Paper Reading #19

Comments:
Comment #1
Comment #2
Reference:
Title: Addressing the Problems of Data-Centric Physiology-Affect Relations Modeling
Author: Roberto Legaspi, Ken-ichi Fukui, Koichi Moriyama, Satoshi Kurihara, Masayuki Numao, Merlin SuarezVenue: IUI 2009/2010
Summary:

"Data-centric affect modeling can render itself restrictive in practical applications for three reasons, namely, it falls short to investigate feature optimality, focuses on inferring discrete, rather than continuous, affect classes and deals with small to average sized datasets." Reporting people's emotions is an incredibly complex problem. People can categorize their emotions into discrete categories, but this does not truly explain what happens neurologically and physiologically. Furthermore, theorists who classify emotion have long been divided on whether to classify emotions as categorical or dimensional. Thus the issues in this paper are not easily explained.

The machine learning described in this paper uses continuous values to describe affective states. This proposes a problem because it leads to regression analysis. Most of the new, state-of-the-art machine learning algorithms have either a O(n^2) or O(n^3) complexity. Thus, large amount of data required for this issue can slow down a system by a huge amount. Music was used for emotion induction, manipulation, and regulation. From the readings collected, feature vectors with 49 attribute values were constructed. An EEG was used to label feature vectors with affect values. The EEG used "emotion spectrum analysis to induce from brainwave signals the user affective states." Essentially, both the affective states and the physiology of the state were recorded at the same time to truly understand each aspect. There were 4 affect types, and thus 4 datasets constructed. These datasets had different continuous affect labels but the same physiological attribute values. Also, a base case, or period of rest, was used at the start of each experiment and in between each sound.

The results of this system were exceptional, but difficult to interpret in the format provides. Even when the number of features were reduced, the results remained very efficient.
Discussion:
I think this was the most difficult paper to read yet. I still have not truly figured out what this paper is about, or more specifically what this system is targeted at ultimately accomplishing. As a result, I have a hard time really saying whether or not I like the article. I don't know what an affect type is, or how it's useful, and don't think the article adequately explained the issue. I even read this article twice. My assumptions are that this is used for some psychological purpose, but that's merely a guess. I really wish there were more of these articles which were easier to read and better explained. I wish I could provide more personal analysis, but I really can't because I simply can not understand the purpose of this paper, or many of the technical details involved.