Data Science Professor Fenglei FAN has won the OlympusMons pioneer award for far-reaching research on model compression techniques.
Launched in 2019, the OlympusMons awards recognize and encourage foundational data-storage research and nurture collaboration among industry, academia, and research institutes. All previous recipients include outstanding scholars in this field such as Professor Weimin Zheng (CAS member) from Tsinghua University and Professor Dr, Onur Mutlu (ACM Fellow) from ETH Zürich. The 2024 accolade was presented to Professor Fenglei FAN’s team at City University of Hong Kong for their study on Hyper Compression. A 2024 document, “Hyper-Compression: A Kind of Model Compression Technique Based on Hyperfunction” shows it involving in four distinct characters: (1) Preferable compression ratio; (2) No post-hoc retraining & recalibration; (3) Affordable inference time; and (4) Short compression time.
The core idea behind Hyper-Compression is inspired by a biological observation: a small human genome can remarkably encode the development of a human brain to the scale of billions of neurons and trillions of synaptic connections. This observation suggests the existence of an implicit mapping from genotype to phenotype, capable of expanding a limited number of genetic instructions into a vast array of biological substances. Inspired by this efficient genetic encoding, Professor FAN’s team generalized the idea of hypernet to a parameterized function (hyperfunction) and successfully uses the ergodicity to construct hyperfunction, as Figure 1 shows. More details of Hyper-Compression are explained in a September 2024 scientific paper, “Hyper-Compression: Model Compression via Hyperfunction”.
The paper’s conclusion reads: “We believe that hyper-compression can contribute to Moore’s law of model compression in the near future, i.e., the compression efficiency can be doubled annually, as a solution for the stagnation of hardware Moore’s law.”
Congratulations to Professor FAN on receiving this highly competitive award. His distinguished achievement in Hyper-Compression research enriches the scholarly landscape.