Artificial intelligence is shaping the future of humanity across various industries. It is already the main driver of emerging technologies like big data, robotics and IoT, and will continue to act as a technological innovator for the foreseeable future. In this workshop, speakers from A*STAR’s Centre for Frontier AI Research (CFAR) and RIKEN Centre for Advanced Intelligence Project (AIP) will gather to share insights on Machine Learning and AI Technology.

Artificial intelligence is shaping the future of humanity across various industries. It is already the main driver of emerging technologies like big data, robotics and IoT, and will continue to act as a technological innovator for the foreseeable future. In this workshop, speakers from A*STAR’s Centre for Frontier AI Research (CFAR) and RIKEN Centre for Advanced Intelligence Project (AIP) will gather to share insights on Machine Learning and AI Technology.

Being able to say no to abnormal/unknown instances is one key capability that machine learning systems should master in broad real-world application domains, especially for security/safety-critical domains, e.g., detecting and rejecting suspicious financial crimes in banking, handling unknown objects in autonomous driving or medical artificial intelligence (AI) systems, etc. Anomaly and out-of-distribution detection are two lines of research that could achieve this capability. While current approaches have resulted in substantially improved detection accuracy in both research lines, uncovering the unknowns still remains an open problem due to the difficulty in obtaining instances of the unknowns and their unbounded distribution. In this talk, Dr. Pang Guansong from the Singapore Management University (SMU) will review relevant progress and introduce some of his team’s recent work on tackling the challenges in anomaly and out-of-distribution detection.

Adversarial attacks can disrupt artificial intelligence (AI) and machine learning (ML) based system functionalities but also provide significant research opportunities. In this talk, Prof Dipankar Dasgupta from The University of Memphis will cover emerging adversarial machine learning (AML) attacks on systems and the state-of-the-art defense techniques. Prof Dasgupta will first discuss how and where adversarial attacks could happen in an AI/ML model and framework. He will then present the classification of adversarial attacks and their severity and applicability in real-world problems, including the steps to mitigate their effects, before illustrating the role of GAN in adversarial attacks and as a defence strategy.Finally, Prof Dasgupta will also discuss a dual-filtering (DF) strategy that could mitigate adaptive or advanced adversarial manipulations for a wide-range of ML attacks with higher accuracy. The developed DF software could be used as a wrapper to any existing ML-based decision support system to prevent a wide variety of adversarial evasion attacks. The DF framework utilises two set of filters based on positive (input filters) and negative (output filters) verification strategies that could communicate with each other for higher robustness.

Agent-based modelling is an old idea that has seen a resurgence of interest over the past decade. The basic idea of agent-based modelling is to model socio-technical systems at the level of individual decision-makers (agents). Agent-based models make it possible to capture aspects of systems (such as social network structures) that cannot be represented with conventional modelling techniques. In this talk, Prof Michael Wooldridge from the University of Oxford will describe his work towards a robust engineering science of agent-based modelling, focusing on four key issues:

How do we capture agent-based models transparently; How do we populate agent-based models with realistic agent behaviours; How do we calibrate agent-based models;How do we verify such models.Prof Wooldridge will illustrate these key points with lessons learned from the COVID-19 models that were central to the UK’s decision to enter the first lockdown in March 2020.

Deepfakes are audio or visual samples manipulated and/or generated using deep learning-based techniques. In spite of a large number of useful applications of synthetic data, misinformation can be easily spread with deepfakes. In this talk, Prof Abhinav Dhall from Monash University will discuss the issues which may arise due to the nefarious use of deepfakes and provide a brief overview on methods to identify deepfakes. He will be sharing his recent work in deepfakes detection inspired by human implicit and explicit signals. Moving forward with classifying deepfakes, Prof Dhall will also discuss the problem of localisation in temporal dimension.

