Maps

Depending upon the random initialization, different features will settle in different parts of the 25*25 plane. However, the topological relations should be preserved in each trained map. We choose the EEG epochs of 4 cases (case-1, case-3, case-4, and case-5) to train the maps. The reason is that the effect of distraction is highest when two things were happened at the same time. In Fig. 4-5(A), there are 4 main clusters in this trained map with 2 phase training. We know that the EEG epoch of 4 cases are clearly recognized by the trained maps and there is some difference Fig. 4-4: The input data.

This was the inputs for the Self-Organizing Map. The input data was pre-processed by the steps described above.

when human response these 4 cases. The filled map by enhanced labeling is showed in Fig. 4-5(B). We can see 4 main clusters clearly in this filled map.

Fig. 4-5: Results of agglomerative clustering of EEG signals by SOM.

The map was trained by 4 cases. Four clusters mark by special colors are showed in the map. (A) The trained map by EEG epochs in 4 cases. (B) The filled map. The interaction of designed two tasks is indicated at the right part. M is the mathematic equation and D means the car deviation. In this map, the EEG epochs of case-2 (M and D are appeared at the same time) are not used.

In Fig. 4-6 we show maps. The investigations of SOM results allow concluding that SOM type algorithm can be used interpreting some mental tasks representing EEG data. We got many maps rapidly and the same phenomenon was found in these maps. We can find that some neurons which located in the middle of particular clusters are not labeled. The winning neurons are activity during training steps, and the neighborhood neurons are also adjusted. The connection among neurons is a key role in this algorithm. Competitive learning algorithm is competition among lateral neurons in a layer (via lateral interconnections) to provide selectivity (or localization) of the learning process. These unlabeled neurons are labeled by the close neighbor labeled neuron. After this computing, we get the better and clearer map in Fig. 4-7.

Fig. 4-6: Results of agglomerative clustering of EEG signals by SOM.

There are four maps in this figure. We verify the results by applying the algorithm many times. Five clusters mark by special colors are showed in each map. The interaction of designed two tasks is indicated at the right part. M is the mathematic equation and D means the car deviation. We combine these two tasks and stimulus onset asynchrony (SOA) to design five cases. The case-1~case-3 are the dual tasks and case-4~case-5 are the single tasks.

Our SOM based exploratory data analysis using EEG suggests existence of distinct signatures among these five cases. Fig. 4-6 and 4-7 show the topology relations of the collected EEG epochs. The EEG epochs of two single tasks are clustered well, but there are several subgroups to the dual tasks, especially the case-2.

Although most neurons labeled case-2 are clustered to some main area, several special neurons are mixed together. The neurons labeled single conditions are mapped to the corners of each map and these two clusters are so congeries. The reason is the changing of brain signals for dealing with single task is consistent with each subject.

When the mathematic question or car deviation appeared, the subject must response

them quickly and correctly. The brain resources are allocated to dealing with just one task during these two single conditions. But these two designed mental tasks are also combined to provide the dual-task situations in our experiment. Each subject is asked to response the tasks as soon as they can. When the subjects must do more than one task at the same time, these tasks scramble brain resources each other. Every subject doesn’t use the same strategy for responding dual-task condition. Someone answer the easy task first then deal with the complex task, but some people can deal with two tasks well.

Fig. 4-7: The filled map.

These four maps come from the maps showed in Fig. 4-1. Each unlabeled neuron in the map is marked a case by the method of flooding. Then all neurons represent one special case to show the topological relations. After flooding, these maps are clearer and easier to understand the distribution of EEG epochs among the five cases.

The three dual-task cases are clustered at the middle of the trained maps. The EEG epochs of case-1 and case-3 are grouped into two main areas, but we can see the

appeared at the same time. When these two tasks are appeared suddenly, subjects chose one tasks to respond first. By the different decision processes, the neurons labeled are close neurons labeled case-4 or labeled case-5. The two tasks are appeared with a 400ms interval and the subjects can have a short time to response one task well.

In our maps, we can see the distribution of the cognitive state and investigate the distraction levels.