Haozhe Sun

- Published 7 papers at prestigious international conferences, including NeurIPS, IJCNN, and PMLR. For the complete list, please visit: https://sunhaozhe.github.io/publications/ - Supervised by Isabelle Guyon (Google Brain, ChaLearn), Felix Mohr, and Hedi Tabia. - Keywords: deep learning, neural network compression (CNNs, Transformers), few-shot learning. - Created OmniPrint, a data synthesizer, and Meta-Album, a meta-dataset, used in over 3 top-tier competitions and various studies, enhancing few-shot learning and meta-learning benchmarks. - Demonstrated improved performance of episodic few-shot learning methods regarding training efficiency. - Developed a "Reuse, Reduce, Recycle" strategy to enhance the utilization efficiency of a pre-trained neural network, reducing the model size and FLOPs by 84%, CPU inference latency by 74%, and GPU inference latency by 57% on average. - Developed a differentiable method to prune a transformer, achieving a 58%-69% increase in inference speed. - Conducted a literature review on modularity in deep learning, categorizing publications across data, task, and model axes. - Organized the NewInML workshops at NeurIPS 2021 and ICML 2022: planning the event and coordination with the main conference, dissemination of the event, invitation and coordination of reviewers and speakers, management of the review process, hosting and moderating the event, etc. - Organized the few-shot learning competition MetaDL at NeurIPS 2021. - Served as a reviewer for NeurIPS and BayLearn.