Abstract:

There are always large demands for more suitable and efficient machine learning ....<more>

Abstract:

There are always large demands for more suitable and efficient machine learning techniques for hi-tech applications. In this talk we will start with a brief review of the architecture of a standard deep learning network for classification, and image to image applications. Many possible approaches can possibly be used to achieve improvement of the architecture of the deep learning structure, which include making novel improvement on the design of the baseline model, information aggregation, resolving the conflict between optimization and generalization and normalization approach of deep learning algorithms, etc. For the present presentation, we will just concentrate on the study of baseline models for deep learning, and will give a brief discussion on the evolution of building blocks for deep learning architectures. We will then proceed with the discussion of some currently proposed baseline models. The techniques can be used as a reference for those who want to design their own deep learning networks for specific applications. At the end of the discussion we will briefly talk about our proposed baseline models making use of joint back projection and residual network for image super-resolution. Demonstrations and experimental results will be provided to show the effect of the new design, and research trend along this direction will also be discussed.

Abstract:

Deep learning has achieved great progress in recent years. ....<more>

Abstract:

Deep learning has achieved great progress in recent years. However, its huge demand for labelling information often overburdens engineers. In this talk few shot learning is focused. First, I will give a short tutorial on Few-Shot Learning. Then I will introduce our works on Few-Shot Learning. 1. We use the structure of dataset to implement a Self-reinforcing Unsupervised Matching algorithm. 2. We use the knowledge of the Chinese characters and encode a character to a vector, and learning the map from a character image to the vector. 3. We separated and build an illumination base, and augment the feature of an image with the base.

Abstract:

With the application of speaker diarization in diversified scenarios ....<more>

Abstract:

With the application of speaker diarization in diversified scenarios, such as telephone, meeting, broadcast, restaurant, clinic, courtroom, etc., a single type of diarization technique could not yield satisfactory performance. In this talk, we will introduce a scenario-dependent speaker diarization pipeline with a divide-and-conquer strategy. In this pipeline, different combinations of techniques are designed, including the deep-learning-based speech enhancement for preprocessing, conventional clustering-based diarization, and iterative neural diarization. As a demonstration, our scenario-dependent pipeline won the first place for all tasks of DIHARD-III challenge.

Abstract:

Bilingualism and biliteracy education help unlock human potential and ....<more>

Abstract:

Bilingualism and biliteracy education help unlock human potential and have been a priority research focus in Singapore for several decades. Bilingualism is a cornerstone of Singapore’s education system with most students taking classes in English and their mother tongue languages (MTL). However, proficiency in MTL can differ greatly. Young adults who received more/earlier MTL exposure in their early years report higher MTL proficiency relative to their peers, and early language exposure also results in changes to the auditory system that modulates later language learning. To help improve the nation’s bilingual education, we have attempted to transform insights gained from the scientific study of cognitive psychology and neuroscience into real-life applications that can be delivered widely and effectively. In this talk, I will highlight the journey of a project dubbed as “Singing and Listening to Improve Our Natural Speaking (SLIONS)” that was transformed from a rehabilitation application to EduTech. The key idea behind the project is to combine teachers’ human intelligence with big data, and to convert them into machine intelligence which will facilitate joyful and effective MTL learning. The SLIONS’ activity and feedback components incorporate features designed to support self-regulated learning during deliberate practice as well as features informed by variation theory and multimedia learning theory. I will also discuss the joys and pains working in such a multidisciplinary environment.

Abstract:

Convolutional Neural Networks in Deep Learning have dominated the research ....<more>

Abstract:

Convolutional Neural Networks in Deep Learning have dominated the research study in the recent 8 to 10 years for object recognition, tracking, inference and image/video signal processing. In the recent three years, the transformer networks which make use of the concepts of attention mechanism and non-local means have been very successfully applied to language processing. Applications of the transformer networks in image/video processing have just been emerging. In this talk we will try to recall some conventional machine learning elements as the hints on coming up with the encoder structure of the transformer model, and subsequently relate this to the basic structures of self-attention, spatial-attention and channel attention, and make use of them to build some standard transformer architectures. At the end of the presentation, we will highlight some applications and our research works relating to this hot topic in deep learning

Abstract:

Recent study shows machine learning models are readily vulnerable to ....<more>

Abstract:

Recent study shows machine learning models are readily vulnerable to adversarial attacks. Substitute attacks, typically black-box ones, employ pre-trained models to generate adversarial examples. It is generally accepted that substitute attacks need to acquire a large amount of real training data combined with model-stealing methods to obtain a substitute model. However, the real training data may be difficult (if not impossible) to be obtained for some practical tasks, e.g., in medical or financial sectors. As the first trial study, the talk will present our proposed model-stealing method that does not require any real training data. The method develops specially designed generative adversarial networks (GANs) for substitute training. The experimental results demonstrate that the substitute models produced by the proposed method without any real training data can achieve competitive performance against the baseline models trained by the same training set as in the attacked models.