The recent successes of deep learning methods have motivated their employment in diverse domains, including physical tasks such as predicting the motion of objects or the behaviour of fluids. However, in order to learn properly, deep learning methods require large amounts of data, which can be costly in certain domains. Moreover, they often do not generalise properly to domains not previously seen in the training data. By embedding traditional physics models into deep learning architectures, Dr. Filipe de Avila Belbute-Peres from Carnegie Mellon University will introduce ways to address these issues that allow for more efficient and robust learning.

Artificial Intelligence (AI) is recognised globally as a highly disruptive technology in recent years. Leveraging opportunities in AI could potentially bring about transformation and benefits to human societies and national economies. In this seminar jointly organised by A*STAR’s Centre for Frontier AI Research (CFAR) and the Nanyang Technological University, Singapore (NTU Singapore), we bring together esteemed speakers from IEEE to share insights on future AI trends.

Deep learning has advanced from fully connected architectures to structured models organised into components, e.g., the transformer composed of positional elements, modular architectures divided into slots, and graph neural nets made up of nodes. Inspired by the human brain, components in such models face communication bottlenecks due to restricted connectivity and attention, which may serve as a useful form of inductive bias. In addition, human consciousness works in a stochastic manner, in which different neurons and sub-structures are activated differently upon different internal and external stimuli. In this talk, Dr Liu Dianbo from Mila - Quebec Artificial Intelligence (AI) Institute will illustrate how the bottleneck of communication among modules and input-dependent stochastic activation could help in out-of-distribution generalisation and uncertainty estimation of the deep learning models. 

With conventional digital computing technology reaching its limits, there has been a renaissance in analog computing across a wide range of physical substrates. In this talk, Prof Peter McMahon from Cornell University will introduce the concept of Physical Neural Networks and describe a method that his group has developed to train any complex physical system to perform as a neural network for machine-learning tasks. They have been able to show MNIST handwritten-digit classification experimentally on i) mechanical, ii) electronic and iii) photonic systems, despite the fact that none were initially designed to do machine learning. He will also describe several possible future research directions on Physical Neural Networks, including: i) the potential to create large-scale photonic accelerators for server-side machine learning, ii) smart sensors that pre-process acoustic, microwave or optical signals in their native domain before digitisation,iii) new kinds of quantum neural network that does not require a carefully engineered quantum computer, iv) and generally the prospect to endow analog physical systems with new, unexpected functionality.

Simulation is important for numerous applications in science and engineering, and there has been increasing interest in using machine learning to produce more efficient learned simulators, distill complex dynamics into a differentiable model, or learn simulators from real world data. In the first part of this talk, Dr Kimberly Stachenfeld from DeepMind will describe her group’s approach to using deep learning for learned physics simulators, with particular focus on work that uses learned models for efficient, accurate, stable simulation of fluids and other complex physical dynamics in particle-, mesh-, and grid-based domains. In the latter part of the talk, Dr Stachenfeld will discuss her group’s recent work using learned simulators for inverse design, where they combined their graph-based learned simulators with gradient-based optimisation in order to optimise designs in a variety of complex physics tasks. These include challenges faced when designing objects in 2D and 3D to direct fluids in complex ways, as well as optimising the shape of an airfoil. The group found that the learned models can not only support design optimisation across 100s of timesteps, these models can in some cases, permit designs that lead to dynamics that are apparently quite different from the training data. Dr Stachenfeld will then conclude the talk with a discussion on some of the future directions for learned simulation.

The first multi-objective evolutionary algorithm was published in 1985. However, it was not until the late 1990s that the so-called evolutionary multi-objective optimisation began to gain popularity as a research area. Throughout these 37 years, there have been several important advances in the area, including the development of different families of algorithms, test problems, performance indicators, hybrid methods and real-world applications, among many others. In this talk, Prof Carlos Artemio Coello Coello from the Center for Research and Advanced Studies of the National Polytechnic Institute (CINVESTAV-IPN) will discuss some of these developments, focusing on the notable recent achievements. This will be followed by a critical analysis on the analogy research that has proliferated in recent years in specialised journals and conferences (perhaps as a side effect of the abundance of publications in this area). Despite having a low level of innovation and almost no scientific input, the research is backed by a large number of statistical tables and analyses. Prof Coello Coello will conclude the talk with some of the future research challenges for this area, which is just beginning to mature after 37 years of existence.

