Information Technology Reference
In-Depth Information
the fMRI data with LS-SVM (Least Square Support Vector Machine) algorithm,
Formisano et al. could predict “who” (male or female) was saying “what” (which
German vowels) through fMRI classification [4]. Haynes et al. even used fMRI
data from visual cortex to classify and predict human's quick stream of con-
sciousness [5]. In a serial of studies, Mitchell and colleagues applied the methods
of fMRI classification in several cognitive tasks: (1) reading a set of words be-
longing to six types of semantics (such as tools, fruits, and so on), (2) reading two
types of sentences (explicit vs. ambiguous sentences), (3) viewing sentences and
pictures [6,7,8,9]. As a type of data-driven method, classification emphasizes the
mapping between observed fMRI data and the cognitive states. Although many
researchers used this machine learning method to explore the states of human
cognition, most of them focused on perception, but not on high-level cognition
with complex information processing, such as problem solving. Different problem
tasks may be with similar visual stimulus but need very different ways to solve
them. If machine learning approach can be used to predict high-level cognition,
this approach might shed light on how the brain processes information, and on
mind reading needed in brain-machine interface development. In this study, we
explore the approaches from fMRI data to identify the type of the problem the
participant was solving.
Several methods of classification have been used in previous studies. Multi-
voxel pattern analysis (MVPA) used in Haxby's study addressed the issue about
classifier selection [1, 2]. In Kendrick's research, a Gabor Wavelet Pyramid model
was adopted to predict the pictures when participants were watching a large
number of nature images [10]. Hasson used the information of intersubject syn-
chronization to judge whether a participant was watching human faces or build-
ings in a film clip [11]. Sato and colleagues compared the methods of SVM and
MLDA's (Maximum Uncertainty Linear Discrimination Analysis) performances
of predicting human cognitive states in hearing, vision, finger exercises and other
tasks [12]. They indicated that all brain areas were required for MLDA, but much
less brain areas for SVM. In addition, with the less areas involved in, the classi-
fication performances were higher in SVM than that in MLDA. So it seems that
the method of SVM had the superiority in classification. We have also tried some
different classifiers (including back propagation of neural network and Gaussian
Naive Bayes) and also find that SVM is better than others in our study. We will
report how we implemented SVM with the emphasis on feature extraction in
later sections.
2M thod
2.1
Task and Materials
Event related fMRI data were recorded while participants were solving simplified
4*4 sudoku tasks. Sudoku is a combinatorial number-placement puzzle, the goal
ofthepuzzleistofilla4
×
4 grid so that each column, each row, and each of the
four 2
2 boxes contains the digits from 1 to 4 only one time each. As shown in
Figure 1, in this study, we simplify the puzzle and ask participants to give the
×
