SUPERINTELLIGENCE Laboratory Artificial Intelligence Engineering

Data preprocessing and manipulation is the most essential part of the recent deep learning technologies. Data in different areas requires various pre-processing tasks depending on their characteristics or purpose of use. For example, to create a sharp macular image from an ultra-wide-angle fundus image, we rotate, transform, and adjust the original image. We aim to automate the process of fine-tuning the position of the macula, nerve papillae, and blood vessels by matching the size and resolution of the photo with the paired dataset of an existing fundus photo and an ultra-wide fundus photo. This process is largely divided into two steps. First, scaling is a process of adjusting the size of the image by cropping a picture centering on the macula. Second, warping is a process of fine-tuning so that the position of the macula, nerve nipples, and blood vessels of a pair of photos are the same. The rapid growth of computing power and wireless technology has spurred new technologies such as the massive adoption of mobile devices, the increase in mobile content and services, and the Internet of Things (IoT). As a result, research fields are being actively conducted to extract information by processing data in numerous mobile and network fields. Our researches include modeling user mobility data for GAN-based predictions and constructing discretization and continuity (comprehensive) network dataset for deep learning. Facial images are particularly challenging subjects in computer vision. A wide spectrum of information including identity, gender, mood, age, and race can be perceived by looking at a single face image. The fact that many contests in the computer vision field are specifically focused on face-related applications (such as face recognition, verification, detection, and so on) is an indication of the distinct difficulties encountered when dealing with facial images. For instance, some topics include transforming face images in terms of attributes, pose, emotion and expressions, predicting facial age progression, and making generated images more realistic. Lifelogging represents a phenomenon whereby people can digitally record their own daily lives in varying amounts of detail, for a variety of purposes. The record contains a comprehensive dataset of a human’s activities. The data could be used to increase knowledge about how people live their lives. Lifelogging may offer benefits to content-based information retrieval, contextual retrieval, browsing, search, linking, summarization and user interaction. However, there are challenges in managing, analyzing, indexing and providing content-based access to streams of multimodal information derived from lifelog sensors which can be noisy, error-prone and with gaps in continuity due to sensor calibration or failure. We study to select, transform, and archive the valuable and informative records from the huge amount of lifelog data.

Medical Imaging refers to handling medical images by using the computer. It is mainly applied to various applications such as diagnosis and analysis using computer vision and machine learning technology. Medical images are good for research because they have been accumulated for a long time and are relatively refined. It is very different from ordinary photographs and requires collaboration with doctors for accurate analysis. Our research has developed a bone age prediction program using x-ray images. We are also working on a macular image transformation system using fundus images. The fundus image can observe symptoms of systemic diseases such as diabetes and high blood pressure. Moreover, the examination is relatively easy since diagnosis is possible with only one fundus image. Our research is to restore the image around the macula from an ultra-wide-angle fundus photograph to the existing fundus photograph taken through mydriasis examination by exploiting deep learning technologies such as Generative Adversarial Networks (GAN). If a desired level of macular image can be obtained, it is expected that diagnosis will be possible with only wide-angle fundus photographs without mydriasis examination that makes the patients uncomfortable. Another research is to predict developmental status from bone age diagnosis. We study a deep feature fusion based on contrast enhancement and superpixels for bone age diagnosis. The key idea is that the area containing information on bone ages in X-ray images is limited, and features extracted from low contrast X-ray images can lead to poor final bone age diagnostic performance. The deep feature fusion based on contrast enhancement and superpixel method is expected to improve the accuracy of bone age diagnosis. Our research is not limited to this, we aim to develop various techniques of creating visual representations of the interior of a body for clinical analysis and medical intervention, as well as visual representation of the function of some organs or tissues.

Generative Adversarial Network (GAN) driven deep learning approaches have gained attention due to their excellent results. Applications of GAN can be found in data augmentation, privacy preserving, anomaly detection, discriminative modeling, but for networking, they are still in nascent stages. GAN consists a generator and a discriminator connected together in an adversary mode. The generator generates new data instances and the discriminator evaluates them for authenticity. The data distributional learning capability of such generator is useful for optimization problems with dynamic and sparse data. We are exploiting these characteristics of GAN to implement intelligent networking solutions with intention to move towards autonomous networks. In particular, we are currently working on proactive mobility for 5G, where GAN is used to predict next attachment point of the user and handover instant from one attachment point to the next. The results of our model show substantial improvement comparing to state-of-the-art deep learning solutions.

