Modeling Early Vision: Probabilistic Computation Using Spiking Neurons, Population Codes, and CUDA

Master's thesis presented by Dan Coates to the Portland State University Computer Science Department, 2009.
Advisor: Melanie Mitchell
With committee member: Dan Hammerstrom

---

Networks of spiking neurons are gaining popularity among neuroscientists for their ability to accurately model biological behavior. From a theoretical standpoint, this abstraction has rich computational power, with many simple nodes and a unique binary communication channel. Although these networks offer a promising new computing paradigm, the question of how to reliably harness their power remains unsolved.

To explore the practical capabilities of spiking neuronal networks, I have implemented an established model of the visual cortex from neuroscience literature. This model leverages notions from biologically-inspired image processing, providing a benchmark task of edge detection in a small grayscale bitmap. To facilitate the multiple trials and network sizes required for my experiments, I parallelized the algorithm for concurrent execution on multiple core hardware. Compute Unified Device Architecture (CUDA) is a new architecture for parallel programming using graphics processing cards (GPUs) with many dozens of cores. A straightforward CUDA port achieves 20x speedup compared to the single core CPU version.

A key feature of the particular model I implemented is that the computations it performs are statistical in nature, with calculations inherently distributed throughout the population of overlapping nodes. This characteristic lends a natural robustness to the network operation, but interpretation of the functional results is an open problem. Utilizing concepts from signal detection theory and machine learning, I demonstrate estimation of a numerical value from the spike output of the network of neurons, a task that parallels efficient signal transmission with an array of noisy binary units.