Afternoon Keynote | Room: 1151
Amir Gandjbakhche, Senior Investigator, National Institutes of Health
Hadis Dashtestani, Pre-Doctorate Fellow, National Institute of Child Health & Human Development
To explore the relationship between two sets of multi-dimensional variables, the coordinate system in which variables are described is crucial. Even strong correlations between sets of variables may not emerge if an inappropriate coordinate system is used. Canonical Correlation Analysis (CCA), is an unsupervised learning method which learns a pair of linear transformations (coordinate systems), one for each set, such that the projections of each set onto these coordinates are maximized. Often in neuroimaging datasets, the number of features is larger than samples, as in our case causing overfitting of the training set. We used regularized CCA (R-CCA) to avoid an overfitting problem. Moreover, the regularization parameter keeps the parameters of the model small, so it is less likely to face high bias problem. To estimate the robustness of R-CCA model, we used the leave-one-out cross validation (LOOCV) method and calculated the mean squared error for each left out data-point.
Dr. Amir Gandjbakhche is a Senior Investigator and Head of the Section on Analytical and Functional Biophotonics of NICHD. He obtained his Ph.D. in physics with a biomedical engineering specialty from the University of Paris in 1989. He is a Fellow of SPIE, the largest society of optical engineers. Dr. Gandjbakhche leads a research group that uses different optical sources of contrast such as endogenous or exogenous fluorescent labels, absorption (e.g., hemoglobin or chromophore concentration) in order to devise quantitative theories at the board, and designs instrumentation at the bench, and brings the imaging system to the bedside. Two areas of interest are the use of near infrared spectroscopy to assess cognitive function in Traumatic Brain Injury and Autistic Spectrum Disorder patients, and using specific fluorescently labeled HER2 imaging agent to monitor monoclonal antibody therapy of breast tumor.
Hadis is PhD student at University of Maryland, Baltimore County. She is doing her PhD project at Section on Analytical and Functional Biophotonics (SAFB)/ National Institute of Health. Her area of expertise is Machine Learning and Pattern Recognition. She is working on brain modeling through non-invasive imaging techniques such as functional Near-Infrared Spectroscopy (fNIRS) and Electroencephalography (EEG).
Recent Comments