June 2025

TechTalk – Minimally Invasive Brain-Computer Interface (BCI) Research @ HKU

June 27, 2025 (Friday) 4:30pm-5:30pm
This talk will present the latest advancements from Professor Chan’s research team and his startup company Brainsmart, dedicated to developing minimally invasive brain-computer interface (BCI) technologies for stroke rehabilitation and neurological recovery. His team focuses on flexible neural probes and high-density ECoG arrays that can be implanted without craniotomy, offering safer options for high-risk patients. Professor Chan will highlight their recent results in decoding neural signals related to hearing perception and motor intentions, and discuss how these findings contribute to future neuroprosthetic applications. The talk will also cover our pathway from technology development to clinical translation, aiming to enhance functional recovery and quality of life for patients.

TechTalk – Building Multi-dimensional Parallel Training Systems for Large AI Models

June 24, 2025 (Tuesday) 4:30-5:30pm
The increasing modeling capacities of large DNNs (e.g., Transformer and GPT) have achieved unprecedented successes in various AI areas, including understanding vision and natural languages. The high modeling power a large DNN mainly stems from its increasing complexity (having more neuron layers and more neuron operators in each layer) and dynamicity (frequently activating/deactivating neuron operators in each layer during training, such as Neural Architecture Search, or NAS). Dr. Cui’s talk will present his recent papers (e.g., [PipeMesh, in revision of a journal], [Fold3D TPDS 2023], [NASPipe ASPLOS 2022], and [vPipe TPDS 2021]), which address major limitations in existing multi-dimensional parallel training systems, including GPipe, Pipedream, and Megatron. Fold3D is now the major thousands-GPU parallel training system on the world-renowned MindSpore AI framework.

TechTalk – More is Less: Dynamic Sparse Processing in the Era of Sustainable AI

June 19, 2025 (Thursday) 4:30-5:30pm
With modern AI seemingly transforming all aspects of our modern society like magic, it is easy to forget the impacts such modern technology marvel is causing to our fragile environment. From massive industry-scale training of large neural networks of epic sizes, to the proliferation of folding AI inference in our everyday activities, AI applications are rapidly increasing the stress to our global energy and water infrastructure. The coming era of AI demands not only the smartest AI models, but also a new generation of sustainable AI mindset that rethinks the when and how to apply AI intelligence.
In this talk, Professor So will discuss one angle that addresses the “how” question with intelligent algorithm-architecture co-designed systems that reduce both energy consumption and computing latency using dynamic sparse processing. Sharing the same “more is less” principle, a series of works from a dynamic sparse processing system for event cameras to token-steering in modern diffusion models will be discussed. Together, these works illustrate that by doing a little bit more upfront intelligently, it is possible to drastically reduce the amount of work necessary to perform the same AI inference operation during run time without affecting the accuracy of a model. The results are solutions that not only are fast, but they are also orders of magnitude more energy-efficient than typical GPU-accelerated systems.