Algorithm Description
First I would like mention the classifier used in this competition followed by more detail about the methodology involved.
Classifier: A single layer feedforward neural network (SLFN) trained by a novel algorithm: Extreme Learning Machine (ELM) developed in [1] [2] is served as a task classifier.
The signal processing methodology here involves feature extraction, model selection, and classification as described as the following:
1, Feature extraction: A 3-order auto-regression model is used to fit into 1/2-second segment of ECoG signal overlapped by 1/4-second. The resulted coefficients are constructed into a feature input vector for each segment, and thereby 11 feature input vector is generated for each train/test trial.
2, Model selection: In order to achieve good generation on the test trial, the optimal number of hidden neurons in the SLFN needs to be pre-determined. And also the optimal set of the channels that provide the crucial information for classification is found out before using the classifier trained by ELM. Here train and validation set approach is applied to implement the objective function in the genetic algorithm [3], which outputs the optimal combination of the number of hidden neurons and the set of the channels.
3, Classification: A SLFN is trained on the set of 11*278 samples generated from train trial ECoG signal based on the optimal combination mentioned in model selection. Then the trained SLFN applies to the set of 11*100 samples generated from test trial ECoG signal. There are 11 output vectors of SLFN for each test trial, and post-processing of these output vectors to produce one label for one test trial is applied at last.
4, 11-sample combination: Voting strategy is applied into 11 output vectors to produce one winner, the label for each test trial.
References
[1] G.-B.
Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme Learning Machine: A New Learning
Scheme of Feedforward Neural Networks”, 2004
International Joint Conference on Neural Networks (IJCNN'2004),
[2] http://www.ntu.edu.sg/eee/icis/cv/egbhuang.htm
[3] http://www.shef.ac.uk/~gaipp/ga-toolbox/
Notes
Dr G-B
Huang is one of my supervisors; the use of ELM is under his sanction. In this
moment, I am a representative of the team who participate this competition.