News

Articles 92

A seminar titled "Green Machine Learning Methods  " is held in KAIST.



Title : Green Machine Learning Methods

Date : July 2 (Monday), 2012 15:00

Location : Wooribyul Seminar Room(#2201), Department of Electrical Engineering
Speaker : Prof. S.Y Kung/ Department of Electrical Engineering of Princeton

Abstract:  Green  machine learning technology  is  vital for many mobile  devices, especially in those applications  where power-on is constantly required.  Such siutations arise in for example BCI, pace makers,  and  mobile security-alarms.   On the other hand,  a kernel-based approach has rapidly positioned itself as a  mainstream  machine learning approach.   To ensure its suitability  for green IT applications,  the kernel approach must  be examined from   the perspective of its computational efficiency and power consumption; especially, how  the kernel method copes with the curse of dimensionality in terms of feature size (M) and/or training dataset (N).   While the kernel approach holds some advantage when M is relatively large, it tends to be more susceptible to high learning/inference cost when N is extremelylarge.  The latter scenario is commonly encountered in biomedical applications, such as seizure or arrhythmia detection, where the number of training data easily approaches 100K or higher.  

In this research, the kernel methods are re-examined, and to a certain extent,  re-designed  so as to arrive at an optimal tradeoff  between design freedom and computational complexity. Two main results of the study are as follows:  First, we propose a fast-PDA (Perturbed Discriminant Analysis)  scheme  and show that it is computationally efficient in the learning phase and operable in low-power in  the inference phase. Second, such an efficient learning scheme can  afford incorporation of an outlier removal scheme so as to improve the prediction accuracy.   This results in a novel Prunned-PDA (PPDA) classifier, in which harmful  “anti-support”  training vectors are identified and removed iteratively.   Based on our simulation on the MIT-BIH ECG dataset, it can be demonstrated that the proposed method has a short time for learning, consumes low power for prediction, and delivers a high detection accuracy. 

No.
Subject
Author
32 [Lab Seminar] Dr. Sungwoong Kim visited SLSP file
Admin
May 04, 2015 33250
31 [Picnic] SLSP will have a picnic for Spring 2015 file
Admin
Apr 07, 2015 39812
30 [Lab Seminar] Youngjoo Seo presented a seminar for action recognition file
Admin
Mar 28, 2015 55898
29 [Lab Seminar] Sanghyuk Park presented a seminar for complex video scene analysis file
Admin
Mar 19, 2015 62982
28 [Defense] Jinho Choi Successfully Defended Ph.D Dissertation file
Admin
Feb 23, 2015 38221
27 [Graduation] Tae-ho Kim, Yongcheon Na file
Admin
Feb 15, 2015 38749
26 [Welcome] New Master's Degree Candidates file
Admin
Feb 02, 2015 36738
25 [Activity] Regular Winter Training file
Admin
Feb 01, 2015 30212
24 [Workshop] Machine Learning Tutorial 2015 file
Admin
Jan 16, 2015 67041
23 [Jinho Choi] Second prize at E5 KAIST file
Admin
Dec 30, 2013 45766
22 [Sanghyuk Park] winning the Best Paper Awards at IEEK Summer Conference 2013 file
Admin
Dec 22, 2013 75965
21 [Seminar] Autoassociative Learning with Modular Recurrent Networks: Brain-Like Cognitive Machine Learning file
Admin
Nov 28, 2013 46067
20 [Seminar] Unsupervised Learning of Representations file
Admin
Nov 01, 2013 64467
19 [Workshop] IEEK Summer Conference 2013 file
Admin
Aug 08, 2013 49180
18 [Seminar] Introduction to Nonparameteric Bayesian Statistical Modeling file
Admin
May 21, 2013 53514
17 [Sanghyuk Park] Samsung Human Tech Paper Contest file
Admin
Jan 25, 2013 69745
16 [Notice] Year-end Party! file
Admin
Nov 26, 2012 72558
15 [Seminar] An Alternative Approach To Handling Missing Data By Deleting Completely Observed Records file
Admin
Nov 23, 2012 75122
14 [Seminar] Overview of Finding Needles in a Haystack file
Admin
Nov 23, 2012 65079
13 [Seminar] Multiple Kernel Learning file
Admin
Oct 19, 2012 165024