Speaker:Professor Cong Fengyu, Dalian University of Technology

Date:April 14, 2015

Time:9:00 a.m.

Location:Lecture Hall, F4, School of Information Science and Engineering

Sponsor:the School of Information Science and Engineering

Abstract:

Event-related potentials (ERPs) are very important tools for the cognitive research including six steps: 1) filtering, 2) segmentation, 3) artifact detection, 4) artifact rejection, 5) averaging and 6) interpretation. The underlying assumptions are that EEG data in each epoch/trial consist of a constant part of ERPs activities and a randomly fluctuating part of brain activities. Hence, through averaging, the constant part is enhanced to represent ERP activities and the randomly fluctuating part is decreased to the degree that it can be ignored.

However, through the conventional data processing method, we have found that ERP activities in the constant part can be overlapped, artifacts cannot be completely rejected, and brain activities in the randomly fluctuating part cannot be sufficiently decreased in practice. This can result in that the interpretation cannot be based on the real ERP activities and the cognitive research can fail.

Thus, before the interpretation, it is necessary to further reject artifacts, remove the randomly fluctuating part, and extract the desired ERP activities out. The conventional methods ignore that ERPs are usually collected using multiple sensors. Therefore, the ERP data are indeed multi-channel signals. We design a systematic tensor decomposition approach to extract the multi-domain feature of an ERP which can be used for group-level analysis of the ERP.

The novel approach mainly includes the following steps

1) implementing the conventional data processing approach to produce ERPs,

2) transforming the time-domain ERPs to the time-frequency domain for producing the tensor including time-frequency representations of ERPs at multiple channels of multiple participants under multiple experimental conditions,

3) defining the number of extracted components for tensor decomposition,

4) performing tensor decomposition on the tensor,

5) selecting the desired multi-domain feature of an ERP for statistical analysis.

Regarding tensor decomposition, both canonical polyadic (CP) and Tucker models can be used. Their difference in extracting multi-domain feature of an ERP is also discussed.

For further information, please visit:

http://202.194.15.203:8080/getNewsDetail.site?newsId=da03675b-47b9-4cb3-a423-0a59a92af7bf