RI: Medium: An Analysis of the Consequences of Cortical Structure on Computation

Indiana University, Bloomington IN

Investigators

Abstract

Networks of cortical neurons are clearly organized into layers and columns, but relatively little is known about how these arrangements affect cortical computations. To approach this issue, a 512 micro-electrode array will be used to stimulate and record activity from hundreds of cortical neurons. With this, the inputs and outputs of a cortical network can be experimentally controlled. A recently-developed framework for understanding neural computation known as "reservoir computing" permits the computational power of neural networks to be quantified based on knowledge of their inputs and outputs. The 512-electrode system allows input stimulation to be localized to different cortical layers or columns. Similarly, outputs can be selected by recording from different layers or columns. Thus, the contributions of layers and columns to computations, and the types of computations they perform, can be measured and compared. The results of this research are expected to increase the understanding of how the cortex attains its remarkable computational power. In addition, the results of this work are expected to inform future designs of brain-like computing circuits. To promote scientific education and outreach, an existing software package called "Simbrain" will be further developed and disseminated. This package will allow students from high school level and above to understand how cortical networks transform inputs into outputs as they perform computations. Three specific aims will be pursued. First, the measurement of computational capacity must be based on realistic levels of random background stimulation. The high-conductance state is a well-known phenomenon in vivo resulting from constant random synaptic inputs, and is also a common feature in many (particularly reservoir computing) neural circuit models. The 512-electrode array will be used to deliver background stimulation to determine levels that will improve computational performance. Second, layer input and output locations will be studied. Using kernel quality and VC-dimension metrics, the computational power and role of each layer taken individually or as a whole will be assessed. It is possible that some layers more strongly generalize input patterns while others separate them. Thus it will be possible to dissect the computational contribution of each layer. Third, the same metrics will be applied to stimulation to one column which feeds to another. Here the computational power and role of multiple columns will be assessed, and any computational differences between columns directly stimulated by the array and columns stimulated by other columns can be observed.