Mr. Chenxiang Ma received the Best Oral Presentation Award and Mr. Yu Zhou received Best Poster Presentation Award in the COMP 50th Anniversary Research Student Conference.
Jun 2024
The COMP 50th Anniversary Research Student Conference was successfully held on 24 June 2024 at the New World Millennium Hong Kong Hotel. Organized by the Department of Computing at The Hong Kong Polytechnic University, the event brought together numerous speakers, scholars, and students to share knowledge and celebrate outstanding academic achievements. MA Chenxiang and ZHOU Yu, PhD students in the MIND Lab, received the Best Presentation Award and the Best Poster Award for their groundbreaking research titled “Scaling Supervised Local Learning with Augmented Auxiliary Networks” [1] and "CausalBench: A Comprehensive Benchmark for Causal Learning Capability of Large Language Models" [2].
â–²Mr. Chenxiang MA(3rd right)
The Best Presentation Award honors outstanding oral presentations at the conference, evaluated by a panel of expert judges. A total of eight students received this accolade, each awarded 5,000 HKD. This recognition highlights the awardees' research excellence and innovation, encouraging more graduate students to strive for academic excellence. MA Chenxiang's research developed a novel method to address the limitations of traditional deep neural network training. Inspired by synaptic plasticity mechanisms in the brain, this approach promises significant improvements in efficiency, parallelism, and memory consumption. Traditional backpropagation (BP) methods, while widely used, face challenges such as biological implausibility, limited parallelization, and high GPU memory demands. Researchers have proposed local learning approaches to overcome these challenges, but these approaches have suffered from significantly lower accuracy compared to BP. Mr. MA's AugLocal method enhances the compatibility of locally trained layers through auxiliary networks that produce representations similar to those of BP. Extensive experiments have demonstrated AugLocal’s effectiveness, achieving comparable accuracy to BP while significantly reducing GPU memory usage. Mr. MA's work offers new opportunities for the efficient training of large-scale models, paving the way for more efficient and scalable AI systems.
â–²Mr. Zhou YU (1st left)
The Best Poster Award, given to eleven exceptional students, acknowledges innovative research presented in poster format. Winners receive a cash prize of 2,000 HKD, celebrating their academic excellence and significant contributions to their respective research areas. ZHOU Yu's work focuses on enhancing causal learning in large language models (LLMs). Causality is crucial for understanding data in real-world contexts, impacting output explanation, model evolution, and counterfactual generation. With LLMs gaining prominence, evaluating their causal learning capabilities is essential. However, existing evaluations are limited and homogeneous. To address this, ZHOU proposed CausalBench, a comprehensive benchmark that rigorously and quantitatively measures LLMs' causal learning abilities. CausalBench includes diverse tasks, and prompt formats, and evaluates causal relations of varying complexities, setting a new standard for LLM evaluation in causal learning.
MA Chenxiang and ZHOU Yu are currently pursuing their PhD degrees with the Department of Computing at The Hong Kong Polytechnic University, Hong Kong SAR, China. Mr. MA’s current research interests include learning algorithms in spiking neural networks and efficient deep learning. Mr. ZHOU’s current research interests include federated learning, causality-based machine learning, and large foundation models.
Reference:
[1] Ma C, Wu J, Si C, et al. Scaling Supervised Local Learning with Augmented Auxiliary Networks[J], 2024, ArXiv preprint: https://arxiv.org/pdf/2402.17318​
[2] Zhou Y, Wu X, Huang B, et al. CausalBench: A Comprehensive Benchmark for Causal Learning Capability of Large Language Models[J], 2024, ArXiv preprint: https://arxiv.org/pdf/2404.06349​