Computational Neuroscience Mentoring

Shawn Barr, Clark University. As part of a larger effort to model the human visual cortex, Shawn Barr implemented a set of psychophysical tasks that are relatively easy for a human to complete but could not be solved by a computer using a simple heuristic. To do so, he used an algorithm to generate a set of images composed of statistically similar line segments. In the first stage of experimentation the participant will be asked whether or not the segments are arranged in such a way that a closed shape can be imaged.

Daniel Coates, Portland State University Computer Science Department. Dan Coates, working with the PetaVision group at Los Alamos National Laboratory (LANL), helped architect the software simulator for neural modeling. The goal was to implement a system that was modular and extensible yet high performance. The current framework supports arbitrary populations of neurons connected in anything-to-anything fashion, giving the researcher freedom to experiment with diverse synaptic connectivity patterns and multiple neuronal types. Although the system will ultimately run on specialized Petascale hardware, the software utilizes multi-core processors and contains extensive optimizations to yield impressive performance on commodity computers. Python and MATLAB may be used to automate network construction and analysis.

The prototypical application models the early stages of visual processing. An image file is presented to a simulated retina, and neural activity propagates through successive biologically plausible layers of the visual pathway. Established techniques from neuroscience research are employed to analyze the results of each run. The processing power of the network is evaluated by studying its performance on psychophysical tasks that are difficult for traditional computer vision systems. Demonstrating proficiency on such tasks may pave the way to understanding how computers might someday exhibit higher-level cognitive ability.

David vanMaanen, Penn State University. In order to complete a computational model of the visual cortex that would detect arbitrary closed curvatures, David vanMaanen implemented a model of the primary visual cortex, which will utilize synchrony to detect arbitrary curvatures. He used weighted connections between neurons to bind the spike timing of each of the neurons on a complete curvature. This synchrony makes the curvature stand out from any clutter that might be in the image. The weight of the connections was based on whether or not the orientations of the connected cells were tangent to the same circle and the radius of that circle.

Ethan Brown. Based on recent experiments in several brain regions, we hypothesized that nonlinear dendritic subunits could enable simple cells in V1 to achieve good orientation selectivity despite the very low aspect ratio of their afferent input from the LGN. We compared the response of model simple cells, both with and without non-linear dendritic subunits, to LGN inputs activated by short oriented line segments and sinusoidal gratings. The model simple cell was based on a previously published multi-compartmental simulation of a reconstructed cat V1 spiny stellate cell. Non-linear dendritic subunits were added using a combination of NMDA- and voltage-gated channels. Interneurons possessing the same morphology and input aspect ratio as the V1 simple cells were also incorporated to provide feedforward flanking inhibition and contrast invariance. Excitatory lateral interactions between V1 simple cells were not modeled, to mimic the conditions present in cortical inactivation experiments.

In the absence of non-linear dendritic subunits, it was impossible to simultaneously obtain realistic orientation and spatial frequency tuning curves when the aspect ratio of the LGN input was close to the mean value measured experimentally; increasing flanking inhibition to improve orientation tuning caused spatial frequency tuning to become unphysiologically sharp. However, by breaking the receptive field into smaller nonlinear dendritic subunits, each with twice the aspect ratio of the total LGN input, both orientation and spatial frequency tuning became consistent with experimental values. We propose that non-linear dendritic subunits provide a general mechanism for generalizing over some stimulus features, such as precise spatial phase, while preserving specificity to other stimulus features, such as orientation.

Stoyana Alexandrova. As part of a larger project to model the human visual cortex, Stoyana Alexandrova was mentored in C programming language, Octave, and neuroscience topics in relation to visual circuit systems. She helped put together a control flow of the main algorithm of the PetaVision code and worked on modeling the connectivity patterns of the brain layers. As any two layers in the visual system are connected, the cells in the LGN of the brain connect to the cells of the lower hierarchical layer, the retina, to communicate input signals and propel them through the visual circuit. To model patterned connectivity, she wrote a subroutine for visual cells in one layer connecting to specific cells in another, according to given appropriate variables. She closely worked with and was mentored by Dr. Craig Rasmussen (LANL).

Contact: Garrett Kenyon gkenyon@lanl.gov