---

Decoding Brain States

We are currently working on several projects in which the goal is to analyze EEG and fMRI data using techniques from machine learning, statistical pattern recognition, and computer vision. Specific projects include:

• analysis of fMRI data during the perceptual decision making (with Luiz Pessoa, Department of Psychology, Indiana University)

• classification of cognitive states (with Bill Heindel and Elena Festa, Department of Psychology, Brown University)

• classification and prediction of voluntary movements, and accelerated learning (with Jerome Sanes, Neuroscience Department, Brown University) 

From the computational and machine learning point of view, the analysis of fMRI data is a very challenging problem because the feature space is  extremely large (e.g. one 3D fMRI image can contain more than 50,000 voxels), the number of training examples is very small, and the data is very noisy. In contrast to slowly varying fMRI signals, the temporal resolution of EEG signals is very high.  Unfortunately like the fMRI, the EEG signals are also very noisy.  As a consequence the analysis of both signals is usually done not in real-time but using averaged values. For example, in analyzing fMRI data researchers are usually concerned with averaged activations of voxels (over repeated trials).  Similarly, EEG data are often averaged using an Event-Related-Potential (ERP) technique. As a consequence, it is very difficult to capture distributed patterns of neuronal activities.  For example, in analyzing  fMRI data the predominant approach consists of using univariate statistics which considers a time series of each voxel in isolation.

Our objective is to develop a system that can analyze the data in real-time using a multivariate approach.  We are currently developing various algorithms using information theoretic and machine learning approaches.  Our future plan is to extend the biologically inspired Bayesian model that we have developed within the domain of computer vision since it can learn new object categories from few examples, and deal with a large number of features. 

---