Introduction:

I am a 3rd-year PhD student in The Institute of Cognitive Science at The University of Louisiana at Lafayette. I am primarily interested computational evolutionary models of neural networks. Most previous work was done using the NEAT (NeuroEvolution of Augmenting Topologies) system, developed by Kenneth Stanley.

Together with Philip Tucker, I helped to implement a Java version of NEAT called ANJI, which is open-source and available for download. We no longer actively develop ANJI, but it includes domains such as Pole Balancing, Tic-Tac-Toe type games, and an Active Vision System. I do continue to moderate a NEAT user group.

Current Research:

I am currently working under the supervision of Dr. Anthony Maida. My focus is on attempting to model aspects of the evolution, development, and function of the neocortex. The model encodes networks in compact genetic representations, which are subject to mutation and crossover operators each generation. Genomes are translated into artificial neural networks, which exhibit both modular and hierarchical organization. The basic modular unit is modeled on the cortical minicolumn, a vertical structure of approximately 100 neurons. Minicolumns are clustered via horizontal connections, and constitute cortical sheets in the model. Sheets are arranged hierarchically, with connectivity converging as it travels up the hierarchy, and diffusing as it travels down the hierarchy. Each minicolumnar unit functions by accumulating input from units in the lower level of the hierarchy, passing inferences up the hierarchy, and propagating predictions down the hierarchy.

Once transcribed into a network phenotype, learning and development take place via Hebbian mechanisms, forming functional aggregates of cells whose interconnectivity is strengthened via the spatial and temporal proximity of input. The model has many influences, from Jeff Hawkins to Gerald Edelman. It is currently in development, and will be presented early next year.