주요서비스 바로가기 안내


세미나

홈으로 > 게시판 > 세미나
[IEEE/GIST EECS Seminar] 12/1(Fri.), 4:00 PM, Considerations for commercializing machine learning applications, Dr. Taesu Kim (Neosapience, Inc.)
작성자관리자 아이피172.26.19.25
작성일17-11-29 14:23 조회수36
파일

IEEE/GIST EECS Seminar

 


Host: Jong Won Shin / Language: English

December 1 (Fri.) 2017, 16:00~17:00

GIST College C-Bldg. 1st Floor #101


Considerations for commercializing machine learning applications

 

Taesu Kim, Ph.D.

Neosapience, Inc.

 

 

[Abstract]

In recent years, deep learning has been showing amazing performance improvement in most of machine learning problems and driving potential applications in multiple industries. For example, it classifies images, recognizes speech, understands natural languages, and answers questions, and even plays Go better than human champion. People often think that deep neural network is a black-box that requires only large dataset and doesn’t need any handicraft work and domain expertise. It is partially true if we build an engineering prototype with wellprepared dataset. However, it may be difficult to collect well-prepared dataset in a large scale. For example, speech synthesis requires tens of hours of speech recordings from a single speaker, which cause high cost and efforts. Speech recognition for kids needs thousands of kids’ speech recordings which may be much more difficult than collecting adults’. Moreover, in reality, the prototype often fails in real usage scenarios because prepared dataset may be different from the dataset collected in real environment. For example, noisy and reverberant acoustic conditions degrade speech recognizer accuracy. In order to develop commercial applications, it is mandatory to collect large dataset easily or augment existing data. It is also necessary to prevent such failure cases and repeatedly debug the models, which often require a lot of handicraft work and domain expertise. It is even more difficult for deep neural networks due to the large amount of data and poor interpretability of the model. In this talk, I will present common practices in commercialization of machine learning products and propose necessary technologies to ease the difficulties in collecting and augmenting data, and discuss how to debug deep learning model for developing commercial applications.


[Bio]

Taesu Kim was received B.S. degree from Hanyang University in 2001, and M.S. and Ph.D. from KAIST in 2003 and 2007, respectively. From 2004 to 2006, he worked for Institute for Neural Computation at UCSD as a visiting scholar, where his research was to develop unsupervised feature learning and source separation. From 2007 to 2010, he worked for LG Electronics as a senior research engineer, where he developed machine learning algorithm for audio signal processing. Then he joined Qualcomm to establish Korea Research Center in 2010. While working for Qualcomm, he developed machine learning algorithm for low power microphone such as environmental sound classification, audio event detection, key phrase detection, and speech recognition, and he has promoted to staff engineer, staff engineer/manager, and senior staff engineer/manager in 2011, 2012, and 2014 respectively. Recently, he founded a startup named Neosapience, Inc. to enable commercial grade AI technology for ordinary developers and companies.


다음글  
이전글  [EECS Colloquium] 11/30(Thu.) 16:00, Cyber Resilience and Metrics, Dr. Jin-Hee Cho (USARL)

카피라이트