Innovation Wing Two

Tech Talk – dPRO: A Generic Performance Diagnosis and Optimization Toolkit for Expediting Distributed DNN Training

Innovation Wing Two, TechTalk / By

Distributed training using multiple devices (i.e., GPU servers) has been widely adopted for learning DNN models over large datasets. However, the performance of large-scale distributed training tends to be far from linear speed-up in practice. Given the complexity of distributed systems, it is challenging to identify the root cause(s) of inefficiency and exercise effective performance optimizations when unexpected low training speed occurs. To date, there exists no software tool which diagnoses performance issues and helps expedite distributed DNN training, while the training can be run using different machine learning frameworks. This paper proposes dPRO, a toolkit that includes: (1) an efficient profiler that collects runtime traces of distributed DNN training across multiple frameworks, especially fine-grained communication traces, and constructs global data flow graphs including detailed communication operations for accurate replay; (2) an optimizer that effectively identifies performance bottlenecks and explores optimization strategies (from computation, communication and memory aspects) for training acceleration. We implement dPRO on multiple deep learning frameworks (PyTorch, TensorFlow, MXNet) and representative communication schemes (AllReduce and Parameter Server architecture). Extensive experiments show that dPRO predicts performance of distributed training in various settings with<5% errors in most cases and finds optimization strategies with up to87.1%speed-up over the baselines.

Tech Talk – Digital Twin-Enabled Synchronization System for Smart Precast Construction

Innovation Wing Two, TechTalk / By

Prefabricated construction is an emerging construction approach to produce prefabricated components in the off-site factory and transport them to the construction site for assembly, which provides enhanced quality, productivity, and sustainability. On-site assembly is an uncertain and complex stage in prefabrication projects, due to high variability of outside conditions, organization of multi-contractors, and geographic dispersion of activities. Information technology is adopted for the management of precast on-site assembly, such as Internet-of-Things (IoT), Cyber-Physical Internet (CPS), and cloud computing, which generate massive digital twins of construction resources and activities. This tech talk introduces a digital twin-enabled real-time synchronization system (DT-SYNC) with a robotic testbed demonstration for smart prefabricated on-site assembly. On-site resources are converted into Smart Construction Objects (SCOs) attaching with UWB and RFID devices to collect and integrate real-time nD data (e.g. identity, location, cost, and construction progress). Through smart mobile gateway, various on-site resources and activities could be real-timely interoperated with their corresponding digital twins. Cloud-based services are provided for real-time monitoring through high-fidelity virtual models, and robotic control with automatic navigations and alerts. Supported by the cyber-physical visibility and traceability provided by digital twins, a real-time synchronization model is designed to organize and coordinate operations and resources with simplicity and resilience, which guarantees that the appropriate resources are spatiotemporally allocated to the appropriate activities.

Biofilm Inhibition in Oral Pathogens by Nanodiamonds​

Inaugural exhibition, Innovation Wing Two / By

Complex microbial communities, e.g., biofilms residing in our oral cavity, have recognized clinical significance, as they are typically the main cause for infections. Diamond nanoparticles, namely, nanodiamonds (NDs) have been demonstrated to work as an effective antibacterial agent against planktonic cells (free-floating state) due to their many promising physico-chemical properties. However, little is known about the behaviors of NDs against biofilms (sessile state).

Tech Talk – Sensing and Data Processing for Human Balancing Evaluation

Innovation Wing Two, TechTalk / By

Balance ability is human’s basic physiological ability that ensures stable standing and walking. Falls are major threats to the health and independent living of the elderly. 10% of the falls in the elderly are associated with fractures, and some can lead to head injuries and deaths. A highly effective human balance sensor is invented, which can capture high-resolution dynamic pressure distribution under human feet. The variation of the pressure distribution can be used to solve the dynamics of human motion while standing on the balance sensor. The algorithms using artificial intelligence are developed to assess the risk of falling for elderly people to help them prevent falls. Furthermore, this sensor can be used to measure balance ability of athletes, such as weightlifting, golf and gymnastics. In medical diagnosis, balance ability tests can provide important information for diagnostics of neural disease. It can also be used to identify drunk driving.

Tech Talk – Short-range exposure during close contact and the environmental interaction

Innovation Wing Two, TechTalk / By

Debate and scientific inquiries regarding airborne transmission of respiratory infections such as COVID-19 and influenza continue. Exposure was investigated under a face-to-face scenario, where people experience the highest risk of respiratory infection. The short-range airborne route was found to dominate exposure during close contact. Based on the fact that most of the outbreaks occurred in indoor environments, we built the link between long-range airborne transmission and short-range airborne route. Results suggest that effective environmental prevention strategies for respiratory infections require appropriate increases in the ventilation rate while maintaining a sufficiently low occupancy.

Tech Talk – Cold-formed Stainless Steel Structures under Concentrated Bearing Loads

Innovation Wing Two, TechTalk / By

The application of stainless steel (SS) as an alternative construction material has been developing in the last decade. SS construction has an outstanding structural performance, excellent corrosion resistance, and long-term durability. Moreover, the SS construction has a relatively low maintenance cost with possibility of a longer occupancy period, and thus it promotes sustainability in the construction industry. In combination with the cold-forming technique during the fabrication process, SS structures offer additional strength and a faster construction speed. However, the available international design standards for cold-formed stainless steel (CFSS) structures have not been developed thoroughly, specifically on the strength prediction of a member under concentrated bearing loads, which causes a web crippling. Therefore, a series of laboratory testing and computational simulations were conducted in this research to evaluate the existing design standards. The reliability analysis shows that the available strength predictions in the design standards are not safely used even though they are conservative. This research proposes new strength predictions that are safe and conservative, and it can be used for an improvement of the design standards.