Abstract:

Depth maps obtained by commercial depth sensors are always in low-resolution, ....<more>

Abstract:

Depth maps obtained by commercial depth sensors are always in low-resolution, making it difficult to be used in various computer vision tasks. Thus, depth map super-resolution (SR) is a practical and valuable task, which up-scales the depth map into high-resolution (HR) space. In this report, I will focus on the recent research works of our research group on depth super-resolution based on deep learning. Limited by the lack of real-world paired low-resolution (LR) and HR depth maps, most existing methods use down-sampling to obtain paired training samples. To this end, a large-scale dataset named “RGB-D-D” is constructed firstly, which can greatly promote the study of depth map SR and even more depth-related real-world tasks. The “D-D” in our dataset represents the paired LR and HR depth maps captured from mobile phone and Lucid Helios respectively ranging from indoor scenes to challenging outdoor scenes. Besides, we provide a fast depth map super-resolution (FDSR) baseline, in which the high-frequency component adaptively decomposed from RGB image to guide the depth map SR. Moreover, for the real-world LR depth maps, our algorithm can produce more accurate HR depth maps with clearer boundaries and to some extent correct the depth value errors.

Abstract:

Video has been widely used in almost every sector such as Internet ....<more>

Abstract:

Video has been widely used in almost every sector such as Internet and security surveillance and the amount of video data is growing explosively. Video summarisation aims to produce a visual summary of a video by distilling the essence of the video. In recent years, there have been significant progress on videos summarization due to various from both conventional data reconstruction perspective and deep learning perspective. This lecture aims to provide the audience an integrated view of both conventional approaches and deep learning based approaches, which will bring the field of video summarization to a new level and broaden the applications of video summarization.

Abstract:

We often capture images and 3D point clouds of a target scene through ....<more>

Abstract:

We often capture images and 3D point clouds of a target scene through glass. The captured data may include undesired reflection artifacts since light passes through and is reflected on a pane of glass simultaneously. Such reflection artifacts may degrade the performance of image processing and computer vision techniques. In this lecture, we talk about our research work on automatic reflection removal for images and 3D point clouds. We first propose reflection removal algorithms for images. We consider multiple glass images and separate the transmission and reflection images by restoring the gradients of the transmission images via low-rank matrix completion. We also propose a semantic based reflection removal network for a single glass image that investigates a non-linear intensity mapping for glass images. Next, we introduce the reflection removal algorithm for large-scale 3D point clouds (LS3DPCs) captured by LiDAR scanners. We estimate multiple glass regions by using the characteristics of LiDAR. Then we detect virtual points by recursively traversing the possible trajectories of light reflection and selecting the optimal trajectory satisfying reflection symmetry and geometric similarity.

Abstract:

Large-scale river disasters have repeatedly caused loss of lives and ....<more>

Abstract:

Large-scale river disasters have repeatedly caused loss of lives and social assets. In this presentation, I will introduce the proposed Cyber-Physical System (CPS) for Active Control of River Channels, which aims to minimize the damage caused by river flooding. This project is being conducted under the interdisciplinary research framework of information science, civil engineering, and physics. In addition to the introduction of a prototype system using an indoor water flow experiment device, topics such as monitoring of real rivers, estimation of riverbed conditions from river surface states using high-dimensional signal processing, and estimation of water level distribution in water systems using graph signal processing will also be mentioned.

Abstract:

The development of 5G and IoT technologies, together with the ....<more>

Abstract:

The development of 5G and IoT technologies, together with the evolution of ubiquitous sensing technologies, has led to smart radio sensing that can monitor, analyze, and react to our daily activities. In this talk, we will present our research work on smart radio sensing, which utilizes the ubiquitous wireless radio signals for various sensing problems including gesture recognition, vital sign tracking, pose estimation, and so on. In essence, the wireless radio signals emitted from the transmitter, before arriving at the receiver, are modulated by different activities of the environment. By analyzing the modulated radio signals at the receiver, we are able to extract the information of different activities for the corresponding sensing problems.

Abstract: Deep learning has become the main solution in many AI ....<more>

Abstract: Deep learning has become the main solution in many AI applications, such as computer vision and speech recognition. Similarly, the potential to apply deep learning in Brain-Computer Interface is also promising. In this talk, I will introduce the deep learning for brain-computer interfaces, including its challenges and opportunities. I will present the state-of-the-art methods and results of the deep-learning BCI research from our group and other BCI labs.

Abstract: Brain-Computer Interfaces (BCIs) have been in the spotlight ....<more>

Abstract: Brain-Computer Interfaces (BCIs) have been in the spotlight due to the announcement of BCI research programs in several big companies. From the academic and practical research perspective, it is important to have a realistic and visionary picture on what BCI can do, where we are today, and what BCI can offer in the future? This talk will provide a brief introduction to brain-computer interfaces and their neuroscience basis. Various examples will be used to illustrate the scope of current and perspective BCI applications.

Abstract: In the past decade, we witnessed a major progress in both research....<more>

Abstract: In the past decade, we witnessed a major progress in both research and novel use of Brain Computer Interfaces (BCIs). With more sophisticated algorithms and innovative protocol development, BCIs have found number of new applications, beyond its long term objective to assist the physically disabled to access computers for communication and control. These new medical applications ranges from the therapy for neurological disorders (for example stroke), to the treatment for mental disorders (for example, ADHD). Researchers in the field are pushing boundary fiercely to expand BCI applications to the new frontiers. In this talk, we will introduce the current medical applications with BCI technologies, and discuss future possibilities and ongoing efforts.