The computer-assisted analysis for better interpreting images has been a longstanding issue in the medical imaging field. Recent advances in machine learning, especially in the way of deep learning, have made a big leap to help identify, classify, and quantify patterns in medical images. Generative Adversarial Networks (GAN) have gained a lot of attention in the computer vision community due to their capability of data generation without explicitly modelling the probability density function. In this talk, Dr Yang Guang from Imperial College London will introduce GAN based methods in medical imaging reconstruction, super-resolution and medical data synthesis. Instead of designing complicated deep neural network architectures for downstream clinical tasks, his research will enhance the importance in medical image data harmonisation and general quality controls to facilitate multi-scanner and multi-centre data collection and analysis. 

There has been a growing emergence of disinformation campaigns using photorealistic fake images in mainstream media outlets and photo hoaxes that land in our e-mails. Accurate detection of such sophisticated and real-look-like fake images presents a challenge in preserving social harmony, security, and privacy-related concerns in our society. A range of advanced methods has been introduced in the literature to detect sophisticated fake images that are often generated by powerful generative adversarial network (GAN) based methods. In this talk, Prof Ajay Kumar from The Hong Kong Polytechnic University will highlight the increasing reliance on such popular and sophisticated fake image detection methods available in the scientific community. He will also discuss such fake image detectors and present a scientific pipeline that can compromise the detection performance when the proposed methods are incorporated. The analysis presented during this talk will caution the forensic community on the reliability of widely employed GAN-generated fake image detectors, and the need for significant advances to address such open challenges.

Deep neural networks have achieved tremendous success across several domains, ranging from computer vision, natural language processing, to audio analysis. However, training neural networks to perform well typically requires a large amount of labelled data. In many cases, this requirement for a large, labelled dataset presents a challenge, as the annotation process can be labour-intensive and thus expensive, especially when specialised expertise is required. To address this challenge, semi-supervised learning methods that could achieve similarly high performance with less labelled data by using unlabelled data have been developed.   In this talk, Mr Arun Raja will address how he and his team at A*STAR’s Institute for Infocomm Research (I2R) have developed a novel technique, named Diverse and Consistent Multi-view Networks or DiCoM based on the multi-view learning framework for semi-supervised regression. The talk will cover the related work and the motivation behind developing the technique and its specifics. He will also discuss the experimental results across various data modalities such as tabular, visual and time-series along with possible future research directions. 

Related to the theme of Scientific Machine Learning, this session describes an on-going research at Caltech that integrates learning into the design of reliable controllers for dynamical systems. To achieve certifiable control-theoretic guarantees while using powerful function classes such as deep neural networks, careful integration of conventional control & planning principles with learning into unified frameworks are required.

Prof Yue Yisong from Argo AI will illustrate the methods that admit relevant behavioural guarantees and are practical to deploy. These methods are demonstrated in a variety of applications, including smooth broadcasting of sports games, agile aerial flight while dealing with perturbations and boundary conditions, and fast planning in resource-limited safety-critical settings such as Mars rover navigation.

The label-noise problem belongs to inaccurate supervision - one of the three typical types of weak supervision. Label noise may exist in many real-world applications where budgets for labeling raw data are limited. However, the famous class-conditional noise (CCN) model, which assumes that the label corruption process (namely, the class-label flipping probability for corrupting the class-posterior probability) is instance-independent and only class-dependent, is not enough in expressing/modeling real-world label noise and thus we need to go beyond it. In this talk, Dr Niu Gang will introduce his recent advances in robust learning against label noise when the noise is significantly harder than CCN. The first noise model is called instance-dependent noise (IDN), where the label flipping probability is conditioned on not only the true label but also the instance itself; IDN is non-identifiable and needs to be approximated with additional assumptions and/or information. The second noise model is called mutually contaminated distributions (MCD), where what has been corrupted is the class-conditional density for sampling instances rather than the class-posterior probability for labeling instances. The learning methods for handling IDN and MCD show that label-noise learning beyond CCN is at least possible and hopefully there will be new methods making it more and more practical.

