies and many industry collaborators, including more than 10 joint labs and 150 industry projects. We constantly push the the boundary of cutting-edge AI research on the global stage, with a strong track record of highly cited publications in top AI conferences and journals, as well as wins in international AI competitions.

A*STAR I²R focus on handling the diversity of data types, as well as the data and computational constraints in real-world problems with deep capabilities in core areas of analytics and AI as shown below.

We develop novel deep learning algorithms that can obtain good predictive performance using only 1-10% of the labels as compared to current approaches. We build upon traditional approaches such as semi-supervised and unsupervised learning, methods to incorporate external knowledge, transfer learning, and active learning, revisiting and updating them in the context of modern deep learning.

Our resource-efficient learning research is focused on improving the efficiency of neural networks through:
1) hardware software co-design of neural networks, 2) neural network compression via quantization and knowledge distillation, and 3) neural network pruning. Our work focuses on efficient inference to support edge deployments of neural networks on inference accelerators, including our own.

We develop methods to analyse sensor data from different internet-of-things (IoT) applications for system condition monitoring, effective decision making and smart system control. Our work also involves scaling machine learning and deep learning techniques to work on smaller devices such as mobile phones and sensors, to realise benefits of computing at the edge.

We specialise in managing real-time streaming data that has spatial temporal attributes, i.e. has time & location tagged to each event e.g., GPS, EZ-Link, Taxi, Telco data. This capability is essential for urban mobility and logistics applications. The varied analyses we cover include movement pattern mining, outlier detection, and trajectory search, amongst others.

We utilise deep learning techniques to analyse and understand three-dimensional (3D) data like point clouds and volumetric data. By leveraging cutting-edge AI methodologies, the objective is to build essential 3D deep learning competencies, i.e., developing sample-efficient 3D, multi-modality 3D, robust 3D, and resource-efficient 3D, to tackle the pain points in current 3D deep learning research and adoption, such that make 3D deep learning more practical for real-world applications.

We develop methods that analyse users' preferences, habits and other behavioral patterns from users' past activities, and then use the learned information to predict users' future move or to provide personalised services. We not only develop machine learning solutions for behavioral analytics applications, but also build AI tools that can produce good machine learning solutions for these applications with much less human efforts (improved productivity) and with less errors (improved quality).

We extract deep insights from unstructured text data from web, social media and documents. We adopt a semantics-aided approach to text analytics whereby we analyse the context and semantics of text together with domain knowledge, using NLP, Deep neural network, Machine learning, Knowledge graphs and Network structure analysis. For instance, we developed Bert-based text classification and keyword extraction tools and Q&A system incorporating domain knowledge, for KPMG.

We develop innovative, robust and flexible optimisation models to aid decision making in complex systems such as modern factory, airlines and healthcare. We combine the strength of both classical optimisation methods and emerging machine learning and artificial intelligence approaches for more accurate parameter estimation, better quantification over uncertainty etc. We also develop online and continual learning approaches for addressing the needs of drafting data characteristics in evolving systems.

We develop model-agnostic interpretability tools to improve our understanding of complex deep learning models. This would facilitate the identification of model overfitting and bias, improve performance and generalisation, and enhance the transparency and trustworthiness for legal and ethical reasons.