Artificial Intelligence & AI Convergence Network »ç¾÷´Ü¿¡¼ °øµ¿À¸·Î
°³ÃÖÇÏ´Â ColloquiumÀ» 4¿ù 29ÀÏ(¿ù) ¿ÀÈÄ 3½Ã¿¡ °³ÃÖÇÏ¿À´Ï ¸¹Àº Âü¿© ºÎŹµå¸³´Ï´Ù.
When : 2024³â 4¿ù 29ÀÏ(¿ù) ¿ÀÈÄ 3½Ã
Where : ÆÈ´Þ°ü 407È£ Speaker : °º´°ï ±³¼ö(SUNY Korea) Title : On the efficiency of pre-trained word embeddings on Transformers
Abstract : In this talk, I will discuss a counter-intuitive phenomenon where state-of-the-art pretrained word embedding vectors perform poorly compared to randomly initialized vectors. More specifically, such a tendency is observed almost exclusively in Transformer architectures, making it a problem to most modern NLP systems that are heavily dependent on Transformers.
I also describe a very simple remedy that somewhat alleviates this shortcoming, as well as numerous empirical results to back this claim.
If time permits, I will also share a failure case of another popular deep learning system (autograd), leading us to the discussion of whether we should take prevalent technologies for granted in this field.
Bio : Byungkon Kang got his PhD at KAIST in 2013 before joining Samsung Advanced Institute of Technology. After that, he spent a few years at Ajou University as a research professor.He is currently an assistant professor at SUNY Korea, working on various AI/ML topics.