Loading...

Brain Decoding Across Subjects

Nasiri Ghosheh Bolagh, Samaneh | 2015

2816 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 47944 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Shamsollahi, Mohammad Bagher
  7. Abstract:
  8. In recent years, techniques in articial intelligence have become an important tool in the analysis of physiological signals. While the application of machine learning techniques has proved useful in other elds, researchers have had difficulty proving its utility for the analysis of physiological signals. A major challenge in applying such techniques to the analysis of physiological signals is dealing effectively with inter-patient differences. The morphology and interpretation of physiological signals can vary dep ending on the patient. This poses a problem, since statistical learning techniques aim to estimate the underlying system that produced the data. If the system (or patient) changes between training and testing, this can cause unpredictable results. More concretely, inter-patient differences mean there is no a priori reason to assume that a classier trained on data from a collection of patients will yield useful results when applied to a previously unseen patient. To cope with the effect of inter-subject, machine-learning researchers have investigated methods to handle mismatch between the training and test domains, with the goal of building a classier using the labeled data in the other subjects that will perform well on the test data in the new domain. This method is called Domain Adaptation or Transfer Learning. In practice, Transfer Learning approaches have been successfully used in many real world applications, such as text classication and image classication. In this project, we address learning target brain decoding model without any labeled data from the target subject by using the Transfer Learning approaches. We propose three novel methods to reduce the inter-subject variability. First, we present a simple and effective algorithm by combining a new subject selection algorithm that it is called Rank of Subject and Riemannian distance based classier for brain decoding. We also use special covariance matrices as features in our classication framework. Second, we propose a method which uses Riemannian Geometry that it has proven to be promising in brain signal processing and uses a new unsupervised domain adaptation to handle mismatching between distribution of training and test data. The basic assumption is that the input domains may share certain knowledge structure, which can be encoded into common latent factors and extracted by preserving important property of original data, e.g., statistical property and geometric structure since different properties of input data can be complementary to each other and exploring them simultaneously can make the learning model robust to the domain difference. Third, we learn a robust decision function (referred to as target classier) for label prediction of instances from the target domain by leveraging a set of source classiers which are pre learned by using labeled instances either from the source domains or from the source domains and the target domain. We demonstrate the success of these algorithms by applying it on real world problems. We compare the performances of different methods on MEG dataset of 16 subjects, from DecMeg2014 competition data sets. Our experiments verify that our proposed method can signicantly outperform state-of-the-art learning methods on M/EEG datasets
  9. Keywords:
  10. Brain Signal ; Decoding Algorithm ; Magnetoencephalography (MEG) ; Decoding Across Subject ; Transfer Learning ; Riemannian Geometry

 Digital Object List

 Bookmark

...see more