An Architecturally Relevant Model for Creating Orientation Maps of Primary Visual Cortex
thesisposted on 28.06.2013 by Jennifer E. Anderson
In order to distinguish essays and pre-prints from academic theses, we have a separate category. These are often much longer text based documents than a paper.
Building biologically relevant models is a area of research experiencing a boom in growth. Specifically, many companies and research teams are trying to understand a very efficient and powerful organ of the body -- the brain. The brain weighs 3 pounds consuming a mere 20 watts of energy, yet it is able to perform seemingly thousands of tasks, whether they be visual, cognitive, or hormonally driven, all simultaneously with the utmost efficiency. Understanding how the brain works is an important milestone for humankind, but it is far from easy. One starting point is to try and model the input-output responses of the brain and somehow formalize these response behaviors. When creating computational models, it is important to try and stay as true to the biological model as possible. This includes constraining the model to the architecture and the wiring of components of the biological model. The reason we constrain ourselves to the architecture is because we need to allow for any unaccounted for properties to develop within the model. Emergence, for example, is a one of these properties; it is the way complex patterns arise out from some multiple of relatively simple interactions. Emergent properties are common in the brain (intelligence being the most commonly cited). The visual system comprises nearly one-third of all of the processing within the human brain and is also one of the most widely studied areas of the nervous system, thus it is appropriate to try and create a model from this plethora of data. In this thesis, I demonstrate that by following the basic rules of brain organization -- sheets of neurons working in parallel organized into a specific hierarchy -- I can achieve a biologically meaningful output. This model is composed solely of leaky integrate-and-fire neurons constrained to a specific layout. In the end, I show that this biologically-inspired model can produce outputs equivalent to empirical data. Additionally, the architecture allows for emergent properties to develop (i.e. Long Term Potentiation between neighboring neurons). I detail the parameters of the model and showcase the parameters that can generate biologically similar maps. These comparisons are done across a multiplicity of dimensionless measures.