Multi-access Edge Computing (MEC) is a paradigm which moves the capabilities of the central cloud closer to users. In this paradigm, miniaturized and distributed clouds are deployed at the network edge to accelerate conventional services such as live video streaming and healthcare, and also to offer innovative services such as online mobile gaming. Challenges regarding MEC management include computation offloading, allocation of computing resource, and mobility management. We employ the reinforcement learning technique for the MEC management since it tolerates noisy inputs and works well under diverse conditions of not only users geographic distribution but also their varying requirements. The general idea of reinforcement learning is that by gradually interaction with the environment, agent will be able to make better decisions using the past experience. At the initial state, the agent rarely has the idea about how to take action or even what the task is. Then according to agent performance of this task, it will obtain a reward as the feedback. Theoretically, agent could make the optimal decisions if it can explore whole states and obtain corresponding values of its action.

IoT systems can be classified into massive IoT, which requires high scalability, and critical IoT, which basically demands high availability, reliability, and low latency. Especially, the 6G paradigm will accommodate a huge number of devices, in which handling data generated from such devices are a lot more challenging than ever. For this reason, integrating Artificial Intelligence (AI) has become a must to boost the capabilities of IoT systems and applications by automating the decision making. The IoT, in turn, improves the value of AI by allowing real-time connectivity and data communication. In this exchange, the data must be concise, therefore, anomaly cognition is one of the most crucial requirements in the Intelligent Internet of Things. Anomaly cognition includes detection, prediction, and identification of the data faults and cyber-attacks.

Network softwarization is an approach in which network control, management, and functions are decoupled from the data plane (i.e., network elements) and softwarized in a central control plane. Software-defined Networking (SDN) and Network Function Virtualization (NFV) technologies together achieves fully softwarized networks. SDN decouples the control plane of underlying network elements and centralized it in a controller. The controller in SDN decides how traffic is controlled in a network, and to do so it requires some method to communicate with the data plane. One such mechanism is OpenFlow, which is the first standard communication interface defined between the control and forwarding layers of an SDN architecture. Our research in SDN consists of but is not limited to network protection, efficient video streaming, flow table management, and software-defined mobility. Network-function virtualization (NFV) uses technologies like virtual machines or containers to virtualize entire classes of network node functions into building blocks that may be connected, or chained, to create communication services. A virtualized network function, or VNF, may consist of one or more virtual machines running different software and processes, on top of standard high-volume servers, switches and storage, or even cloud computing infrastructure, instead of having custom hardware appliances for each network function. Our research in NFV consists of but is not limited to network slicing based 5G architecture, admission control in network slicing, empirical comparison of virtualization technologies, and efficient monitoring system for VNFs.

A wireless ad hoc network is a decentralized type of wireless network, in which the network does not rely on a pre-existing infrastructure. Instead, each node participates in routing by forwarding data for other nodes, and so the determination of which nodes forward data is made dynamically based on the network connectivity. Ad hoc sensor networks can be very effective in disseminating and collecting sensory data, providing a digital sense of the environmental conditions to smart applications. Herein, spatially distributed autonomous sensors cooperatively pass their sensed data through the network to a base station. The research challenges are now driven toward the speed of collection and reliability under realistic changeable wireless medium condition, given diverse underlying emerging technologies, such as energy harvesting and multi-channel. Today, such networks play a vital role to realize the coming 5th industrial revolution.

Technology is becoming ubiquitous in our daily lives, and human interaction is no longer limited with only computers but has extended to all kind of devices like smart phones, smart homes, smart virtual assistants, and smart vending machines etc. Form factors and interfaces of all these devices are very different from each other and that requires more innovative ways of human interaction to enhance user experience. With widespread adoption of smart phones, users’ intuitive nature towards forms of interactions has significantly evolved. In Human-X-Interaction studies we investigate this evolution of human nature to understand what kind of interactions will be perceived intuitive in future and how interfaces must be evolved to provide seamless experience. In particular, research in intuitive interaction interfaces for variety of smart homes IoT devices and industry 4.0 robotics has significant importance in near future.