|
|
|
|
Background: Since information present in the raw sensory data is inherently incomplete, ambiguous, and noisy, visual system has to integrate available sources of information to perform inferential reasoning about a scene. Architecturally, visual cortex appears to be designed for integrating multiple sources of information, yet the underlying computational mechanisms remain mostly unexplained. Recently, much focus has been on probabilistic frameworks for understanding the neural mechanisms and computational principles underlying inference within the brain. They naturally provide the ability to deal with complex uncertainty associated with ambiguous and noisy signal, and to integrate multiple sources of information across space and modality.
Research Project: We have developed a generative network model for the cortical architecture in which the unknown scene properties are inferred by integrating visual information non-locally. The model constructs probability distributions for local scene features and integrates these local features via message passing using belief propagation. Although not biologically realistic the architecture of our model exploits several organizational principles seen in visual cortex, such as local lateral connectivity between cortical hypercolumns and intralaminar connectivity within a hypercolumn. We have demonstrated how the generative network model might represent an appropriate theoretical framework for understanding visual processing for inferring intermediate-level visual representations, such as "direction of figure" (DOF) and motion, by integrating visual cues across space and modality. The simulation results show that the network model can account for several examples of perceptual ambiguity in DOF as well as the bias in perceived motion reported from psychophysical experiments.
An example of DOF estimates: an ambiguous spiral figure
Examples of DOF estimates: barberpole motions modulated by occluders
My previous research involved modeling the expert object recognition in the ventral visual pathway, comparative study of PCA and ICA on recognizing facial actions and identities, adaptive object recognition, and several projects in computer graphics.
|
