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.