Second exhibition

MindPipe: High-performance and Carbon-efficient Four-dimensional Parallel Training System for Large AI Models

MindPipe, the first 4D parallel training system for large DNN models, has the following objectives:
1. Greatly reducing load imbalance in GPU pipeline parallel stages; 2. Effectively resolving contention of the 3D parallel communication tasks; 3. Deterministically scheduling multiple subnets to be trained in supernet parallelism, a novel parallel dimension proposed by MindPipe; and 4. Automatic near-optimal 4D configuration of GPUs considering both DNN converging efficiency and GPU utilization.

Wireless AI Perception: A New Sense for Machine Intelligence

Computer vision enables machines to “see”. The capability of machine vision based on cameras, however, is fundamentally limited to a certain field-of-view with good lighting conditions – they cannot see through any occlusions or in the dark.
Wireless sensing opens a new sense for machine perception to decipher the physical world, even in absolute darkness and through walls and obstacles.
It can capture human activities invisibly in a contactless and sensorless way.

Augmented Reality based Robot Control System for Customized Garment Production

Garment industry needs to embrace flexible automation and robotics to lower cost and to enhance human productivity. As it is too complex to achieve full automation in garment production, collaborative works between robots and humans will be foreseen. The project demonstrates the feasibility of deploying multi-robots by combining computer vision and robotic technologies.