TechTalk

TechTalk – Thermal Insulation in Materials for Efficient Energy Conversion

Innovation Wing Two, TechTalk / By

December 19, 2024 (Thursday) 4:30-5:30pm
To enhance the thermoelectric conversion efficiency of materials, the thermal conduction needs to be suppressed, and the lattice dynamics and the thermal transport mechanisms must be better understood. Lattice thermal conduction of conventional solids is dominated by phonon propagation; however, diffuson-like thermal transport can become predominant in materials with ultralow thermal conductivity. New strategies for a simultaneous suppression of both propagative and diffusive thermal transports will be discussed in this talk based on state-of-the-art theories. Zintl compounds were recently found to exhibit exceptional thermoelectric properties. A thorough experimental study of the thermoelectric transport and carrier properties of Zintl compounds will be discussed. It will be shown that a high figure of merit over a broad temperature range can be realized through the suppression of the intrinsic carrier excitation.

TechTalk – Optimizing Distributed Large Model Training in AI Clouds

Innovation Wing Two, TechTalk / By

December 12, 2024 (Thursday) 4:30-5:30pm
Distributed training using a large number of devices has been widely adopted for learning large deep learning models. Improving distributed training efficiency is critical for time, resource and energy consumption of large model learning. In this talk, I will introduce recent research works in my group on optimizing distributed training parallelisms for effective training acceleration and maximal resource utilization. Especially, we have designed optimized strategies and systems for operator sharding, computation and communication scheduling for SPMD parallelism (e.g., in Mixture-of-Experts model training) in both homogeneous and heterogeneous AI clusters, as well as dynamic micro-batching and pipelining to tackle sequence length variation in multi-task model training (e.g., Large Language Model training).

TechTalk – Medical Image Representation Learning via Cross-supervision between Images and Text Reports

Innovation Wing Two, TechTalk / By

December 5, 2024 (Thursday) 4:30-5:30pm
Pre-training lays the foundation for recent successes in radiograph analysis supported by deep learning. It learns transferable image representations by conducting large-scale fully- or self-supervised learning on a source domain; however, supervised pre-training requires a complex and labour-intensive two-stage human-assisted annotation process, whereas self-supervised learning cannot compete with the supervised paradigm. To tackle these issues, we propose a cross-supervised methodology called reviewing free-text reports for supervision (REFERS), which acquires free supervision signals from the original radiology reports accompanying the radiographs. The proposed approach employs a vision transformer and is designed to learn joint representations from multiple views within every patient study. REFERS outperforms its transfer learning and self-supervised learning counterparts on four well-known X-ray datasets under extremely limited supervision. Moreover, REFERS even surpasses methods based on a source domain of radiographs with human-assisted structured labels; it therefore has the potential to replace canonical pre-training methodologies.

TechTalk – Wafer-scale Structural Coloration Using Gray-scale Lithographic Fabrication

Innovation Wing Two, TechTalk / By

November 28, 2024 (Thursday) 4:30-5:30pm
Structural colors use nanostructured building blocks or thin films to resonantly reflect or scatter light to generate colors and can exhibit higher resolution, saturation, and durability than pigment-based colors. To create structural color based paintings, it is essential to develop a capability of spatially varying the dimensions of these nanosized structures. Recently we reported a high-throughput and wafer-scale nanopatterning method by combining interference lithography and grayscale-patterned secondary exposure (IL-GPSE) to spatially modulate nanostructure feature sizes on large scale while maintaining sufficiently high resolution. Here, we employ the IL-GPSE method in the fabrication of wafer-scale structural color paintings, which can improve the patterning efficiency by orders of magnitude when compared with e-beam lithography. The fabrication techniques developed in this work have unique potentials for broader applications in biomedical sensing, spectral filtering, anti-counterfeiting or encryption, etc.

HKAE TechTalk – The Grain Boundary Ratchet: How to Engineer Grain Size

Innovation Wing Two, TechTalk / By

January 9, 2024 (Thursday) 4:30-5:30pm
We demonstrate that grain boundaries (GBs) behave as Brownian ratchets, exhibiting direction-dependent mobilities and unidirectional motion under oscillatory driving forces or cyclic thermal annealing. We observed these phenomena for nearly all nonsymmetric GBs but not for symmetric ones. Our observations build on molecular dynamics and phase-field crystal simulations for a wide range of GB types and driving forces in both bicrystal and polycrystalline microstructures. We corroborate these simulation results through in situ experimental observations. We analyze these results with a Markov chain model and explore the implications of GB ratchet behavior for materials processing and microstructure tailoring.