Tech Talk – Light or Fiber Touch

Innovation Wing Two, TechTalk / By

Precision manipulation of various liquids is essential in many fields, including DNA analysis, proteomics, cell assay and clinical diagnosis, chemical synthesis, and drug discovery. Their divisible, sticky, and sometime infectious features impose, however, great challenges on processing them, particularly when their volume is down to nano-/subnano-liter. A blood droplet from an Ebola patient can for example infect medical workers through the skin. For diagnosis, medial workers have to crash, filter, and purify a patient’s blood sample to obtain the virus’s genetic materials. This series of operations, very often in a fluidic medium, is highly infectious. Moreover, fluids stick to surfaces, which will contaminate containers and handling tools, causing potential dangers if the medical wastes are not properly managed. In this talk, Prof. Wang shall demonstrate how a simple light or fiber touch functions as a “magic” wetting-proof hand to navigate, fuse, pinch, and cleave fluids on demand, being capable of reducing and even replacing the usage of disposable plastics in the biomedical and pharmaceutical industries.

Tech Talk – HINCare: Using Heterogenous Information Networks for Elderly Care Recommendation

Innovation Wing Two, TechTalk / By

In Hong Kong, the number of elderly citizens is estimated to rise to one third of the population, or 2.37 million, in year 2037. As they age and become more frail, the demand for formal support services (e.g., providing domestic or escort services) will increase significantly in the coming years. However, there is a severe lack of manpower to meet these needs. Some elderly-care homes reported a 70% shortage of employees. There is thus a strong need of voluntary or part-time helpers for taking care of elders.
In this talk, Prof. Cheng will introduce HINCare, a software platform that encourages mutual-help and volunteering culture in the community. HINCare uses the HIN (Heterogeneous Information Network) to recommend helpers to elders or other service recipients. The algorithms that use HINs and AI technologies for matching elders and helpers are based on our recent research results. This is the first time that HIN is used to support elderly care.
HINCare is now downloadable in Apple and Google Play Store, and has been serving more than a thousand of elders and helpers in NGOs (e.g., SKH and CSFC). The app is originally designed for elderly users, but has now expanded its services to support the Community Investment and Inclusion Fund (CIIF) and 10 NGOs engaged in teenage and family services. The system won the HKICT Award 2021, Asia Smart App Award 2020, and the HKU Faculty Knowledge Exchange Awards 2021 HKU.

Tech Talk – Smarter, smaller and softer robots for medical image-guided minimally invasive surgery

Innovation Wing Two, TechTalk / By

In recent years, there has been a trend towards integrating small, soft and deformable structures into surgical robot systems. Target applications include endoscopy or magnetic resonance imaging (MRI)-guided intervention, where researchers take advantage of soft and flexible robots for their inherent mechanical compliance. However, these flexible robotic systems are often controlled in an open loop or with positional feedback from 3D tracking devices. Not only the real-time feedback of flexible/soft robot configuration or morphology itself is of importance, but also the robot manipulation modelling, as well as its intelligent control, become an area of interest in the field. To this end, this talk will present various robot prototypes, which attempt to resolve unmet clinical and technical challenges for image-guided intervention or surgery, either in strong magnetic field (1.5-3T) by magnetic resonance imaging (MRI) scanner or in confined anatomical space through endoscopy. Machine intelligent approaches, and also the recent advances in continuum robot design and learning-based sensing/control will also be overviewed. These robots have to incorporate with efficient mechanical transmission, thus enabling delicate mechanical force/motion transmitted from actuators to surgical tools in a long and flexible route. The ultimate goal is to provide high-performance control of robotics instruments for safe, precise and effective surgical manipulation. The speaker will not only share his research outcome, but also various difficulties in his up-and-down research journey, from R&D in university, (pre-)clinical trials in hospital, then technology transfer for clinical applications.

Tech Talk – Plastic Response of Cells: from Embryo Development to Disease Detection

Innovation Wing Two, TechTalk / By

Living cells need to undergo significant shape changes during processes such as cell division, migration and tissue formation. Therefore, it is commonly believed that the deformability of cells is intimately related to their capability in executing different biological duties as well as the progression of diseases. In this talk, I will discuss how irreversible deformation of cells ensures proper axial extension of embryos during their development and how the plastic response of tumor cells can be used in monitoring the progression of cancer. Specifically, I will show that the presence of active intracellular/intercellular contraction will trigger the severing and re-bundling of actin filaments in cells (leading to cellular anisotropy and plasticity), elevate the internal hydrostatic pressure of embryo and eventually drive its elongation. In particular, the gradual re-alignment of F-actins must be synchronized with the development of intracellular forces for the embryo to elongate, which is then further sustained by muscle contraction-triggered plastic deformation of cells. In addition, I will also introduce a microfluidic setup developed in our lab allowing us to impose precisely controlled cyclic deformation on cells and therefore probe their plastic characteristics. Interestingly, we found that significant plastic strain can accumulate rapidly in highly invasive cancer cell lines and circulating tumor cells (CTCs) from late-stage lung cancer patients with a characteristic time of a few seconds. In comparison, very little irreversible deformation was observed in the less invasive cell lines and CTCs from early-stage lung cancer patients, highlighting the potential of using the plastic response of cells as a novel marker in future cancer prognosis and monitoring.