Feature Space Quantization for Data-Driven Search

There is a growing need to be able to detect arbitrary patterns and trends in large data sets. Practically detecting arbitrary patterns in neuroimaging databases requires a fast and computationally inexpensive method. Here we present an unsupervised and fast alternative to existing methods of recognition for brain activity. We suggest transforming the decoding-relevant features from brain activity data into signatures represented by binary vectors, to enable computationally inexpensive comparison. We then apply this method to ECoG data recorded from two human subjects and we introduce the results of a binary classification task. We then compare the accuracy of SVM classifications based on spectral power features to those using the binary signatures. Our results demonstrate that SVM classifications using binary signatures can perform significantly above chance level and are comparable to classifications based on feature vectors, for some criteria.