HKAE TechTalk – Vision-driven Robots: Challenges, Technologies and Applications

Innovation Wing Two, TechTalk / By

November 19 2024 (Tuesday) 4:00-5:00pm
Robotics and AI technologies are rapidly advancing in recent years, but intelligence levels of robots are still far below humans’ expectations. One of the major reasons is that a robot is not good at coordinating visual information captured by its vision systems, i.e. eyes, with motions of its arms, hands, and legs or wheels. Accurate, robust and efficient perception and effective use of the visual information is crucial for robots to successfully perform tasks in natural environments. This talk presents technical challenges in vision-driven robotics, our on-going work and latest results in vision-driven robot manipulation and vision-guided robot navigation, and applications of the technologies in manufacturing, logistics and robotic surgery.

TechTalk – Lifecycle Wake Mitigation Strategies of Wind Farms using Machine Learning Techniques: Layout Optimization and Cooperative Yaw Control

Innovation Wing Two, TechTalk / By

November 7, 2024 (Thursday) 4:30-5:30pm
The turbine wake refers to the trail left by the upstream turbines with the decreased wind speed and the increased turbulence intensity. The adverse wake effect correspondingly results in less energy production and triggers earlier fatigue failure of downstream turbines. The talk aims to propose solutions to lifecycle wake mitigation strategies, i.e., layout optimization in the design stage and cooperative yaw control in the operation stage, through advanced machine learning techniques and optimization methods. A machine learning wake model accurately predicts wake characteristics and demonstrates its advantages when applied to the freeform optimization and renovation of wind farms. A novel double-layer machine learning framework involving Bayesian optimization for cooperative wind farm control is established to improve the overall power output in the maintenance stage. Overall, these proposed solutions offer a promising path forward for the robust development of offshore wind farms in the long term.

TechTalk – MRI at 0.05 Tesla for Accessible Healthcare: Back to the Future?

Innovation Wing Two, TechTalk / By

October 10, 2024 (Thursday) 4:30-5:30pm
Despite a half-century of advancements, global MRI accessibility remains limited, hindering its full potential in health care. Initially, MRI development focused on low fields around 0.05 Tesla, but progress halted after the introduction of the 1.5 Tesla whole-body superconducting scanner in 1983. Using permanent 0.05 Tesla magnets and deep learning for electromagnetic interference elimination, we developed highly simplified 0.05 Tesla MRI systems that operate using a standard wall power outlet and without radiofrequency and magnetic shielding. We demonstrated its wide-ranging applicability for imaging both brain (Nature Communications 2021) and various anatomical structures at whole-body level (Science 2024). Furthermore, we developed three-dimensional deep learning reconstruction methodologies to boost image quality by harnessing extensive high-field MRI data (MRM 2023, Science Advances 2023, Science 2024). We hope these advances will eventually lead to a new class of affordable and computing–powered ultra-low-field MRI scanners for point-of-care applications, addressing unmet clinical needs in diverse health care settings.

HKAE TechTalk – Towards 6G: The Impacts of Wireless Communications Technologies Research to Society

Innovation Wing Two, TechTalk / By

October 22 2024 (Tuesday) 4:00-5:00pm
The rapid advancement of mobile communications technology is profoundly reshaping society and paving the way for the future. This talk will first guide the audience through the innovations and adoption progress of 5G technology, highlighting its transformative impacts. The discussion will then shift to ASTRI’s initiatives aiming at addressing critical pain points and customizing applications to enhance user experiences, showcasing how 5G continues to evolve and unlock its full potential. As we transition from 5G to 6G, a new landscape of opportunities and challenges is emerging, driven by academic research and operator requirements. Furthermore, the integration of artificial intelligence in 6G presents exciting possibilities for enhancing everyday life. The presentation will also introduce ASTRI’s research efforts on shaping a connected future, illustrating how these advancements can foster a more integrated and intelligent society.

Young Scholar TechTalk – Learning Out-of-Distribution Object Detectors from Foundation Models

Innovation Wing Two, TechTalk / By

September 16 2024 (Monday) 4:30-5:30pm
Out-of-distribution (OOD) object detection is a challenging task due to the absence of open-set OOD data. Inspired by recent advancements in text-to-image generative models, such as Stable Diffusion, we study the potential of generative models trained on large-scale open-set data to synthesize OOD samples, thereby enhancing OOD object detection. We introduce SyncOOD, a simple data curation method that capitalizes on the capabilities of large foundation models to automatically extract meaningful OOD data from text-to-image generative models. This offers the model access to open-world knowledge encapsulated within off-the-shelf foundation models. The synthetic OOD samples are then employed to augment the training of a lightweight, plug-and-play OOD detector, thus effectively optimizing the in-distribution (ID)/OOD decision boundaries. Extensive experiments across multiple benchmarks demonstrate that SyncOOD significantly outperforms existing methods, establishing new state-of-the-art performance with minimal synthetic data usage.