|Class of 2001
Graduated in 2008
|Undergraduate Institution: University of California, Berkeley
Major: Applied Math, Physics and Molecular & Cell Biology
Origin: Redwood City, CA
|Lab: Jack Gallant
Nonlinear Spatiotemporal Receptive Field (STRF) Estimation by Neural Computational and Machine Learning Methods
My research will focus on the computational analysis and modeling of neurophysiological data collected from experiments that are designed to unravel the visual processing in the cortex. If one treats a neuron as a computing unit that transforms input (stimuli) into action potentials, then its transfer function, or spatiotemporal receptive field (STRF), can be estimated using mathematical techniques. The kernel might reveal the computations mediating extraction of visual features to which the cell is tuned. More sophisticated nonlinear kernel estimation methods could enable us to understand processing in extra-striate visual areas, such as V2 and V4.
I will use techniques in statistics (principal component analysis, independent component analysis and data manifold modeling), engineering (linear systems analysis and neural networks), and machine learning (support vector machine and boosting) to analyze the physiological data in order to model the complex processing of the visual system. By using natural scene stimuli and simulated saccades in the stimulus set, I should drive visual neurons efficiently so that sufficient data can be collected during an experiment. I will first attempt to characterize the linear properties of cortical cells using Wiener kernel analysis. To capture the nonlinear properties of neuron, I also plan to use neural network, support vector machines and other nonlinear techniques to study the visual pathways. With knowledge extraction techniques and unsupervised learning, I hope to visualize the kernel of extra-striate cells.
Ever since Hubel and Wiesel (1959), we have made much progress in understanding the function of the visual system. However, progress in fully characterizing neurons in primary visual cortex (V1) and extra-striate visual areas has been limited by the complex nonlinearities within the system. Further progress in this area will depend critically on finding an effective method to estimate the input-output relationship of nonlinear visual neurons. Applying the more sophisticated nonlinear kernel estimation methods to extra-striate cortex could lead to the understanding of higher visual functions, such as object recognition, form vision and visual attention.
Although there are no efficient algorithms for object recognition and form vision, our brain does it reliably in almost real time. Successful models of the visual cortex could lead to novel algorithms that are superior to conventional computer vision. The human brain is by no means faster than a computer, but it is definitely superior in certain tasks. Understanding the working of our brain could lead to novel computers that could perform tasks that currently can only be performed by humans.
- Michael CK Wu, Stephen V David, and Jack L Gallant, "Complete Functional Characterization of Sensory Neurons by System Identification," Annual Review of Neuroscience 29, pp 477-505. (2006)
- Ryan Prenger, Michael CK Wu, Stephen V David, and Jack L Gallant, "Nonlinear V1 Responses to Natural Scenes Revealed by Neural Network Analysis," Neural Networks 17, pp663-679, 2004.