STATEMENT OF PURPOSE
My motivation to study machine perception is a long-term passion for mathematics, programming, and the human brain. Before knowing much about machine intelligence, my primary motivation was to take part and securing the first place in Bangladesh Mathematical Olympiad for three times. Following this achievement, I graduated in Electrical Engineering taking major classes on theoretical signal processing at Bangladesh University of Engineering and Technology. Additionally, I accomplished a thesis on the modeling of brain hemodynamics and became enthusiastic for data-driven sciences. Having had later experiences in data analysis at Johns Hopkins, Cambridge University, and The MathWorks, I am now pursuing M.Sc. in Neuroscience …show more content…
Using fMRI, we discovered that several areas in the perisylvian cortex show within-modality adaptation. Interestingly, multiple areas show adaptation to both modalities. For example, parietal operculum (which is primarily somatosensory) and planum temporale (which is primarily auditory) show adaptation properties to both tactile and audio stimuli separately. Their relative response magnitudes were higher to the stimuli that are primarily supposed to process. This result suggests that there are likely one or more areas between the primary auditory (A1) and primary somatosensory (S1) cortices that respond equally to both modalities. Performing a principal component analysis, we confirmed that the response size of the secondary somatosensory cortex (S2) is same to both tactile and auditory …show more content…
We also used nonlinear regressions like kernel ridge and kernel support vector machines. Further, to capture the spatiotemporal patterns of ECoG data, we introduced a space-time kernel - A product of radial basis function (RBF) kernel with a sum of the polynomial (PLY), periodic (PER) and rational quadratic (RQ) kernels. The three components (RBFxPLY, RBFxPER, and RBFxRQ) capture the long-term nonlinear trend, local periodicity, and short-term irregularities in the data, respectively. We also implemented a neural network trained with Bayesian regularization for the same prediction
My motivation to study machine perception is a long-term passion for mathematics, programming, and the human brain. Before knowing much about machine intelligence, my primary motivation was to take part and securing the first place in Bangladesh Mathematical Olympiad for three times. Following this achievement, I graduated in Electrical Engineering taking major classes on theoretical signal processing at Bangladesh University of Engineering and Technology. Additionally, I accomplished a thesis on the modeling of brain hemodynamics and became enthusiastic for data-driven sciences. Having had later experiences in data analysis at Johns Hopkins, Cambridge University, and The MathWorks, I am now pursuing M.Sc. in Neuroscience …show more content…
Using fMRI, we discovered that several areas in the perisylvian cortex show within-modality adaptation. Interestingly, multiple areas show adaptation to both modalities. For example, parietal operculum (which is primarily somatosensory) and planum temporale (which is primarily auditory) show adaptation properties to both tactile and audio stimuli separately. Their relative response magnitudes were higher to the stimuli that are primarily supposed to process. This result suggests that there are likely one or more areas between the primary auditory (A1) and primary somatosensory (S1) cortices that respond equally to both modalities. Performing a principal component analysis, we confirmed that the response size of the secondary somatosensory cortex (S2) is same to both tactile and auditory …show more content…
We also used nonlinear regressions like kernel ridge and kernel support vector machines. Further, to capture the spatiotemporal patterns of ECoG data, we introduced a space-time kernel - A product of radial basis function (RBF) kernel with a sum of the polynomial (PLY), periodic (PER) and rational quadratic (RQ) kernels. The three components (RBFxPLY, RBFxPER, and RBFxRQ) capture the long-term nonlinear trend, local periodicity, and short-term irregularities in the data, respectively. We also implemented a neural network trained with Bayesian regularization for the same prediction