We are investigating representation, computation and response properties in both biologically realistic and theoretical network architectures.  For example, we have developed a biophysically realistic model of primary visual cortex, demonstrating how both classical and extra-classical response properties (i.e. contextual effects) arise from network dynamics and neuronal circuitry.   We have also developed more theoretical models focused on how networks of cortical hypercolumns in visual cortex might integrate visual cues, across space and stream, using the machinery of Bayesian Inference.  A subset of these principles have been incorporated into artificial neural networks architectures and applied to problems in medical image analysis—e.g. cancer detection. Jim Wielaard, Kyungim Baek, Jeremy Lewi, Jianing Shi. 

Our group uses neuroimaging to undercover the neural origins of behavior, particularly within the context of processing visual stimuli.  For example we have used single-trial analysis of high-density EEG to undercover the components of object recognition during rapid serial visual presentation.  More recently we have integrated this approach with fMRI, correlating the high-spatial resolution hemodynamic response with high-temporal resolution electroencephalography. Adam Gerson, An Luo, Marios Philiastides, Robin Goldman.

Our group is investigating statistical representations of signals and images, as a means to better understand how these representations might be developed by visual networks, as well as how they can be exploited by subsequent processing downstream (e.g. in recognition/classification).  We have developed a hierarchical image probability model for learning local and non-local (i.e. contextual) relationships of image structure.  The result is a generative model, able to optimally classify, synthesize and compress images, and which has been tested for medical image analysis applications.  We have also developed techniques, based on non-negative matrix factorization, for improving brain cancer detection in magnetic resonance spectroscopy image.  The approach uses statistical properties of the signal measured across the image (or volume) to blindly recovery spectral signatures of metabolites and tissues.  The method is based on a factorization that, for natural images, yields basis vectors that are very similar to receptive field structure of simple cells in primary visual cortex. Kyungim Baek, Shuyan Du, Jianing Shi.

Machine Learning

At the core of much of our research is understanding and characterizing learning and adaptive systems, whether biological or artificial. In our research machine learning is used both to process data as well as serving as a theoretical model for biological learning. A major focus of our research is toward developing new approaches to machine learning, including new architectures, algorithms, and optimization strategies.  Particular areas of research include blind source separation, probabilistic graphical models, and adaptive subspace reduction. Adam Gerson, Shuyan Du, Jeremy Lewi.