Current approaches for (multi-horizon) time-series forecasting using recurrent neural networks (RNNs) focus on issuing point estimates, which are insufficient for informing decision-making in critical application domains wherein uncertainty estimates are also required. Most existing methods (such as Bayesian recurrent neural networks, quantile regression models, latent variable models with deep state-space architectures, and post-hoc uncertainty estimates) share at least one of the two major drawbacks: (1) they require substantial modifications to the underlying model architecture, and (2) they provide no theoretical guarantees on frequentist coverage, with exceptions being computationally intractable. In this talk, Ms Kamilė Stankevičiūtė from the University of Cambridge will present a conformal forecasting architecture which extends inductive conformal prediction (ICP) to the multi-horizon time-series forecasting problem, providing a lightweight algorithm for computing forecast uncertainty intervals with frequentist coverage guarantees.

Digital transformation is an enterprise view of “smart manufacturing” or “Industry 4.0” which aims to optimise the entire product development process by integrating individual steps of the value-adding chain of the business with meaningful data sharing and cross-organisational communication. However, the current practices of collecting, structuring, and sharing of data in a product life cycle (from marketing, design, manufacturing, sales to service, and so on) result in isolated and siloed data platforms which hinder true digital transformation of a factory. One of the two major barriers for the data integration is the top-down decision push structure of most organisations, while the other is the absence of data representation tools for archiving product development decisions, including functional descriptions of design. Cross industry studies show that on average, less than half of an organisation’s structured data is actively used in making decisions and less than 1% of its unstructured data is analysed or used at all. In this talk, Prof Kim Sang-Gook will present a novel approach on how to represent conceptual and decision reasoning data that AI tools can semantically understand, store, and retrieve when necessary. The integrated use of big decision reasoning data across the product life cycle will enable true digital transformation of the factory. 

Research in cognitive science has provided extensive evidence of human cognitive ability in performing physical reasoning of objects from noisy perceptual inputs. Such a cognitive ability is commonly known as intuitive physics. With advancements in deep learning, there is an increasing interest in building intelligent systems capable of performing physical reasoning from a given scene to build better artificial intelligence (AI) systems. As a result, many contemporary approaches to modelling intuitive physics for machine cognition have been inspired by literature from cognitive science. Existing models of intuitive physics based on vision can predict the results of physical interactions. However, they primarily concentrate on end-to-end deep learning methods to infer physical properties (such as mass, friction, and velocity) in latent space for performing physical reasoning. In this talk, Mr Duan Jiafei will first briefly introduce the recent advancements in machine learning approaches for modelling intuitive physics. Then, he will talk about some of his most recent initiatives to create fresh AI models for physical reasoning via perception using cognitively inspired methods, as well as how these models might further improve explainability and interpretability in learning.

The upcoming talk covers a brief overview of fundamentals in optimal design construction and how metaheuristics can be used to design more efficient biomedical studies. Prof Wong Weng Kee will demonstrate the usefulness of nature-inspired metaheuristic algorithms for finding various hard-to-find optimal designs for complex models. They include locally optimal designs, minimax or standardised maximin optimal designs or Bayesian optimal designs.

As applications, Prof Wong will discuss search strategies to estimate some or all parameters in a compartment model, estimate the biological optimal dose in a continuation ratio model, construct multiple-objective optimal designs and extend Simon’s two-stage adaptive designs for Phase II trials to incorporate testing one of several pre-specified alternative hypotheses.